From Newsgroup: sci.logic

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly. https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

This is quite fun, how some TLA+ guy fears
the full state of queue like the devil in
itself. But I guess if a service rate is

low and the producer has not much to do to
produce its work items, the arrival rate
has nevertheless to adapt, and dealing

with "full states", which are wrongly
called deadlock here, is the normal:

Tutorial-style talk - BlockingQueue https://github.com/lemmy/BlockingQueue/tree/main

Prolog is in good position. The bird box
model has a redo port. So sometimes switching
from push to pull, can help without doing

Deadlock Exorcism. You can also translate
the bird box ports into pi-calculus:

A pi-calculus Specification of Prolog
Benjamin Z. Li - University of Pennsylvania
11 Apr 1994, European Symposium on Programming,
Prolog, Unification, Backtracking https://scispace.com/pdf/a-pi-calculus-specification-of-prolog-3qf2pf04ud.pdf

Have Fun!

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly. https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Because I use WebGPU and not WebGL. And
because WebGPU can adresss modern GPU
developed with the NVIDIA Volta evolution,

which happened in 2017. Namley that compute
shaders are not any more subject to the
realization restriction of lock step

execution, but have independent thread state.
And because there is independent thread state
there is also independent time spent for a

a work item by each logical thread, if the
submitted logical thread uses a lot of branching
logic or even loops. But the use of branching

and loops is encouraged in independent thread
state programming of compute shaders. The variables
that can drive such logic are the scalar variables:

Tour of WGSL - Control Flow
https://google.github.io/tour-of-wgsl/control-flow/

Then not to waste GPU compute time, by logical
threads doing nothing. You will need to
introduce some load balancing among multiple

logical threads. And MPMC queues are one way to
readize load balancing. Compute shaders with
producer and consumer entry points are proposed

as fundamental architecture by Thunder Kittens:

ThunderKittens: Simple, Fast, and Adorable AI Kernels https://arxiv.org/abs/2410.20399

They are used by this SpaceX acquisition:

Composer 2 Technical Report
https://arxiv.org/abs/2603.24477

Thunder Kittens uses Hardware support, i.e. tma_expect().

Bye

Chris M. Thomasson schrieb:

never meant to be used in a GPU.
Dmitry CAS version can be used, but

Why do you even need a mpmc queue
in your compute shader anyway?

Mild Shock schrieb:

Hi,

This is quite fun, how some TLA+ guy fears
the full state of queue like the devil in
itself. But I guess if a service rate is

low and the producer has not much to do to
produce its work items, the arrival rate
has nevertheless to adapt, and dealing

with "full states", which are wrongly
called deadlock here, is the normal:

Tutorial-style talk - BlockingQueue https://github.com/lemmy/BlockingQueue/tree/main

Prolog is in good position. The bird box
model has a redo port. So sometimes switching
from push to pull, can help without doing

Deadlock Exorcism. You can also translate
the bird box ports into pi-calculus:

A pi-calculus Specification of Prolog
Benjamin Z. Li - University of Pennsylvania
11 Apr 1994, European Symposium on Programming,
Prolog, Unification, Backtracking https://scispace.com/pdf/a-pi-calculus-specification-of-prolog-3qf2pf04ud.pdf

Have Fun!

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Is a trivial control construct for(),
when used in a compute shader with
NVIDIA Volta evolution, i.e. MIMD,

can lead to different time spend by
individual compute shaders:

fn main(global_id : i32) {
i : i32 = 0;
while (i < globa_id) {
i++;
}
}

You can visiualize as the time spent
by each logical thread as follows:

global id, logical thread life line
1 [ ]
2 [ ]
3 [ ]
4 [ ]
5 [ ]
Etc..

With work items and load balancing you
could run the above with a lower number
of logical threads, I am writing the

work item number now inside the sub life
line inside the overall life line of
the logical thread:

worker , worker work items
A [3 ]
B [4 ][2 ]
C [5 ][1 ]

The overall time slightly increased by 1,
i.e. the case global_id = k combined
with the case global_id = n-k+1 . Also

one worker didn't have two work items,
only one work item. But the number of
logical threads needed was halfed.

Ok, a mpmc queue will be not that
intelligent, concerning the work sheduling.
But one could experiment with mpmc queue

priority queues etc.. etc..

Have Fun!

Bye

Mild Shock schrieb:

Hi,

Because I use WebGPU and not WebGL. And
because WebGPU can adresss modern GPU
developed with the NVIDIA Volta evolution,

which happened in 2017. Namley that compute
shaders are not any more subject to the
realization restriction of lock step

execution, but have independent thread state.
And because there is independent thread state
there is also independent time spent for a

a work item by each logical thread, if the
submitted logical thread uses a lot of branching
logic or even loops. But the use of branching

and loops is encouraged in independent thread
state programming of compute shaders. The variables
that can drive such logic are the scalar variables:

Tour of WGSL - Control Flow https://google.github.io/tour-of-wgsl/control-flow/

Then not to waste GPU compute time, by logical
threads doing nothing. You will need to
introduce some load balancing among multiple

logical threads. And MPMC queues are one way to
readize load balancing. Compute shaders with
producer and consumer entry points are proposed

as fundamental architecture by Thunder Kittens:

ThunderKittens: Simple, Fast, and Adorable AI Kernels https://arxiv.org/abs/2410.20399

They are used by this SpaceX acquisition:

Composer 2 Technical Report
https://arxiv.org/abs/2603.24477

Thunder Kittens uses Hardware support, i.e. tma_expect().

Bye

Chris M. Thomasson schrieb:

never meant to be used in a GPU.
Dmitry CAS version can be used, but

Why do you even need a mpmc queue
in your compute shader anyway?

Mild Shock schrieb:

Hi,

This is quite fun, how some TLA+ guy fears
the full state of queue like the devil in
itself. But I guess if a service rate is

low and the producer has not much to do to
produce its work items, the arrival rate
has nevertheless to adapt, and dealing

with "full states", which are wrongly
called deadlock here, is the normal:

Tutorial-style talk - BlockingQueue
https://github.com/lemmy/BlockingQueue/tree/main

Prolog is in good position. The bird box
model has a redo port. So sometimes switching
from push to pull, can help without doing

Deadlock Exorcism. You can also translate
the bird box ports into pi-calculus:

A pi-calculus Specification of Prolog
Benjamin Z. Li - University of Pennsylvania
11 Apr 1994, European Symposium on Programming,
Prolog, Unification, Backtracking
https://scispace.com/pdf/a-pi-calculus-specification-of-prolog-3qf2pf04ud.pdf


Have Fun!

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

For graphics, rendering, in the worst case your
FPS can go down. Then you have a 2x as performant
graphic card, and sundently the FPS is ok again!

Or you render into a smaller screen, with less
number of pixels, and things turn good again.
So for my pixel phone AI experiment, or what I

will toy around on a AI laptop, I am singing:

I'm a spinner, I'm a sinner
I spin on CAS loops for my dinner
Some call it busy-wait, I call it fate
When the queue is empty, I just rotate

Bye

Whats better Steve Miller or Muddy Waters?

The Joker
https://www.youtube.com/watch?v=dV3AziKTBUo

Hoochie Coochie Man
https://www.youtube.com/watch?v=e_l6A7krjrQ

Mild Shock schrieb:

Hi,

Is a trivial control construct for(),
when used in a compute shader with
NVIDIA Volta evolution, i.e. MIMD,

can lead to different time spend by
individual compute shaders:

fn main(global_id : i32) {
-a-a i : i32 = 0;
-a-a while (i < globa_id) {
-a-a-a-a-a i++;
-a-a }
}

You can visiualize as the time spent
by each logical thread as follows:

global id, logical thread life line
1-a-a-a-a [-a-a-a ]
2-a-a-a-a [-a-a-a-a-a-a-a ]
3-a-a-a-a [-a-a-a-a-a-a-a-a-a-a-a ]
4-a-a-a-a [-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a ]
5-a-a-a-a [-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a ]
Etc..

With work items and load balancing you
could run the above with a lower number
of logical threads, I am writing the

work item number now inside the sub life
line inside the overall life line of
the logical thread:

worker , worker work items
A-a-a-a-a [3-a-a-a-a-a-a-a-a-a-a ]
B-a-a-a-a [4-a-a-a-a-a-a-a-a-a-a-a-a-a-a ][2-a-a-a-a-a-a ]
C-a-a-a-a [5-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a ][1-a-a ]

The overall time slightly increased by 1,
i.e. the case global_id = k combined
with the case global_id = n-k+1 . Also

one worker didn't have two work items,
only one work item. But the number of
logical threads needed was halfed.

Ok, a mpmc queue will be not that
intelligent, concerning the work sheduling.
But one could experiment with mpmc queue

priority queues etc.. etc..

Have Fun!

Bye

Mild Shock schrieb:

Hi,

Because I use WebGPU and not WebGL. And
because WebGPU can adresss modern GPU
developed with the NVIDIA Volta evolution,

which happened in 2017. Namley that compute
shaders are not any more subject to the
realization restriction of lock step

execution, but have independent thread state.
And because there is independent thread state
there is also independent time spent for a

a work item by each logical thread, if the
submitted logical thread uses a lot of branching
logic or even loops. But the use of branching

and loops is encouraged in independent thread
state programming of compute shaders. The variables
that can drive such logic are the scalar variables:

Tour of WGSL - Control Flow
https://google.github.io/tour-of-wgsl/control-flow/

Then not to waste GPU compute time, by logical
threads doing nothing. You will need to
introduce some load balancing among multiple

logical threads. And MPMC queues are one way to
readize load balancing. Compute shaders with
producer and consumer entry points are proposed

as fundamental architecture by Thunder Kittens:

ThunderKittens: Simple, Fast, and Adorable AI Kernels
https://arxiv.org/abs/2410.20399

They are used by this SpaceX acquisition:

Composer 2 Technical Report
https://arxiv.org/abs/2603.24477

Thunder Kittens uses Hardware support, i.e. tma_expect().

Bye

Chris M. Thomasson schrieb:

never meant to be used in a GPU.
Dmitry CAS version can be used, but

Why do you even need a mpmc queue
in your compute shader anyway?

Mild Shock schrieb:

Hi,

This is quite fun, how some TLA+ guy fears
the full state of queue like the devil in
itself. But I guess if a service rate is

low and the producer has not much to do to
produce its work items, the arrival rate
has nevertheless to adapt, and dealing

with "full states", which are wrongly
called deadlock here, is the normal:

Tutorial-style talk - BlockingQueue
https://github.com/lemmy/BlockingQueue/tree/main

Prolog is in good position. The bird box
model has a redo port. So sometimes switching
from push to pull, can help without doing

Deadlock Exorcism. You can also translate
the bird box ports into pi-calculus:

A pi-calculus Specification of Prolog
Benjamin Z. Li - University of Pennsylvania
11 Apr 1994, European Symposium on Programming,
Prolog, Unification, Backtracking
https://scispace.com/pdf/a-pi-calculus-specification-of-prolog-3qf2pf04ud.pdf


Have Fun!

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Since the main goal is to change the
geometry of how work items are assigned
and processed, and since work item processing

costs much more time then queue API calls.
It is irrelevant whether the queue API are
ultra fast or not, or whether they spin or

not. Usually one can affort both things,
since balancing outweights these small extra
costs or silly occupation of a platform thread.

It could be also feasible, in the total
balance sheet of execution and energy consumption
of your work items along workers to use

Atomic.pause() equivalents, special instructions,
for spinning. This will be added in 2027 to JavaScript:

For example, in Intel x86, it may be a
pause instruction as per Intel's optimization manual. https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Atomics/pause

But I have to see what works and what doesn't
work. The above are only theoretical
considerations about the friction that queues

induce. Usually friction hurts much less
than balant scheduling mistakes, i.e. unfavorable
schedulings. golang is very much bugged by

unfavorable schedulings. While on paper golang
looks like brilliant idea, examples that create
similar patterns of runtimes as in the below

for (int i=0; i<global_id; i++) can create
horrible scheduling and very bad an erratic
behaviour, especially if the synchronization

mechanism and backpressure signals have to travel
long chains of channel waits. So you will get not
into the terrain of "dead lock" questions, but

basically enter the terrain of "fairness" questions.

Bye

Mild Shock schrieb:

Hi,

For graphics, rendering, in the worst case your
FPS can go down. Then you have a 2x as performant
graphic card, and sundently the FPS is ok again!

Or you render into a smaller screen, with less
number of pixels, and things turn good again.
So for my pixel phone AI experiment, or what I

will toy around on a AI laptop, I am singing:

-a I'm a spinner, I'm a sinner
-a I spin on CAS loops for my dinner
-a Some call it busy-wait, I call it fate
-a When the queue is empty, I just rotate

Bye

Whats better Steve Miller or Muddy Waters?

The Joker
https://www.youtube.com/watch?v=dV3AziKTBUo

Hoochie Coochie Man
https://www.youtube.com/watch?v=e_l6A7krjrQ

Mild Shock schrieb:

Hi,

Is a trivial control construct for(),
when used in a compute shader with
NVIDIA Volta evolution, i.e. MIMD,

can lead to different time spend by
individual compute shaders:

fn main(global_id : i32) {
-a-a-a i : i32 = 0;
-a-a-a while (i < globa_id) {
-a-a-a-a-a-a i++;
-a-a-a }
}

You can visiualize as the time spent
by each logical thread as follows:

global id, logical thread life line
1-a-a-a-a [-a-a-a ]
2-a-a-a-a [-a-a-a-a-a-a-a ]
3-a-a-a-a [-a-a-a-a-a-a-a-a-a-a-a ]
4-a-a-a-a [-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a ]
5-a-a-a-a [-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a ]
Etc..

With work items and load balancing you
could run the above with a lower number
of logical threads, I am writing the

work item number now inside the sub life
line inside the overall life line of
the logical thread:

worker , worker work items
A-a-a-a-a [3-a-a-a-a-a-a-a-a-a-a ]
B-a-a-a-a [4-a-a-a-a-a-a-a-a-a-a-a-a-a-a ][2-a-a-a-a-a-a ]
C-a-a-a-a [5-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a ][1-a-a ]

The overall time slightly increased by 1,
i.e. the case global_id = k combined
with the case global_id = n-k+1 . Also

one worker didn't have two work items,
only one work item. But the number of
logical threads needed was halfed.

Ok, a mpmc queue will be not that
intelligent, concerning the work sheduling.
But one could experiment with mpmc queue

priority queues etc.. etc..

Have Fun!

Bye

Mild Shock schrieb:

Hi,

Because I use WebGPU and not WebGL. And
because WebGPU can adresss modern GPU
developed with the NVIDIA Volta evolution,

which happened in 2017. Namley that compute
shaders are not any more subject to the
realization restriction of lock step

execution, but have independent thread state.
And because there is independent thread state
there is also independent time spent for a

a work item by each logical thread, if the
submitted logical thread uses a lot of branching
logic or even loops. But the use of branching

and loops is encouraged in independent thread
state programming of compute shaders. The variables
that can drive such logic are the scalar variables:

Tour of WGSL - Control Flow
https://google.github.io/tour-of-wgsl/control-flow/

Then not to waste GPU compute time, by logical
threads doing nothing. You will need to
introduce some load balancing among multiple

logical threads. And MPMC queues are one way to
readize load balancing. Compute shaders with
producer and consumer entry points are proposed

as fundamental architecture by Thunder Kittens:

ThunderKittens: Simple, Fast, and Adorable AI Kernels
https://arxiv.org/abs/2410.20399

They are used by this SpaceX acquisition:

Composer 2 Technical Report
https://arxiv.org/abs/2603.24477

Thunder Kittens uses Hardware support, i.e. tma_expect().

Bye

Chris M. Thomasson schrieb:

never meant to be used in a GPU.
Dmitry CAS version can be used, but

Why do you even need a mpmc queue
in your compute shader anyway?

Mild Shock schrieb:

Hi,

This is quite fun, how some TLA+ guy fears
the full state of queue like the devil in
itself. But I guess if a service rate is

low and the producer has not much to do to
produce its work items, the arrival rate
has nevertheless to adapt, and dealing

with "full states", which are wrongly
called deadlock here, is the normal:

Tutorial-style talk - BlockingQueue
https://github.com/lemmy/BlockingQueue/tree/main

Prolog is in good position. The bird box
model has a redo port. So sometimes switching
from push to pull, can help without doing

Deadlock Exorcism. You can also translate
the bird box ports into pi-calculus:

A pi-calculus Specification of Prolog
Benjamin Z. Li - University of Pennsylvania
11 Apr 1994, European Symposium on Programming,
Prolog, Unification, Backtracking
https://scispace.com/pdf/a-pi-calculus-specification-of-prolog-3qf2pf04ud.pdf


Have Fun!

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

For those who didn't get it. Its not about programming
languages Python versus C/C++ per se. Its about the
runtime enviroments they deliver and/or require.

Python is traditionally bugged by the GIL (Global
Interpreter Lock). I tested this here with Python,
using just a porting of the Java class Thread:

Parallel -C-WAM: 1.7 Giga Lips on a CPU https://medium.com/@janburse_2989/parallel-%CF%80-wam-1-9-giga-lips-on-a-cpu-8a984e75af44

It didn't work. Horrible performance of threading.Thread.
But most of AI does utilize both CPU and GPU, and
bare metal access to system threads, with the avoidance

some annoying GIL nonsense, is essential. But there
is a silver lining on the horizon. Namely free
threading, Starting with the 3.13 release,

but I havent tested it yet:

Python support for free threading https://docs.python.org/3/howto/free-threading-python.html

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly. https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

While HBM and RDMA happen outside of a the main
silicon chip. Amazing things are now happening
inside a silicon chip as found in AI laptops.

Basically XILINX later acquired by AMD, had
already the Versal architecture. Where FGPA was
used to custom wire chips. The Versal area
had already Network-on-Chip (NoC): https://www.adiuvoengineering.com/post/microzed-chronicles-versal-part-two-device-architecture

While a Ryzen AI 7 350 /w Radeon 860M does not
really have a versal area anymore. But the
Network-on-Chip (NoC) survived, with twist:

GEMM Performance Generations of Ryzen AI NPUs
4.3 On-The-Fly Tensor Transformations
We extensively exploit the multi-dimensional
addressing feature of DMAs to reorganize data into
tiled layouts, as needed by the NPU cores.
https://arxiv.org/abs/2512.13282v1

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly. https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

How it started:

Captain: Throw the switch, Scotty!
Enterprise: Cloaking Device makes it invisible
Spock: Military secrets are the most fleeting of all.
Kirk Escapes the Romulans - The Enterprise Incident https://www.youtube.com/watch?v=AusAGjwlql8

How its going:

CEO Jensen Huang said the company has rCLlargely
concededrCY ChinarCOs artificial intelligence chip
market to Huawei, as U.S. export restrictions
continue to reshape the global AI semiconductor landscape. https://www.cnbc.com/2026/05/21/nvidia-jensen-huang-china-ai-chip-market-huawei.html

Bye

P.S.: What does China do?

HuaweirCOs semiconductor chief He Tingbo at the IEEE
ISCAS 2026 conference, Huawei's Tau Scaling Law is a newly
introduced semiconductor design framework that
shifts the industryrCOs optimization focus from
geometric scaling (shrinking physical transistor
sizes) to temporal scaling (compressing signal
propagation delay).
Nvidia Gave Up China - 4 Days Later THIS Happened https://www.youtube.com/watch?v=dLLw-qADKSU

Mild Shock schrieb:

Hi,

While HBM and RDMA happen outside of a the main
silicon chip. Amazing things are now happening
inside a silicon chip as found in AI laptops.

Basically XILINX later acquired by AMD, had
already the Versal architecture. Where FGPA was
used to custom wire chips. The Versal area
had already Network-on-Chip (NoC): https://www.adiuvoengineering.com/post/microzed-chronicles-versal-part-two-device-architecture

While a Ryzen AI 7 350 /w Radeon 860M does not
really have a versal area anymore. But the
Network-on-Chip (NoC) survived, with twist:

GEMM Performance Generations of Ryzen AI NPUs
4.3 On-The-Fly Tensor Transformations
We extensively exploit the multi-dimensional
addressing feature of DMAs to reorganize data into
tiled layouts, as needed by the NPU cores.
https://arxiv.org/abs/2512.13282v1

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic


Hi,

My mother is worried that I fucked Lane W.
aka Micro Penis mother 24 hours straight.
She was screaming, basically singing all

the arias from operas that Luciano Pavarotti
usually sings. You Lane W. aka Micro Penis
should have heard it, since you

live in the basement of your mothers house.

Bye

Lane W schrieb:

Mild Shock wrote:

Hi,

How it started:

Captain: Throw the switch, Scotty!
Enterprise: Cloaking Device makes it invisible

You stupid ass. You posted this twice.

Mild Shock wrote:

So a few days later comes out the LLaMA, I do some calculations and I
figure out rCLOkay, 65 billion parameters. You probably need about 40 gigs
of RAM, with 4-bit quantization. So this can run on a MacBook. Why not
do it?rCY

you are a shame to your mother

Perplexity Increase: Quantizing to 4-bit typically increases perplexity

Reasoning & Coding: Complex reasoning chains and coding tasks suffer

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic


Hi,

Micro penis brain is in constant hiatus.
He can even not detect a trope.

LoL

Bye

Lane W schrieb:

Mild Shock wrote:

Hi,

My mother is worried that I fucked Lane W.
aka Micro Penis mother 24 hours straight.
She was screaming, basically singing all

the arias from operas that Luciano Pavarotti
usually sings. You Lane W. aka Micro Penis
should have heard it, since you

live in the basement of your mothers house.

No, actually remarkably, I don't. According to google I live 433

miles away from her.

Strike!

See, what i said about you was spot on.

What you said about me was generic and incorrect.

You really suck, man.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

If any of you guys do not understand what
is meant by or what the implications are:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Well I wouldn't care less. There are two
outcomes for numb nuts:

- Ignoramus: They don't understand it, but
they will understand it before they die.

- Ignorabimus: They don't understand it, and
will never understand it, and they die.

So who cares, its not my problem, you people
are stupid as fuck, and slow as fuck...

Bye

Mild Shock schrieb:

Hi,

Micro penis brain is in constant hiatus.
He can even not detect a trope.

LoL

Bye

Lane W schrieb:

Mild Shock wrote:

Hi,

My mother is worried that I fucked Lane W.
aka Micro Penis mother 24 hours straight.
She was screaming, basically singing all

the arias from operas that Luciano Pavarotti
usually sings. You Lane W. aka Micro Penis
should have heard it, since you

live in the basement of your mothers house.

No, actually remarkably, I don't. According to google I live 433

miles away from her.

Strike!

See, what i said about you was spot on.

What you said about me was generic and incorrect.

You really suck, man.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Now you can compare this here from 2008
with modern AI Laptops for 500-1000 USD:

Google spotlights data center inner workings https://web.archive.org/web/20131019063218/http://news.cnet.com/8301-10784_3-9955184-7.html

There is a striking similarity, only what
once occupied a rack, has now the size
of your plam, all inside one silicon chip:

- Multiple CPU cores on the same chip
- Multiple GPU units on the same chip
- Network on the same chip communication
- Crossbar caches on the same chip
- Disk controllers on the same chip
- Multi channel RAM access on the same chip

Pretty cool!

Bye

P.S.: Example such devices with iGPU:

Intel(R) Core(TM) Ultra 7 258V
AMD Ryzen AI 7 350 w/ Radeon 860M
Apple A18 Pro, Darwin Kernel Version 25.5.0
Snapdragon(R) X - X126100 - Qualcomm(R) Oryon(TM) CPU


Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly. https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

You are a moron, and you represent putin payed
trolls from the army of brainless troll morons.

Bye

Lane W schrieb:

Mild Shock has no idea who I am or what I represent

Mild Shock schrieb:

Hi,

If any of you guys do not understand what
is meant by or what the implications are:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Well I wouldn't care less. There are two
outcomes for numb nuts:

- Ignoramus: They don't understand it, but
-a they will understand it before they die.

- Ignorabimus: They don't understand it, and
-a will never understand it, and they die.

So who cares, its not my problem, you people
are stupid as fuck, and slow as fuck...

Bye

Mild Shock schrieb:

Hi,

Micro penis brain is in constant hiatus.
He can even not detect a trope.

LoL

Bye

Lane W schrieb:

Mild Shock wrote:

Hi,

My mother is worried that I fucked Lane W.
aka Micro Penis mother 24 hours straight.
She was screaming, basically singing all

the arias from operas that Luciano Pavarotti
usually sings. You Lane W. aka Micro Penis
should have heard it, since you

live in the basement of your mothers house.

No, actually remarkably, I don't. According to google I live 433

miles away from her.

Strike!

See, what i said about you was spot on.

What you said about me was generic and incorrect.

You really suck, man.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Feature 2008: 4 Blades + Tesla S1070
CPU Cores 16
CPU Clock (GHz) 2.5
CPU IPC (est.) 1
CPU Throughput (units) 16 x 2.5 |u1.0=40
GPU Cores 960 (4x T10P)
GPU Clock (GHz) 1.5
GPU IPC (est.) ~1.0 (FMA)
GPU Throughput (units) 960 x 1.5x1.0=1440
Total Compute (CPU+GPU) 40 + 1440 = 1480
Memory Capacity 16-20 GB (DDR2)
Storage I/O ~400 MB/s (HDDs)
Power Consumption ~1500 W
Physical Size 8-12 RU + 1U GPU
Cost (2008 USD) ~$33,000

Feature 2026: 1 Al Laptop
CPU Cores Aug 16
CPU Clock (GHz) 4.5
CPU IPC (est.) 2
CPU Throughput (units) 16 x 4.5 |u2.0=144 (or 72 for 8c)
GPU Cores 4096
GPU Clock (GHz) ~2.0
GPU IPC (est.) ~1.5 (modern)
GPU Throughput (units) 4096 x 2.0x1.5=12288
Total Compute (CPU+GPU) 144 + 12288 = 12432 (or 72+12288 for 8c)
Memory Capacity 16-32 GB (DDR5)
Storage I/O ~7000 MB/s (NVMe)
Power Consumption ~50-100 W
Physical Size 1 laptop bag
Cost (2008 USD) ~$500-1000

Feature Winner
CPU Cores Tie
CPU Clock (GHz) Laptop (1.8x faster)
CPU IPC (est.) Laptop (2x better)
CPU Throughput (units) Laptop: 1.8-3.6x faster
GPU Cores Laptop: 4.3x more cores
GPU Clock (GHz) Laptop (1.33x faster)
GPU IPC (est.) Laptop (1.5x better)
GPU Throughput (units) Laptop: 8.5x more GPU throughput
Total Compute (CPU+GPU) Laptop: 8.4x more total compute
Memory Capacity Laptop (more, faster)
Storage I/O Laptop: 17x faster
Power Consumption Laptop: 15-30x more efficient
Physical Size Laptop
Cost (2008 USD) Laptop: 33-66x cheaper

Bye

Mild Shock schrieb:

Hi,

Now you can compare this here from 2008
with modern AI Laptops for 500-1000 USD:

Google spotlights data center inner workings https://web.archive.org/web/20131019063218/http://news.cnet.com/8301-10784_3-9955184-7.html

There is a striking similarity, only what
once occupied a rack, has now the size
of your plam, all inside one silicon chip:

- Multiple CPU cores on the same chip
- Multiple GPU units on the same chip
- Network on the same chip communication
- Crossbar caches on the same chip
- Disk controllers on the same chip
- Multi channel RAM access on the same chip

Pretty cool!

Bye

P.S.: Example such devices with iGPU:

Intel(R) Core(TM) Ultra 7 258V
AMD Ryzen AI 7 350 w/ Radeon 860M
Apple A18 Pro, Darwin Kernel Version 25.5.0
Snapdragon(R) X - X126100 - Qualcomm(R) Oryon(TM) CPU

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Yeah keep reading my posts, uninspired fool.
Seems you got a glimps of imagination from my post:

From: Mild Shock <janburse@fastmail.fm>
Subject: NVIDIA evacuated its Chinese market [Tau Scaling]
Date: Thu, 23 Jul 2026 19:13:51 +0200

How it started:

Captain: Throw the switch, Scotty!
Enterprise: Cloaking Device makes it invisible
Spock: Military secrets are the most fleeting of all.
Kirk Escapes the Romulans - The Enterprise Incident https://www.youtube.com/watch?v=AusAGjwlql8

But copying others in trope, is not the same
as jolting a trope into a conservation.
It still makes you a lame copist. Maybe you

don't know with whom you are dealing with, right?
I don't know who you are, but I will look for you,
I will find you and I will let you run my pi-WAM

on your sputnik commodore c64 with 8088.

Bye

Mild Shock schrieb:

Hi,

You are a moron, and you represent putin payed
trolls from the army of brainless troll morons.

Bye

Lane W schrieb:

Mild Shock has no idea who I am or what I represent

Mild Shock schrieb:

Hi,

If any of you guys do not understand what
is meant by or what the implications are:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Well I wouldn't care less. There are two
outcomes for numb nuts:

- Ignoramus: They don't understand it, but
-a-a they will understand it before they die.

- Ignorabimus: They don't understand it, and
-a-a will never understand it, and they die.

So who cares, its not my problem, you people
are stupid as fuck, and slow as fuck...

Bye

Mild Shock schrieb:

Hi,

Micro penis brain is in constant hiatus.
He can even not detect a trope.

LoL

Bye

Lane W schrieb:

Mild Shock wrote:

Hi,

My mother is worried that I fucked Lane W.
aka Micro Penis mother 24 hours straight.
She was screaming, basically singing all

the arias from operas that Luciano Pavarotti
usually sings. You Lane W. aka Micro Penis
should have heard it, since you

live in the basement of your mothers house.

No, actually remarkably, I don't. According to google I live 433

miles away from her.

Strike!

See, what i said about you was spot on.

What you said about me was generic and incorrect.

You really suck, man.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Again I posted everything here:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

The repo says, same time when I posted
the link first time:

This repository was archived by the
owner on Jul 9, 2026. It is now read-only.

Now a USENET user, who had already entitled
himself for a couple of irrational accusations

towards my side, is asking this question:

Chris M. Thomasson schrieb, Jul 24, 2026

Show an outline of what you
need you compute shader to do?

Bravo, thats a delay of a wooping 15 days.

Bye

Mild Shock schrieb:

Hi,

If any of you guys do not understand what
is meant by or what the implications are:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Well I wouldn't care less. There are two
outcomes for numb nuts:

- Ignoramus: They don't understand it, but
-a they will understand it before they die.

- Ignorabimus: They don't understand it, and
-a will never understand it, and they die.

So who cares, its not my problem, you people
are stupid as fuck, and slow as fuck...

Bye

Mild Shock schrieb:

Hi,

Micro penis brain is in constant hiatus.
He can even not detect a trope.

LoL

Bye

Lane W schrieb:

Mild Shock wrote:

Hi,

My mother is worried that I fucked Lane W.
aka Micro Penis mother 24 hours straight.
She was screaming, basically singing all

the arias from operas that Luciano Pavarotti
usually sings. You Lane W. aka Micro Penis
should have heard it, since you

live in the basement of your mothers house.

No, actually remarkably, I don't. According to google I live 433

miles away from her.

Strike!

See, what i said about you was spot on.

What you said about me was generic and incorrect.

You really suck, man.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Ride the snake
He's old and his skin is cold
The west is the best
The west is the best
Get here and we'll do the rest
The blue bus is calling us
The blue bus is calling us
Driver, where you taking us?

Apocalypse Now intro: The Doors, The End {1979} https://www.youtube.com/watch?v=CIrvSJwwJUE

Bye

Hi,

Again I posted everything here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

The repo says, same time when I posted
the link first time:

This repository was archived by the
owner on Jul 9, 2026. It is now read-only.

Now a USENET user, who had already entitled
himself for a couple of irrational accusations

towards my side, is asking this question:

Chris M. Thomasson schrieb, Jul 24, 2026

Show an outline of what you
need you compute shader to do?

Bravo, thats a delay of a wooping 15 days.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly. https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Its not tested on some Single Instruction/
Multiple Data (SIMD) GPU. It was only tested on
AI Laptops with Multiple instruction, Multiple

Data (GPU) architecture for the scalar registers
per logical thread. As introduced by NVIDIA Volta
in around 2017:

the first product was not announced until May 2017 https://en.wikipedia.org/wiki/Volta_%28microarchitecture%29

Although I wrote the code of Hack VM with SIMD
in mind, I never tested it on a pure SIMD GPU,
and I never ported boot.mjs or boot2.mjs to

WebGL2 / GLSL. I uploaded WebGPU / WGSL. Among the
tester I had were these AI Laptops, that could all
run WebGPU / WGSL in a browser:

Intel(R) Core(TM) Ultra 7 258V
AMD Ryzen AI 7 350 w/ Radeon 860M
Apple A18 Pro, Darwin Kernel Version 25.5.0
Snapdragon(R) X - X126100 - Qualcomm(R) Oryon(TM) CPU

Some AI Laptops had WebGPU / WGSL still behind
a browser flag, since its relatively new on ARM.
Also the above AI Laptops have all a iGPU and

not a separate GPU card.

Bye

Mild Shock schrieb:

Hi,

Ride the snake
He's old and his skin is cold
The west is the best
The west is the best
Get here and we'll do the rest
The blue bus is calling us
The blue bus is calling us
Driver, where you taking us?

Apocalypse Now intro: The Doors, The End {1979} https://www.youtube.com/watch?v=CIrvSJwwJUE

Bye

Hi,

Again I posted everything here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

The repo says, same time when I posted
the link first time:

This repository was archived by the
owner on Jul 9, 2026. It is now read-only.

Now a USENET user, who had already entitled
himself for a couple of irrational accusations

towards my side, is asking this question:

Chris M. Thomasson schrieb, Jul 24, 2026

Show an outline of what you
need you compute shader to do?

Bravo, thats a delay of a wooping 15 days.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

It could one take 3-4 months to find a suitable
machine and suitable browser, so that MIMD is
utilized, especially if you have:

A Sputnik Commodore C64 with 8088
from the basement of your mother

But maybe somebody figures out it also runs on
SIMD. Its not in my focus to test the SIMD
platform, I do not intend to go back in time

to 2008, and a Tesla S1070:

Tesla S1070 was an professional graphics card by NVIDIA
Its price at launch was 7999 US Dollars. https://www.techpowerup.com/gpu-specs/tesla-s1070.c1540

Also not in 2026, a RTX 5090:

NVIDIA-< RTXrao 5090 most powerful GeForce GPU ever made
With Boost Clock Speed its at 3779 USD https://marketplace.nvidia.com/en-us/consumer/graphics-cards/?locale=en-us&page=1&limit=15&gpu=RTX+5090&has_offer=is_bestselling

The title of the experiment is really Budget Laptop.
What is a litte unspoken in the title, that the Laptop
is an AI Laptop. But you see it in the description:

11.4 Giga Lips with a Budget Laptop
At the end of 2025 we acquired a couple of AI Laptops https://github.com/Jean-Luc-Picard-2021/gigabudget

These AI Laptops are quite affordable ,
500 USD to 1000 USD.

Bye

Mild Shock schrieb:

Hi,

Its not tested on some Single Instruction/
Multiple Data (SIMD) GPU. It was only tested on
AI Laptops with Multiple instruction, Multiple

Data (GPU) architecture for the scalar registers
per logical thread. As introduced by NVIDIA Volta
in around 2017:

the first product was not announced until May 2017 https://en.wikipedia.org/wiki/Volta_%28microarchitecture%29

Although I wrote the code of Hack VM with SIMD
in mind, I never tested it on a pure SIMD GPU,
and I never ported boot.mjs or boot2.mjs to

WebGL2 / GLSL. I uploaded WebGPU / WGSL. Among the
tester I had were these AI Laptops, that could all
run WebGPU / WGSL in a browser:

Intel(R) Core(TM) Ultra 7 258V
AMD Ryzen AI 7 350 w/ Radeon 860M
Apple A18 Pro, Darwin Kernel Version 25.5.0
Snapdragon(R) X - X126100 - Qualcomm(R) Oryon(TM) CPU

Some AI Laptops had WebGPU / WGSL still behind
a browser flag, since its relatively new on ARM.
Also the above AI Laptops have all a iGPU and

not a separate GPU card.

Bye

Mild Shock schrieb:

Hi,

Ride the snake
He's old and his skin is cold
The west is the best
The west is the best
Get here and we'll do the rest
The blue bus is calling us
The blue bus is calling us
Driver, where you taking us?

Apocalypse Now intro: The Doors, The End {1979}
https://www.youtube.com/watch?v=CIrvSJwwJUE

Bye

Hi,

Again I posted everything here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

The repo says, same time when I posted
the link first time:

This repository was archived by the
owner on Jul 9, 2026. It is now read-only.

Now a USENET user, who had already entitled
himself for a couple of irrational accusations

towards my side, is asking this question:

Chris M. Thomasson schrieb, Jul 24, 2026

Show an outline of what you
need you compute shader to do?

Bravo, thats a delay of a wooping 15 days.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

The primary motivation behind SpaceX's acquisition
of Cursor (and its underlying Composer model tech stack)
comes down to vertical integration of the entire AI
stackrCocompute, models, and application surface.

Bridging the Coding & Reasoning Gap: While xAIrCOs Grok
had massive raw compute backing it (via massive clusters
like Colossus), its training data heavily leaned on
social media feeds from X.

That made it great for conversational chat, but left
it lagging behind competitors (like OpenAI and Anthropic)
in pure coding, reasoning efficiency, and long-horizon
agentic execution.

Owning the Developer Workflow: Instead of just building
a model and hoping people use it, acquiring Cursor
gives SpaceX the premier developer-facing application
layer that millions of developers already use
for daily coding.

Coupling Cursor's model architecture with SpaceX/xAI's
massive compute infrastructure allows them to aggressively
scale up training for next-gen models.

So its a Win-Win. Got it?

Bye

Mild Shock schrieb:

Hi,

Because I use WebGPU and not WebGL. And
because WebGPU can adresss modern GPU
developed with the NVIDIA Volta evolution,

which happened in 2017. Namley that compute
shaders are not any more subject to the
realization restriction of lock step

execution, but have independent thread state.
And because there is independent thread state
there is also independent time spent for a

a work item by each logical thread, if the
submitted logical thread uses a lot of branching
logic or even loops. But the use of branching

and loops is encouraged in independent thread
state programming of compute shaders. The variables
that can drive such logic are the scalar variables:

Tour of WGSL - Control Flow https://google.github.io/tour-of-wgsl/control-flow/

Then not to waste GPU compute time, by logical
threads doing nothing. You will need to
introduce some load balancing among multiple

logical threads. And MPMC queues are one way to
readize load balancing. Compute shaders with
producer and consumer entry points are proposed

as fundamental architecture by Thunder Kittens:

ThunderKittens: Simple, Fast, and Adorable AI Kernels https://arxiv.org/abs/2410.20399

They are used by this SpaceX acquisition:

Composer 2 Technical Report
https://arxiv.org/abs/2603.24477

Thunder Kittens uses Hardware support, i.e. tma_expect().

Bye

Chris M. Thomasson schrieb:

never meant to be used in a GPU.
Dmitry CAS version can be used, but

Why do you even need a mpmc queue
in your compute shader anyway?

Mild Shock schrieb:

Hi,

This is quite fun, how some TLA+ guy fears
the full state of queue like the devil in
itself. But I guess if a service rate is

low and the producer has not much to do to
produce its work items, the arrival rate
has nevertheless to adapt, and dealing

with "full states", which are wrongly
called deadlock here, is the normal:

Tutorial-style talk - BlockingQueue
https://github.com/lemmy/BlockingQueue/tree/main

Prolog is in good position. The bird box
model has a redo port. So sometimes switching
from push to pull, can help without doing

Deadlock Exorcism. You can also translate
the bird box ports into pi-calculus:

A pi-calculus Specification of Prolog
Benjamin Z. Li - University of Pennsylvania
11 Apr 1994, European Symposium on Programming,
Prolog, Unification, Backtracking
https://scispace.com/pdf/a-pi-calculus-specification-of-prolog-3qf2pf04ud.pdf


Have Fun!

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Why I forget something, that was never
on my mind. This here is hardly
about rockets:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

So stop glue sniffing. Got it? Your are
just confused. I don't care about SpaceX,
that Composer was acquired by

SpaceX was just a factual thing.

Bye

Disclaimer: Who knows, maybe somebody picks
up pi-calculus and/or WAM for rocket engineering,
I do not primarily exclude it.

But its not on my mind, so I cannot forget something
which I don't care about.

Ross Finlayson schrieb:

Forget Space-X and forget that heil-throwing fat-ass, too,
the rockets they bought are wearing out and the ones they
built are blowing up, moving fast and breaking things.

Mild Shock schrieb:

Hi,

The primary motivation behind SpaceX's acquisition
of Cursor (and its underlying Composer model tech stack)
comes down to vertical integration of the entire AI
stackrCocompute, models, and application surface.

Bridging the Coding & Reasoning Gap: While xAIrCOs Grok
had massive raw compute backing it (via massive clusters
like Colossus), its training data heavily leaned on
social media feeds from X.

That made it great for conversational chat, but left
it lagging behind competitors (like OpenAI and Anthropic)
in pure coding, reasoning efficiency, and long-horizon
agentic execution.

Owning the Developer Workflow: Instead of just building
a model and hoping people use it, acquiring Cursor
gives SpaceX the premier developer-facing application
layer that millions of developers already use
for daily coding.

Coupling Cursor's model architecture with SpaceX/xAI's
massive compute infrastructure allows them to aggressively
scale up training for next-gen models.

So its a Win-Win. Got it?

Bye

Mild Shock schrieb:

Hi,

Because I use WebGPU and not WebGL. And
because WebGPU can adresss modern GPU
developed with the NVIDIA Volta evolution,

which happened in 2017. Namley that compute
shaders are not any more subject to the
realization restriction of lock step

execution, but have independent thread state.
And because there is independent thread state
there is also independent time spent for a

a work item by each logical thread, if the
submitted logical thread uses a lot of branching
logic or even loops. But the use of branching

and loops is encouraged in independent thread
state programming of compute shaders. The variables
that can drive such logic are the scalar variables:

Tour of WGSL - Control Flow
https://google.github.io/tour-of-wgsl/control-flow/

Then not to waste GPU compute time, by logical
threads doing nothing. You will need to
introduce some load balancing among multiple

logical threads. And MPMC queues are one way to
readize load balancing. Compute shaders with
producer and consumer entry points are proposed

as fundamental architecture by Thunder Kittens:

ThunderKittens: Simple, Fast, and Adorable AI Kernels
https://arxiv.org/abs/2410.20399

They are used by this SpaceX acquisition:

Composer 2 Technical Report
https://arxiv.org/abs/2603.24477

Thunder Kittens uses Hardware support, i.e. tma_expect().

Bye

Chris M. Thomasson schrieb:

never meant to be used in a GPU.
Dmitry CAS version can be used, but

Why do you even need a mpmc queue
in your compute shader anyway?

Mild Shock schrieb:

Hi,

This is quite fun, how some TLA+ guy fears
the full state of queue like the devil in
itself. But I guess if a service rate is

low and the producer has not much to do to
produce its work items, the arrival rate
has nevertheless to adapt, and dealing

with "full states", which are wrongly
called deadlock here, is the normal:

Tutorial-style talk - BlockingQueue
https://github.com/lemmy/BlockingQueue/tree/main

Prolog is in good position. The bird box
model has a redo port. So sometimes switching
from push to pull, can help without doing

Deadlock Exorcism. You can also translate
the bird box ports into pi-calculus:

A pi-calculus Specification of Prolog
Benjamin Z. Li - University of Pennsylvania
11 Apr 1994, European Symposium on Programming,
Prolog, Unification, Backtracking
https://scispace.com/pdf/a-pi-calculus-specification-of-prolog-3qf2pf04ud.pdf


Have Fun!

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Mostlikely this is faster with MIMD,
than with SIMD. Draw a Mandel Brot Figure:

mandelbrot
https://compute.toys/view/213

Why, because each pixel has a different
result i, since this loop has a break:

for (i = 0; i < count; i++) {
if (z_re * z_re + z_im * z_im > 4.0) {
break;
}
let new_re = z_re * z_re - z_im * z_im;
let new_im = 2.0f * z_re * z_im;
z_re = c_re + new_re;
z_im = c_im + new_im;
}

So a sheduler balancer, that uses independent
thread state (MIMD), from a post NVIDIA Volta
type GPU, can squeeze out more computation,

than a lock step (SIMD) scheduler, from a
pre NVIDIA Volta GPU. I guess I will use that
as a balancing example for pi-WAM.

Bye

Mild Shock schrieb:

Hi,

Because I use WebGPU and not WebGL. And
because WebGPU can adresss modern GPU
developed with the NVIDIA Volta evolution,

which happened in 2017. Namley that compute
shaders are not any more subject to the
realization restriction of lock step

execution, but have independent thread state.
And because there is independent thread state
there is also independent time spent for a

a work item by each logical thread, if the
submitted logical thread uses a lot of branching
logic or even loops. But the use of branching

and loops is encouraged in independent thread
state programming of compute shaders. The variables
that can drive such logic are the scalar variables:

Tour of WGSL - Control Flow https://google.github.io/tour-of-wgsl/control-flow/

Then not to waste GPU compute time, by logical
threads doing nothing. You will need to
introduce some load balancing among multiple

logical threads. And MPMC queues are one way to
readize load balancing. Compute shaders with
producer and consumer entry points are proposed

as fundamental architecture by Thunder Kittens:

ThunderKittens: Simple, Fast, and Adorable AI Kernels https://arxiv.org/abs/2410.20399

They are used by this SpaceX acquisition:

Composer 2 Technical Report
https://arxiv.org/abs/2603.24477

Thunder Kittens uses Hardware support, i.e. tma_expect().

Bye

Chris M. Thomasson schrieb:

never meant to be used in a GPU.
Dmitry CAS version can be used, but

Why do you even need a mpmc queue
in your compute shader anyway?

Mild Shock schrieb:

Hi,

This is quite fun, how some TLA+ guy fears
the full state of queue like the devil in
itself. But I guess if a service rate is

low and the producer has not much to do to
produce its work items, the arrival rate
has nevertheless to adapt, and dealing

with "full states", which are wrongly
called deadlock here, is the normal:

Tutorial-style talk - BlockingQueue
https://github.com/lemmy/BlockingQueue/tree/main

Prolog is in good position. The bird box
model has a redo port. So sometimes switching
from push to pull, can help without doing

Deadlock Exorcism. You can also translate
the bird box ports into pi-calculus:

A pi-calculus Specification of Prolog
Benjamin Z. Li - University of Pennsylvania
11 Apr 1994, European Symposium on Programming,
Prolog, Unification, Backtracking
https://scispace.com/pdf/a-pi-calculus-specification-of-prolog-3qf2pf04ud.pdf


Have Fun!

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

You see it all boils down to find your inner peace
by an immaculate inception of some queue datatype.

KOAN/Fortran-S was an early 1990s research programming
system for distributed-memory multiprocessors . Developed
at ENS Lyon in the early 1990s . Often listed alongside
other historical parallel programming efforts.

The Message Passing: The research explicitly
compared the SVM approach against message passing
on the same hardware . The finding was that SVM
could achieve good performance without the low-level

complexity of managing explicit messages, though
the best results often came from a hybrid approach (sic!)
Here is an interesting baseline, from Java,
a class ElevenSingle that only does:

public static void run() {
for (int A = 1; A < 192; A++) {
int Y = (771-A)/3;
for (int B = A; B < Y; B++) {
int Z = (771-A-B)/2;
for (int C = B; C < Z; C++) {
int D = 711-A-B-C;
if (A*B*C == 711000000/D &&
711000000 % D == 0)
System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
}
}
}
}

And then compare it to ElevenMulti, doing some
Work Balancing Scheduler Tetris Game with 8 cores:

ElevenSingle
A=120, B=125, C=150, D=316
6.628 ms

ElevenMulti
A=120, B=125, C=150, D=316
1.941 ms

Not great, not terrible!

Bye

Mild Shock schrieb:

Hi,

Its not tested on some Single Instruction/
Multiple Data (SIMD) GPU. It was only tested on
AI Laptops with Multiple instruction, Multiple

Data (GPU) architecture for the scalar registers
per logical thread. As introduced by NVIDIA Volta
in around 2017:

the first product was not announced until May 2017 https://en.wikipedia.org/wiki/Volta_%28microarchitecture%29

Although I wrote the code of Hack VM with SIMD
in mind, I never tested it on a pure SIMD GPU,
and I never ported boot.mjs or boot2.mjs to

WebGL2 / GLSL. I uploaded WebGPU / WGSL. Among the
tester I had were these AI Laptops, that could all
run WebGPU / WGSL in a browser:

Intel(R) Core(TM) Ultra 7 258V
AMD Ryzen AI 7 350 w/ Radeon 860M
Apple A18 Pro, Darwin Kernel Version 25.5.0
Snapdragon(R) X - X126100 - Qualcomm(R) Oryon(TM) CPU

Some AI Laptops had WebGPU / WGSL still behind
a browser flag, since its relatively new on ARM.
Also the above AI Laptops have all a iGPU and

not a separate GPU card.

Bye

Mild Shock schrieb:

Hi,

Ride the snake
He's old and his skin is cold
The west is the best
The west is the best
Get here and we'll do the rest
The blue bus is calling us
The blue bus is calling us
Driver, where you taking us?

Apocalypse Now intro: The Doors, The End {1979}
https://www.youtube.com/watch?v=CIrvSJwwJUE

Bye

Hi,

Again I posted everything here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

The repo says, same time when I posted
the link first time:

This repository was archived by the
owner on Jul 9, 2026. It is now read-only.

Now a USENET user, who had already entitled
himself for a couple of irrational accusations

towards my side, is asking this question:

Chris M. Thomasson schrieb, Jul 24, 2026

Show an outline of what you
need you compute shader to do?

Bravo, thats a delay of a wooping 15 days.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Mostlikely we see the turbo capping of certain
CPU designs, that have turbo mode when CPU
is used with mostly only one thread active,

but throttles down when more threads are active.
This has even resulted in designs with performance
cores and economy cores.

So the factor below for 8 cores is only:

6.628 ms / 1.941 ms = 3.415

But if you discount for turbo versus non-turbo,
i.e. 5 GHz versus 3 GHz or so. You see that
the machine was not utilized very badly:

3.415 * 5 / 3 = 5.691

The class ElevenMulti does use 6 workers,
and 1 producer and 1 consumer.

Bye

Mild Shock schrieb:

Hi,

You see it all boils down to find your inner peace
by an immaculate inception of some queue datatype.

KOAN/Fortran-S was an early 1990s research programming
system for distributed-memory multiprocessors . Developed
at ENS Lyon in the early 1990s . Often listed alongside
other historical parallel programming efforts.

The Message Passing: The research explicitly
compared the SVM approach against message passing
on the same hardware . The finding was that SVM
could achieve good performance without the low-level

complexity of managing explicit messages, though
the best results often came from a hybrid approach (sic!)
Here is an interesting baseline, from Java,
a class ElevenSingle that only does:

-a-a-a public static void run() {
-a-a-a-a-a-a-a for (int A = 1; A < 192; A++) {
-a-a-a-a-a-a-a-a-a-a-a int Y = (771-A)/3;
-a-a-a-a-a-a-a-a-a-a-a for (int B = A; B < Y; B++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int Z = (771-A-B)/2;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a for (int C = B; C < Z; C++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int D = 711-A-B-C;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a if (A*B*C == 711000000/D &&
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a 711000000 % D == 0)
-a-a-a System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a }
-a-a-a }

And then compare it to ElevenMulti, doing some
Work Balancing Scheduler Tetris Game with 8 cores:

ElevenSingle
A=120, B=125, C=150, D=316
6.628 ms

ElevenMulti
A=120, B=125, C=150, D=316
1.941 ms

Not great, not terrible!

Bye

Mild Shock schrieb:

Hi,

Its not tested on some Single Instruction/
Multiple Data (SIMD) GPU. It was only tested on
AI Laptops with Multiple instruction, Multiple

Data (GPU) architecture for the scalar registers
per logical thread. As introduced by NVIDIA Volta
in around 2017:

the first product was not announced until May 2017
https://en.wikipedia.org/wiki/Volta_%28microarchitecture%29

Although I wrote the code of Hack VM with SIMD
in mind, I never tested it on a pure SIMD GPU,
and I never ported boot.mjs or boot2.mjs to

WebGL2 / GLSL. I uploaded WebGPU / WGSL. Among the
tester I had were these AI Laptops, that could all
run WebGPU / WGSL in a browser:

Intel(R) Core(TM) Ultra 7 258V
AMD Ryzen AI 7 350 w/ Radeon 860M
Apple A18 Pro, Darwin Kernel Version 25.5.0
Snapdragon(R) X - X126100 - Qualcomm(R) Oryon(TM) CPU

Some AI Laptops had WebGPU / WGSL still behind
a browser flag, since its relatively new on ARM.
Also the above AI Laptops have all a iGPU and

not a separate GPU card.

Bye

Mild Shock schrieb:

Hi,

Ride the snake
He's old and his skin is cold
The west is the best
The west is the best
Get here and we'll do the rest
The blue bus is calling us
The blue bus is calling us
Driver, where you taking us?

Apocalypse Now intro: The Doors, The End {1979}
https://www.youtube.com/watch?v=CIrvSJwwJUE

Bye

Hi,

Again I posted everything here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

The repo says, same time when I posted
the link first time:

This repository was archived by the
owner on Jul 9, 2026. It is now read-only.

Now a USENET user, who had already entitled
himself for a couple of irrational accusations

towards my side, is asking this question:

Chris M. Thomasson schrieb, Jul 24, 2026

Show an outline of what you
need you compute shader to do?

Bravo, thats a delay of a wooping 15 days.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Whats this "forget" trope of glue sniffing
Rossy Boy with his herpes blisters?

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

Why should I forget Bulgarians,
they are never on my mind. Do you
see me doing ggml stuff?

I only hypothesized that it is
over for Python as the machine
learning language or AI inferencing

locally on AI laptops language, and
made the ggml case, so I already forgot
about them. Which might give you a glimps,

why WebGPU was used for this here:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Is an interesting choice. Even
github has some Languages statistics,
giving an account what I used:

HTML 67.5% JavaScript 23.1% CSS 9.4%

Have Fun!

Bye

P.S.: The example below is not p-adics,
you complete imbecil moron. Its just:

7-11 cubic Solution by Pritchard & Gries https://www.cs.cornell.edu/gries/TechReports/83-574.pdf

Ross Finlayson schrieb:

On 07/26/2026 10:52 AM, Mild Shock wrote:

Hi,

You see it all boils down to find your inner peace
by an immaculate inception of some queue datatype.

KOAN/Fortran-S was an early 1990s research programming
system for distributed-memory multiprocessors . Developed
at ENS Lyon in the early 1990s . Often listed alongside
other historical parallel programming efforts.

The Message Passing: The research explicitly
compared the SVM approach against message passing
on the same hardware . The finding was that SVM
could achieve good performance without the low-level

complexity of managing explicit messages, though
the best results often came from a hybrid approach (sic!)
Here is an interesting baseline, from Java,
a class ElevenSingle that only does:

public static void run() {
for (int A = 1; A < 192; A++) {
int Y = (771-A)/3;
for (int B = A; B < Y; B++) {
int Z = (771-A-B)/2;
for (int C = B; C < Z; C++) {
int D = 711-A-B-C;
if (A*B*C == 711000000/D &&
711000000 % D == 0)
System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
}
}
}
}

And then compare it to ElevenMulti, doing some
Work Balancing Scheduler Tetris Game with 8 cores:

ElevenSingle
A=120, B=125, C=150, D=316
6.628 ms

ElevenMulti
A=120, B=125, C=150, D=316
1.941 ms

Not great, not terrible!

Bye

Oh, that's just "tricks of p-adic arithmetic".

Like other sock-puppet howler trolls, when confronted
with its base incredulity, it will descend to its
lower levers of the pathos variety.

You might be happier learning about Julia trees and
raster ops, instead of shilling yet another Ramanujan
series without saying how it's made.

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Whats this "forget" trope of glue sniffing
Rossy Boy with his herpes blisters?

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

Why should I forget Bulgarians,
they are never on my mind. Do you
see me doing ggml stuff?

I only hypothesized that it is
over for Python as the machine
learning language or AI inferencing

locally on AI laptops language, and
made the ggml case, so I already forgot
about them. Which might give you a glimps,

why WebGPU was used for this here:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Is an interesting choice. Even
github has some Languages statistics,
giving an account what I used:

HTML 67.5% JavaScript 23.1% CSS 9.4%

Have Fun!

Bye

P.S.: The example below is not p-adics,
you complete imbecil moron. Its just:

7-11 cubic Solution by Pritchard & Gries https://www.cs.cornell.edu/gries/TechReports/83-574.pdf

Ross Finlayson schrieb:

On 07/26/2026 10:52 AM, Mild Shock wrote:

Hi,

You see it all boils down to find your inner peace
by an immaculate inception of some queue datatype.

KOAN/Fortran-S was an early 1990s research programming
system for distributed-memory multiprocessors . Developed
at ENS Lyon in the early 1990s . Often listed alongside
other historical parallel programming efforts.

The Message Passing: The research explicitly
compared the SVM approach against message passing
on the same hardware . The finding was that SVM
could achieve good performance without the low-level

complexity of managing explicit messages, though
the best results often came from a hybrid approach (sic!)
Here is an interesting baseline, from Java,
a class ElevenSingle that only does:

public static void run() {
for (int A = 1; A < 192; A++) {
int Y = (771-A)/3;
for (int B = A; B < Y; B++) {
int Z = (771-A-B)/2;
for (int C = B; C < Z; C++) {
int D = 711-A-B-C;
if (A*B*C == 711000000/D &&
711000000 % D == 0)
System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
}
}
}
}

And then compare it to ElevenMulti, doing some
Work Balancing Scheduler Tetris Game with 8 cores:

ElevenSingle
A=120, B=125, C=150, D=316
6.628 ms

ElevenMulti
A=120, B=125, C=150, D=316
1.941 ms

Not great, not terrible!

Bye

Oh, that's just "tricks of p-adic arithmetic".

Like other sock-puppet howler trolls, when confronted
with its base incredulity, it will descend to its
lower levers of the pathos variety.

You might be happier learning about Julia trees and
raster ops, instead of shilling yet another Ramanujan
series without saying how it's made.

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Some counter PyTorch Python trends are
for example OpenAIs Triton. And the variant
miniTriton CUDA vibe produced by Kimi K3 (sic!):

"We further tested whether Kimi K3 could build
a GPU programming system from scratch. Kimi K3
developed MiniTriton, a compact Triton-like
compiler with its own tile-level IR layer over
MLIR, optimization passes, and a PTX code-
generation pipeline.

Across supported roofline benchmarks, MiniTriton
delivers performance on par with or better than
Triton and torch.compile rCo beating Triton on
certain workloads. Beyond microbenchmarks,
MiniTriton sustains end-to-end nanoGPT training
with stable convergence, the loss curve

closely tracking the reference with only minor
divergence rCo validating the full pipeline on a
realistic workload. These results demonstrate
that Kimi K3 can build a coherent end-to-end
compiler rCo from DSL frontend and IR passes to
PTX codegen and runtime rCo rather than isolated

kernels; its from-scratch Tensor Core path
already rivals TritonrCOs extensively optimized stack."

GPU Compiler Development
https://www.kimi.com/blog/kimi-k3

Although many GPU corporate stuff is anonymized,
and some AI papers have lists of 30 authors. Here
nanoGPT is mentioned which is tied to the name

Andrej Karpathy. See also here:

Update Nov 2025 nanoGPT has a new and
improved cousin called nanochat.
https://github.com/karpathy/nanogpt

But as can be seen, he moved on to another project.

Bye

Mild Shock schrieb:

Hi,

Whats this "forget" trope of glue sniffing
Rossy Boy with his herpes blisters?

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

Why should I forget Bulgarians,
they are never on my mind. Do you
see me doing ggml stuff?

I only hypothesized that it is
over for Python as the machine
learning language or AI inferencing

locally on AI laptops language, and
made the ggml case, so I already forgot
about them. Which might give you a glimps,

why WebGPU was used for this here:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Is an interesting choice. Even
github has some Languages statistics,
giving an account what I used:

HTML 67.5% JavaScript 23.1% CSS 9.4%

Have Fun!

Bye

P.S.: The example below is not p-adics,
you complete imbecil moron. Its just:

7-11 cubic Solution by Pritchard & Gries https://www.cs.cornell.edu/gries/TechReports/83-574.pdf

Ross Finlayson schrieb:

On 07/26/2026 10:52 AM, Mild Shock wrote:

Hi,

You see it all boils down to find your inner peace
by an immaculate inception of some queue datatype.

KOAN/Fortran-S was an early 1990s research programming
system for distributed-memory multiprocessors . Developed
at ENS Lyon in the early 1990s . Often listed alongside
other historical parallel programming efforts.

The Message Passing: The research explicitly
compared the SVM approach against message passing
on the same hardware . The finding was that SVM
could achieve good performance without the low-level

complexity of managing explicit messages, though
the best results often came from a hybrid approach (sic!)
Here is an interesting baseline, from Java,
a class ElevenSingle that only does:

-a-a-a-a-a public static void run() {
-a-a-a-a-a-a-a-a-a for (int A = 1; A < 192; A++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a int Y = (771-A)/3;
-a-a-a-a-a-a-a-a-a-a-a-a-a for (int B = A; B < Y; B++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int Z = (771-A-B)/2;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a for (int C = B; C < Z; C++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int D = 711-A-B-C;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a if (A*B*C == 711000000/D &&
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a 711000000 % D == 0)
-a-a-a-a-a System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a }

And then compare it to ElevenMulti, doing some
Work Balancing Scheduler Tetris Game with 8 cores:

ElevenSingle
A=120, B=125, C=150, D=316
6.628 ms

ElevenMulti
A=120, B=125, C=150, D=316
1.941 ms

Not great, not terrible!

Bye

Oh, that's just "tricks of p-adic arithmetic".

Like other sock-puppet howler trolls, when confronted
with its base incredulity, it will descend to its
lower levers of the pathos variety.

You might be happier learning about Julia trees and
raster ops, instead of shilling yet another Ramanujan
series without saying how it's made.

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Andrej Karpathy was bascially the original gangster
of doing not only AI inferencing but also AI
learning on a Budget Laptop. The nanoGPT project

states the following:

"I only have a macbook (or other cheap
computer). No worries, we can still train a
GPT but we want to dial things down a notch.
I recommend getting the bleeding edge PyTorch
nightly (select it here when installing) as
it is currently quite likely to make your
code more efficient."
https://github.com/karpathy/nanogpt

But meanwhile he has moved to a higher price
segment. Not sure whether he will climbe
down to a lower price segment again:

For example, you can train your own GPT-2
capability LLM (which cost ~$43,000 to train in
2019) for only $48 (~2 hours of 8XH100 GPU node)
and then talk to it over a simple CLI. On a spot
instance, the total cost can be closer to ~$15. https://github.com/karpathy/nanochat

Bt he taps into the model to rent GPU which
is available with prices in the range of 1-2 $
per hour. Even in Switzerland one can do that,

for example using the provider Exoscale. Since
he rents a cluster of 8 cards of type H100, this
explains his training price still in the 2 digit range.

Bye

P.S.: I could also do my experiment here with
rented GPU cards, and then draw a comparison
from budget laptop to the rented GPU time market:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

But testing rented GPU is not high priority.

Mild Shock schrieb:

Hi,

Some counter PyTorch Python trends are
for example OpenAIs Triton. And the variant
miniTriton CUDA vibe produced by Kimi K3 (sic!):

"We further tested whether Kimi K3 could build
a GPU programming system from scratch. Kimi K3
developed MiniTriton, a compact Triton-like
compiler with its own tile-level IR layer over
MLIR, optimization passes, and a PTX code-
generation pipeline.

Across supported roofline benchmarks, MiniTriton
delivers performance on par with or better than
Triton and torch.compile rCo beating Triton on
certain workloads. Beyond microbenchmarks,
MiniTriton sustains end-to-end nanoGPT training
with stable convergence, the loss curve

closely tracking the reference with only minor
divergence rCo validating the full pipeline on a
realistic workload. These results demonstrate
that Kimi K3 can build a coherent end-to-end
compiler rCo from DSL frontend and IR passes to
PTX codegen and runtime rCo rather than isolated

kernels; its from-scratch Tensor Core path
already rivals TritonrCOs extensively optimized stack."

GPU Compiler Development
https://www.kimi.com/blog/kimi-k3

Although many GPU corporate stuff is anonymized,
and some AI papers have lists of 30 authors. Here
nanoGPT is mentioned which is tied to the name

Andrej Karpathy. See also here:

Update Nov 2025 nanoGPT has a new and
improved cousin called nanochat.
https://github.com/karpathy/nanogpt

But as can be seen, he moved on to another project.

Bye

Mild Shock schrieb:

Hi,

Whats this "forget" trope of glue sniffing
Rossy Boy with his herpes blisters?

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

Why should I forget Bulgarians,
they are never on my mind. Do you
see me doing ggml stuff?

I only hypothesized that it is
over for Python as the machine
learning language or AI inferencing

locally on AI laptops language, and
made the ggml case, so I already forgot
about them. Which might give you a glimps,

why WebGPU was used for this here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Is an interesting choice. Even
github has some Languages statistics,
giving an account what I used:

HTML 67.5% JavaScript 23.1% CSS 9.4%

Have Fun!

Bye

P.S.: The example below is not p-adics,
you complete imbecil moron. Its just:

7-11 cubic Solution by Pritchard & Gries
https://www.cs.cornell.edu/gries/TechReports/83-574.pdf

Ross Finlayson schrieb:

On 07/26/2026 10:52 AM, Mild Shock wrote:

Hi,

You see it all boils down to find your inner peace
by an immaculate inception of some queue datatype.

KOAN/Fortran-S was an early 1990s research programming
system for distributed-memory multiprocessors . Developed
at ENS Lyon in the early 1990s . Often listed alongside
other historical parallel programming efforts.

The Message Passing: The research explicitly
compared the SVM approach against message passing
on the same hardware . The finding was that SVM
could achieve good performance without the low-level

complexity of managing explicit messages, though
the best results often came from a hybrid approach (sic!)
Here is an interesting baseline, from Java,
a class ElevenSingle that only does:

-a-a-a-a-a public static void run() {
-a-a-a-a-a-a-a-a-a for (int A = 1; A < 192; A++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a int Y = (771-A)/3;
-a-a-a-a-a-a-a-a-a-a-a-a-a for (int B = A; B < Y; B++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int Z = (771-A-B)/2;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a for (int C = B; C < Z; C++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int D = 711-A-B-C;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a if (A*B*C == 711000000/D && >> -a>>-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a 711000000 % D == 0)
-a-a-a-a-a System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a }

And then compare it to ElevenMulti, doing some
Work Balancing Scheduler Tetris Game with 8 cores:

ElevenSingle
A=120, B=125, C=150, D=316
6.628 ms

ElevenMulti
A=120, B=125, C=150, D=316
1.941 ms

Not great, not terrible!

Bye

Oh, that's just "tricks of p-adic arithmetic".

Like other sock-puppet howler trolls, when confronted
with its base incredulity, it will descend to its
lower levers of the pathos variety.

You might be happier learning about Julia trees and
raster ops, instead of shilling yet another Ramanujan
series without saying how it's made.

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

While Huggingfaces hired GG in 2026,
AK was hired by Anthropic in 2026:

Andrej Karpathy (born 23 October 1986[3])
is a Slovak-Canadian AI researcher, who
co-founded and formerly worked at OpenAI
In 2026 he joined Anthropic as part of
the pretraining team.
https://en.wikipedia.org/wiki/Andrej_Karpathy

But his nanochat archivement has an
interesting time line:

168 hours , Original OpenAI GPT-2 checkpoint, 2019
3 hours , d24 baseline, slightly overtrained, Jan 29 2026
1 1/2 hour, autoresearch round 2, Mar 14 2026
The best ChatGPT that $100 can buy.
https://github.com/karpathy/nanochat

But what hardware was the enabler. What is the
NVIDIA H100 GPU even. Well the thingy is surely not
a Budget Laptop, performance pretty much

dependence on data elememt size, the H100 NVL
version (*), and when using tensor operations,
and not only scalar operations:

8-bit towards 3000 tera flops
16-bit towards 1500 tera flops
32-bit towards 900 tera flops

Cool! I guess this experiment would tap into 60
tera flops, since it only uses scalar operations so far:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

You could perform it by migration the web application
using WebGPU into a node.js standalone application
using the dawn library for GPU access.

Bye

(*) https://resources.nvidia.com/en-us-hopper-architecture/nvidia-tensor-core-gpu-datasheet


Mild Shock schrieb:

Hi,

Whats this "forget" trope of glue sniffing
Rossy Boy with his herpes blisters?

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

Why should I forget Bulgarians,
they are never on my mind. Do you
see me doing ggml stuff?

I only hypothesized that it is
over for Python as the machine
learning language or AI inferencing

locally on AI laptops language, and
made the ggml case, so I already forgot
about them. Which might give you a glimps,

why WebGPU was used for this here:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Is an interesting choice. Even
github has some Languages statistics,
giving an account what I used:

HTML 67.5% JavaScript 23.1% CSS 9.4%

Have Fun!

Bye

P.S.: The example below is not p-adics,
you complete imbecil moron. Its just:

7-11 cubic Solution by Pritchard & Gries https://www.cs.cornell.edu/gries/TechReports/83-574.pdf

Ross Finlayson schrieb:

On 07/26/2026 10:52 AM, Mild Shock wrote:

Hi,

You see it all boils down to find your inner peace
by an immaculate inception of some queue datatype.

KOAN/Fortran-S was an early 1990s research programming
system for distributed-memory multiprocessors . Developed
at ENS Lyon in the early 1990s . Often listed alongside
other historical parallel programming efforts.

The Message Passing: The research explicitly
compared the SVM approach against message passing
on the same hardware . The finding was that SVM
could achieve good performance without the low-level

complexity of managing explicit messages, though
the best results often came from a hybrid approach (sic!)
Here is an interesting baseline, from Java,
a class ElevenSingle that only does:

-a-a-a-a-a public static void run() {
-a-a-a-a-a-a-a-a-a for (int A = 1; A < 192; A++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a int Y = (771-A)/3;
-a-a-a-a-a-a-a-a-a-a-a-a-a for (int B = A; B < Y; B++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int Z = (771-A-B)/2;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a for (int C = B; C < Z; C++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int D = 711-A-B-C;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a if (A*B*C == 711000000/D &&
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a 711000000 % D == 0)
-a-a-a-a-a System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a }

And then compare it to ElevenMulti, doing some
Work Balancing Scheduler Tetris Game with 8 cores:

ElevenSingle
A=120, B=125, C=150, D=316
6.628 ms

ElevenMulti
A=120, B=125, C=150, D=316
1.941 ms

Not great, not terrible!

Bye

Oh, that's just "tricks of p-adic arithmetic".

Like other sock-puppet howler trolls, when confronted
with its base incredulity, it will descend to its
lower levers of the pathos variety.

You might be happier learning about Julia trees and
raster ops, instead of shilling yet another Ramanujan
series without saying how it's made.

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

One could critisize that my -C-WAM doesn't
utilize GPU to the fullest, since its GPU
backend prototype only uses scalar operations

and no vector or matrix operations. And
modern GPUs thrive on vector and matrix
operations. Especially matrix operations giving

a boost of a factor 15x or so. There are
many papers already showing how Prolog can be
mapped to matrix operations. Only this research

is completely ignored by Prolog systems such as
SICStus, Ciao, SWI, ECLiPSe etc.. But lets
illustrate what vector operations could do

for -C-WAM, take this compilation of the Prolog
goal between(0,1023,X), Y is X*2+3:

int X;
int Y;
for (X=0; X < 1024; X++) {
Y=X*2+3;
[...]
}

With vector operations, and vectors of size
32 one could do:

int X1;
int[] X = new int[32];
int X3;
int[] Y = new int[32];
for (X1 = 0; X1 < 1024 / 32; X1++) {
for (int X2 = 0; X2 < 32; X2++)
X[X2] = X1*32+X2;
vec_mul_add(X, 2, 3, Y);
[..]
}

Have Fun!

Bye

Mild Shock schrieb:

Hi,

While Huggingfaces hired GG in 2026,
AK was hired by Anthropic in 2026:

Andrej Karpathy (born 23 October 1986[3])
is a Slovak-Canadian AI researcher, who
co-founded and formerly worked at OpenAI
In 2026 he joined Anthropic as part of
the pretraining team.
https://en.wikipedia.org/wiki/Andrej_Karpathy

But his nanochat archivement has an
interesting time line:

168 hours , Original OpenAI GPT-2 checkpoint, 2019
3 hours , d24 baseline, slightly overtrained, Jan 29 2026
1 1/2 hour, autoresearch round 2, Mar 14 2026
The best ChatGPT that $100 can buy.
https://github.com/karpathy/nanochat

But what hardware was the enabler. What is the
NVIDIA H100 GPU even. Well the thingy is surely not
a Budget Laptop, performance pretty much

dependence on data elememt size, the H100 NVL
version (*), and when using tensor operations,
and not only scalar operations:

8-bit towards 3000 tera flops
16-bit towards 1500 tera flops
32-bit towards 900 tera flops

Cool! I guess this experiment would tap into 60
tera flops, since it only uses scalar operations so far:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

You could perform it by migration the web application
using WebGPU into a node.js standalone application
using the dawn library for GPU access.

Bye

(*) https://resources.nvidia.com/en-us-hopper-architecture/nvidia-tensor-core-gpu-datasheet

Mild Shock schrieb:

Hi,

Whats this "forget" trope of glue sniffing
Rossy Boy with his herpes blisters?

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

Why should I forget Bulgarians,
they are never on my mind. Do you
see me doing ggml stuff?

I only hypothesized that it is
over for Python as the machine
learning language or AI inferencing

locally on AI laptops language, and
made the ggml case, so I already forgot
about them. Which might give you a glimps,

why WebGPU was used for this here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Is an interesting choice. Even
github has some Languages statistics,
giving an account what I used:

HTML 67.5% JavaScript 23.1% CSS 9.4%

Have Fun!

Bye

P.S.: The example below is not p-adics,
you complete imbecil moron. Its just:

7-11 cubic Solution by Pritchard & Gries
https://www.cs.cornell.edu/gries/TechReports/83-574.pdf

Ross Finlayson schrieb:

On 07/26/2026 10:52 AM, Mild Shock wrote:

Hi,

You see it all boils down to find your inner peace
by an immaculate inception of some queue datatype.

KOAN/Fortran-S was an early 1990s research programming
system for distributed-memory multiprocessors . Developed
at ENS Lyon in the early 1990s . Often listed alongside
other historical parallel programming efforts.

The Message Passing: The research explicitly
compared the SVM approach against message passing
on the same hardware . The finding was that SVM
could achieve good performance without the low-level

complexity of managing explicit messages, though
the best results often came from a hybrid approach (sic!)
Here is an interesting baseline, from Java,
a class ElevenSingle that only does:

-a-a-a-a-a public static void run() {
-a-a-a-a-a-a-a-a-a for (int A = 1; A < 192; A++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a int Y = (771-A)/3;
-a-a-a-a-a-a-a-a-a-a-a-a-a for (int B = A; B < Y; B++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int Z = (771-A-B)/2;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a for (int C = B; C < Z; C++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int D = 711-A-B-C;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a if (A*B*C == 711000000/D && >> -a>>-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a 711000000 % D == 0)
-a-a-a-a-a System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a }

And then compare it to ElevenMulti, doing some
Work Balancing Scheduler Tetris Game with 8 cores:

ElevenSingle
A=120, B=125, C=150, D=316
6.628 ms

ElevenMulti
A=120, B=125, C=150, D=316
1.941 ms

Not great, not terrible!

Bye

Oh, that's just "tricks of p-adic arithmetic".

Like other sock-puppet howler trolls, when confronted
with its base incredulity, it will descend to its
lower levers of the pathos variety.

You might be happier learning about Julia trees and
raster ops, instead of shilling yet another Ramanujan
series without saying how it's made.

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

The nice thing about AI accelerators, pioneered
maybe by Apple Silicon and their unified memory.
The AMD APU model can be extended so that

vector and matrix operations become uniformly
available for GPU and CPU. With unified memory
already a vector operation such as:

vec_mul_add(X, 2, 3, Y)

Only needs the X and Y address. But I havent
got my head around yet how this is all organized.
Maybe a GPU has still its own GEMM cores,

but you find Apple Silicon C++/C source code,
that taps into vector and matrix operations
by Zero Copying. The Copying is left to the DMA

of the vector or matrix operation. And moderated
by the various caches. Leading to the slogan, that
multiple floating point operations become zero cost:

Some teaching can be found here https://www.hpc-ch.org/category/topics/course-workshop/

Bye

Mild Shock schrieb:

Hi,

One could critisize that my -C-WAM doesn't
utilize GPU to the fullest, since its GPU
backend prototype only uses scalar operations

and no vector or matrix operations. And
modern GPUs thrive on vector and matrix
operations. Especially matrix operations giving

a boost of a factor 15x or so. There are
many papers already showing how Prolog can be
mapped to matrix operations. Only this research

is completely ignored by Prolog systems such as
SICStus, Ciao, SWI, ECLiPSe etc.. But lets
illustrate what vector operations could do

for -C-WAM, take this compilation of the Prolog
goal between(0,1023,X), Y is X*2+3:

int X;
int Y;
for (X=0; X < 1024; X++) {
-a-a-a Y=X*2+3;
-a-a-a [...]
}

With vector operations, and vectors of size
32 one could do:

int X1;
int[] X = new int[32];
int X3;
int[] Y = new int[32];
for (X1 = 0; X1 < 1024 / 32; X1++) {
-a-a-a for (int X2 = 0; X2 < 32; X2++)
-a-a-a-a-a-a X[X2] = X1*32+X2;
-a-a-a vec_mul_add(X, 2, 3, Y);
-a-a-a [..]
}

Have Fun!

Bye

Mild Shock schrieb:

Hi,

While Huggingfaces hired GG in 2026,
AK was hired by Anthropic in 2026:

Andrej Karpathy (born 23 October 1986[3])
is a Slovak-Canadian AI researcher, who
co-founded and formerly worked at OpenAI
In 2026 he joined Anthropic as part of
the pretraining team.
https://en.wikipedia.org/wiki/Andrej_Karpathy

But his nanochat archivement has an
interesting time line:

168 hours , Original OpenAI GPT-2 checkpoint, 2019
3 hours , d24 baseline, slightly overtrained, Jan 29 2026
1 1/2 hour, autoresearch round 2, Mar 14 2026
The best ChatGPT that $100 can buy.
https://github.com/karpathy/nanochat

But what hardware was the enabler. What is the
NVIDIA H100 GPU even. Well the thingy is surely not
a Budget Laptop, performance pretty much

dependence on data elememt size, the H100 NVL
version (*), and when using tensor operations,
and not only scalar operations:

8-bit towards 3000 tera flops
16-bit towards 1500 tera flops
32-bit towards 900 tera flops

Cool! I guess this experiment would tap into 60
tera flops, since it only uses scalar operations so far:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

You could perform it by migration the web application
using WebGPU into a node.js standalone application
using the dawn library for GPU access.

Bye

(*)
https://resources.nvidia.com/en-us-hopper-architecture/nvidia-tensor-core-gpu-datasheet



Mild Shock schrieb:

Hi,

Whats this "forget" trope of glue sniffing
Rossy Boy with his herpes blisters?

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

Why should I forget Bulgarians,
they are never on my mind. Do you
see me doing ggml stuff?

I only hypothesized that it is
over for Python as the machine
learning language or AI inferencing

locally on AI laptops language, and
made the ggml case, so I already forgot
about them. Which might give you a glimps,

why WebGPU was used for this here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Is an interesting choice. Even
github has some Languages statistics,
giving an account what I used:

HTML 67.5% JavaScript 23.1% CSS 9.4%

Have Fun!

Bye

P.S.: The example below is not p-adics,
you complete imbecil moron. Its just:

7-11 cubic Solution by Pritchard & Gries
https://www.cs.cornell.edu/gries/TechReports/83-574.pdf

Ross Finlayson schrieb:

On 07/26/2026 10:52 AM, Mild Shock wrote:

Hi,

You see it all boils down to find your inner peace
by an immaculate inception of some queue datatype.

KOAN/Fortran-S was an early 1990s research programming
system for distributed-memory multiprocessors . Developed
at ENS Lyon in the early 1990s . Often listed alongside
other historical parallel programming efforts.

The Message Passing: The research explicitly
compared the SVM approach against message passing
on the same hardware . The finding was that SVM
could achieve good performance without the low-level

complexity of managing explicit messages, though
the best results often came from a hybrid approach (sic!)
Here is an interesting baseline, from Java,
a class ElevenSingle that only does:

-a-a-a-a-a public static void run() {
-a-a-a-a-a-a-a-a-a for (int A = 1; A < 192; A++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a int Y = (771-A)/3;
-a-a-a-a-a-a-a-a-a-a-a-a-a for (int B = A; B < Y; B++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int Z = (771-A-B)/2;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a for (int C = B; C < Z; C++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int D = 711-A-B-C;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a if (A*B*C == 711000000/D && >>> -a>>-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a 711000000 % D == 0)
-a-a-a-a-a System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a }

And then compare it to ElevenMulti, doing some
Work Balancing Scheduler Tetris Game with 8 cores:

ElevenSingle
A=120, B=125, C=150, D=316
6.628 ms

ElevenMulti
A=120, B=125, C=150, D=316
1.941 ms

Not great, not terrible!

Bye

Oh, that's just "tricks of p-adic arithmetic".

Like other sock-puppet howler trolls, when confronted
with its base incredulity, it will descend to its
lower levers of the pathos variety.

You might be happier learning about Julia trees and
raster ops, instead of shilling yet another Ramanujan
series without saying how it's made.

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

But the example gives also way to vector
and matrix registers. The int[] X and
int[] Y could be also held in vector

registers. Compilers can also optimize
away int[] Y, and use a inline modification,
in case X isn't used later, then playing

the role of Y:

vec_mul_add(X, 2, 3, X)

Vector and matrix registers in modern GPUs
emerged from distinct architectural milestones:
vector-like register files developed with
early programmable 3D vertex/pixel pipelines

in the late 1990s to early 2000s. While
dedicated multi-dimensional matrix registers
(Tensor Cores/Matrix Cores) were invented by
NVIDIA in 2017, starting with the Tesla

V100 (Volta microarchitecture):

From Volta To Blackwell https://newsletter.semianalysis.com/p/nvidia-tensor-core-evolution-from-volta-to-blackwell

You see the scheduling of tensure core occupation
scheduling in the above article, including memory
and register flow, following the section:

MMA Instruction Overview

It went through a couple of generations, leading
to Tensor Memory (TMEM) and collective operations,
basically realizing the PIM idea:

Processing-in-Memory Tutorials https://www.sigarch.org/processing-in-memory-tutorials-experiences-from-past-two-years-and-thoughts-looking-forward/

Have Fun!

Bye


Mild Shock schrieb:

Hi,

The nice thing about AI accelerators, pioneered
maybe by Apple Silicon and their unified memory.
The AMD APU model can be extended so that

vector and matrix operations become uniformly
available for GPU and CPU. With unified memory
already a vector operation such as:

vec_mul_add(X, 2, 3, Y)

Only needs the X and Y address. But I havent
got my head around yet how this is all organized.
Maybe a GPU has still its own GEMM cores,

but you find Apple Silicon C++/C source code,
that taps into vector and matrix operations
by Zero Copying. The Copying is left to the DMA

of the vector or matrix operation. And moderated
by the various caches. Leading to the slogan, that
multiple floating point operations become zero cost:

Some teaching can be found here https://www.hpc-ch.org/category/topics/course-workshop/

Bye

Mild Shock schrieb:

Hi,

One could critisize that my -C-WAM doesn't
utilize GPU to the fullest, since its GPU
backend prototype only uses scalar operations

and no vector or matrix operations. And
modern GPUs thrive on vector and matrix
operations. Especially matrix operations giving

a boost of a factor 15x or so. There are
many papers already showing how Prolog can be
mapped to matrix operations. Only this research

is completely ignored by Prolog systems such as
SICStus, Ciao, SWI, ECLiPSe etc.. But lets
illustrate what vector operations could do

for -C-WAM, take this compilation of the Prolog
goal between(0,1023,X), Y is X*2+3:

int X;
int Y;
for (X=0; X < 1024; X++) {
-a-a-a-a Y=X*2+3;
-a-a-a-a [...]
}

With vector operations, and vectors of size
32 one could do:

int X1;
int[] X = new int[32];
int X3;
int[] Y = new int[32];
for (X1 = 0; X1 < 1024 / 32; X1++) {
-a-a-a-a for (int X2 = 0; X2 < 32; X2++)
-a-a-a-a-a-a-a X[X2] = X1*32+X2;
-a-a-a-a vec_mul_add(X, 2, 3, Y);
-a-a-a-a [..]
}

Have Fun!

Bye

Mild Shock schrieb:

Hi,

While Huggingfaces hired GG in 2026,
AK was hired by Anthropic in 2026:

Andrej Karpathy (born 23 October 1986[3])
is a Slovak-Canadian AI researcher, who
co-founded and formerly worked at OpenAI
In 2026 he joined Anthropic as part of
the pretraining team.
https://en.wikipedia.org/wiki/Andrej_Karpathy

But his nanochat archivement has an
interesting time line:

168 hours , Original OpenAI GPT-2 checkpoint, 2019
3 hours , d24 baseline, slightly overtrained, Jan 29 2026
1 1/2 hour, autoresearch round 2, Mar 14 2026
The best ChatGPT that $100 can buy.
https://github.com/karpathy/nanochat

But what hardware was the enabler. What is the
NVIDIA H100 GPU even. Well the thingy is surely not
a Budget Laptop, performance pretty much

dependence on data elememt size, the H100 NVL
version (*), and when using tensor operations,
and not only scalar operations:

8-bit towards 3000 tera flops
16-bit towards 1500 tera flops
32-bit towards 900 tera flops

Cool! I guess this experiment would tap into 60
tera flops, since it only uses scalar operations so far:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

You could perform it by migration the web application
using WebGPU into a node.js standalone application
using the dawn library for GPU access.

Bye

(*)
https://resources.nvidia.com/en-us-hopper-architecture/nvidia-tensor-core-gpu-datasheet



Mild Shock schrieb:

Hi,

Whats this "forget" trope of glue sniffing
Rossy Boy with his herpes blisters?

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

Why should I forget Bulgarians,
they are never on my mind. Do you
see me doing ggml stuff?

I only hypothesized that it is
over for Python as the machine
learning language or AI inferencing

locally on AI laptops language, and
made the ggml case, so I already forgot
about them. Which might give you a glimps,

why WebGPU was used for this here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Is an interesting choice. Even
github has some Languages statistics,
giving an account what I used:

HTML 67.5% JavaScript 23.1% CSS 9.4%

Have Fun!

Bye

P.S.: The example below is not p-adics,
you complete imbecil moron. Its just:

7-11 cubic Solution by Pritchard & Gries
https://www.cs.cornell.edu/gries/TechReports/83-574.pdf

Ross Finlayson schrieb:

On 07/26/2026 10:52 AM, Mild Shock wrote:

Hi,

You see it all boils down to find your inner peace
by an immaculate inception of some queue datatype.

KOAN/Fortran-S was an early 1990s research programming
system for distributed-memory multiprocessors . Developed
at ENS Lyon in the early 1990s . Often listed alongside
other historical parallel programming efforts.

The Message Passing: The research explicitly
compared the SVM approach against message passing
on the same hardware . The finding was that SVM
could achieve good performance without the low-level

complexity of managing explicit messages, though
the best results often came from a hybrid approach (sic!)
Here is an interesting baseline, from Java,
a class ElevenSingle that only does:

-a-a-a-a-a public static void run() {
-a-a-a-a-a-a-a-a-a for (int A = 1; A < 192; A++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a int Y = (771-A)/3;
-a-a-a-a-a-a-a-a-a-a-a-a-a for (int B = A; B < Y; B++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int Z = (771-A-B)/2;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a for (int C = B; C < Z; C++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a int D = 711-A-B-C;
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a if (A*B*C == 711000000/D && >>>> -a>>-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a 711000000 % D == 0)
-a-a-a-a-a System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a }

And then compare it to ElevenMulti, doing some
Work Balancing Scheduler Tetris Game with 8 cores:

ElevenSingle
A=120, B=125, C=150, D=316
6.628 ms

ElevenMulti
A=120, B=125, C=150, D=316
1.941 ms

Not great, not terrible!

Bye

Oh, that's just "tricks of p-adic arithmetic".

Like other sock-puppet howler trolls, when confronted
with its base incredulity, it will descend to its
lower levers of the pathos variety.

You might be happier learning about Julia trees and
raster ops, instead of shilling yet another Ramanujan
series without saying how it's made.

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

On 07/27/2026 04:22 AM, Mild Shock wrote:

Hi,

But the example gives also way to vector
and matrix registers. The int[] X and
int[] Y could be also held in vector

registers. Compilers can also optimize
away int[] Y, and use a inline modification,
in case X isn't used later, then playing

the role of Y:

vec_mul_add(X, 2, 3, X)

Vector and matrix registers in modern GPUs
emerged from distinct architectural milestones:
vector-like register files developed with
early programmable 3D vertex/pixel pipelines

in the late 1990s to early 2000s. While
dedicated multi-dimensional matrix registers
(Tensor Cores/Matrix Cores) were invented by
NVIDIA in 2017, starting with the Tesla

V100 (Volta microarchitecture):

From Volta To Blackwell https://newsletter.semianalysis.com/p/nvidia-tensor-core-evolution-from-volta-to-blackwell

You see the scheduling of tensure core occupation
scheduling in the above article, including memory
and register flow, following the section:

MMA Instruction Overview

It went through a couple of generations, leading
to Tensor Memory (TMEM) and collective operations,
basically realizing the PIM idea:

Processing-in-Memory Tutorials https://www.sigarch.org/processing-in-memory-tutorials-experiences-from-past-two-years-and-thoughts-looking-forward/

Have Fun!

Bye

Mild Shock schrieb:

Hi,

The nice thing about AI accelerators, pioneered
maybe by Apple Silicon and their unified memory.
The AMD APU model can be extended so that

vector and matrix operations become uniformly
available for GPU and CPU. With unified memory
already a vector operation such as:

vec_mul_add(X, 2, 3, Y)

Only needs the X and Y address. But I havent
got my head around yet how this is all organized.
Maybe a GPU has still its own GEMM cores,

but you find Apple Silicon C++/C source code,
that taps into vector and matrix operations
by Zero Copying. The Copying is left to the DMA

of the vector or matrix operation. And moderated
by the various caches. Leading to the slogan, that
multiple floating point operations become zero cost:

Some teaching can be found here
https://www.hpc-ch.org/category/topics/course-workshop/

Bye

Mild Shock schrieb:

Hi,

One could critisize that my -C-WAM doesn't
utilize GPU to the fullest, since its GPU
backend prototype only uses scalar operations

and no vector or matrix operations. And
modern GPUs thrive on vector and matrix
operations. Especially matrix operations giving

a boost of a factor 15x or so. There are
many papers already showing how Prolog can be
mapped to matrix operations. Only this research

is completely ignored by Prolog systems such as
SICStus, Ciao, SWI, ECLiPSe etc.. But lets
illustrate what vector operations could do

for -C-WAM, take this compilation of the Prolog
goal between(0,1023,X), Y is X*2+3:

int X;
int Y;
for (X=0; X < 1024; X++) {
Y=X*2+3;
[...]
}

With vector operations, and vectors of size
32 one could do:

int X1;
int[] X = new int[32];
int X3;
int[] Y = new int[32];
for (X1 = 0; X1 < 1024 / 32; X1++) {
for (int X2 = 0; X2 < 32; X2++)
X[X2] = X1*32+X2;
vec_mul_add(X, 2, 3, Y);
[..]
}

Have Fun!

Bye

Mild Shock schrieb:

Hi,

While Huggingfaces hired GG in 2026,
AK was hired by Anthropic in 2026:

Andrej Karpathy (born 23 October 1986[3])
is a Slovak-Canadian AI researcher, who
co-founded and formerly worked at OpenAI
In 2026 he joined Anthropic as part of
the pretraining team.
https://en.wikipedia.org/wiki/Andrej_Karpathy

But his nanochat archivement has an
interesting time line:

168 hours , Original OpenAI GPT-2 checkpoint, 2019
3 hours , d24 baseline, slightly overtrained, Jan 29 2026
1 1/2 hour, autoresearch round 2, Mar 14 2026
The best ChatGPT that $100 can buy.
https://github.com/karpathy/nanochat

But what hardware was the enabler. What is the
NVIDIA H100 GPU even. Well the thingy is surely not
a Budget Laptop, performance pretty much

dependence on data elememt size, the H100 NVL
version (*), and when using tensor operations,
and not only scalar operations:

8-bit towards 3000 tera flops
16-bit towards 1500 tera flops
32-bit towards 900 tera flops

Cool! I guess this experiment would tap into 60
tera flops, since it only uses scalar operations so far:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

You could perform it by migration the web application
using WebGPU into a node.js standalone application
using the dawn library for GPU access.

Bye

(*)
https://resources.nvidia.com/en-us-hopper-architecture/nvidia-tensor-core-gpu-datasheet



Mild Shock schrieb:

Hi,

Whats this "forget" trope of glue sniffing
Rossy Boy with his herpes blisters?

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

Why should I forget Bulgarians,
they are never on my mind. Do you
see me doing ggml stuff?

I only hypothesized that it is
over for Python as the machine
learning language or AI inferencing

locally on AI laptops language, and
made the ggml case, so I already forgot
about them. Which might give you a glimps,

why WebGPU was used for this here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Is an interesting choice. Even
github has some Languages statistics,
giving an account what I used:

HTML 67.5% JavaScript 23.1% CSS 9.4%

Have Fun!

Bye

P.S.: The example below is not p-adics,
you complete imbecil moron. Its just:

7-11 cubic Solution by Pritchard & Gries
https://www.cs.cornell.edu/gries/TechReports/83-574.pdf

Ross Finlayson schrieb:

On 07/26/2026 10:52 AM, Mild Shock wrote:

Hi,

You see it all boils down to find your inner peace
by an immaculate inception of some queue datatype.

KOAN/Fortran-S was an early 1990s research programming
system for distributed-memory multiprocessors . Developed
at ENS Lyon in the early 1990s . Often listed alongside
other historical parallel programming efforts.

The Message Passing: The research explicitly
compared the SVM approach against message passing
on the same hardware . The finding was that SVM
could achieve good performance without the low-level

complexity of managing explicit messages, though
the best results often came from a hybrid approach (sic!)
Here is an interesting baseline, from Java,
a class ElevenSingle that only does:

public static void run() {
for (int A = 1; A < 192; A++) {
int Y = (771-A)/3;
for (int B = A; B < Y; B++) {
int Z = (771-A-B)/2;
for (int C = B; C < Z; C++) {
int D = 711-A-B-C;
if (A*B*C == 711000000/D &&
711000000 % D == 0)
System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
}
}
}
}

And then compare it to ElevenMulti, doing some
Work Balancing Scheduler Tetris Game with 8 cores:

ElevenSingle
A=120, B=125, C=150, D=316
6.628 ms

ElevenMulti
A=120, B=125, C=150, D=316
1.941 ms

Not great, not terrible!

Bye

Oh, that's just "tricks of p-adic arithmetic".

Like other sock-puppet howler trolls, when confronted
with its base incredulity, it will descend to its
lower levers of the pathos variety.

You might be happier learning about Julia trees and
raster ops, instead of shilling yet another Ramanujan
series without saying how it's made.

Bulgarians, that's some real Boris and Natasha crap,
forget Hungarians and Bulgarians.

That's bullshit, and alike those talking heads that
sniff their way into talking about many-core jumbo-trons,
the super-scalar is as old as the scalar and Cray and examples alike
the Connection Machine what made all the craze of neural nets
is old-wrapped-as-new.

Fabless chips did it already.


Data centers should pay a 10000% excise on electricity,
wherever it comes from, a natural regulator of inverted economies.

And by ten thousand percent I really mean a ten thousand percent.


--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Slowly I start understanding numbnuts like
Rossy Boy who don't understand tech, although
they are from UK and not from a 3rd world

country, and also I start understanding morons
like Micro Penis, who are behind a curtain,
and cannot access a lot of tech.

The same holds for SWI Prologs newest campaign
that probably adresses some poor indians that
have neither 5G nor Macs:

1:38:01 The Kyiv keynote disaster
https://www.youtube.com/watch?v=U8goS6B3BbI

Woa! Real time download of Scala, Closure,
etc.. Whats the magic behind that? Some SWI
point of sale, downloading it via its

keyboard and some telephathy module ?

Bye

Mild Shock schrieb:

Hi,

Ride the snake
He's old and his skin is cold
The west is the best
The west is the best
Get here and we'll do the rest
The blue bus is calling us
The blue bus is calling us
Driver, where you taking us?

Apocalypse Now intro: The Doors, The End {1979} https://www.youtube.com/watch?v=CIrvSJwwJUE

Bye

Hi,

Again I posted everything here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

The repo says, same time when I posted
the link first time:

This repository was archived by the
owner on Jul 9, 2026. It is now read-only.

Now a USENET user, who had already entitled
himself for a couple of irrational accusations

towards my side, is asking this question:

Chris M. Thomasson schrieb, Jul 24, 2026

Show an outline of what you
need you compute shader to do?

Bravo, thats a delay of a wooping 15 days.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

I guess Rossy Boys mother was so disappointed
in the 50's that is son didn't become the
next Einstain, physics was the ultinate idol,

so that Rossy Boy was left rotting in the
basement. But Rossy Boys indoctrination was
not spurious, he now is conditioned on

Einstein. Maybe NVIDIA should have named
its Tesla V100 card NVIDIA Einstein. You would
then see Rossy Boy toe sucking the graphic

card, in his pyjamas in the basement.

Bye

Ross Finlayson schrieb:

On 07/27/2026 04:22 AM, Mild Shock wrote:

Hi,

But the example gives also way to vector
and matrix registers. The int[] X and
int[] Y could be also held in vector

registers. Compilers can also optimize
away int[] Y, and use a inline modification,
in case X isn't used later, then playing

the role of Y:

vec_mul_add(X, 2, 3, X)

Vector and matrix registers in modern GPUs
emerged from distinct architectural milestones:
vector-like register files developed with
early programmable 3D vertex/pixel pipelines

in the late 1990s to early 2000s. While
dedicated multi-dimensional matrix registers
(Tensor Cores/Matrix Cores) were invented by
NVIDIA in 2017, starting with the Tesla

V100 (Volta microarchitecture):

-aFrom Volta To Blackwell
https://newsletter.semianalysis.com/p/nvidia-tensor-core-evolution-from-volta-to-blackwell



You see the scheduling of tensure core occupation
scheduling in the above article, including memory
and register flow, following the section:

MMA Instruction Overview

It went through a couple of generations, leading
to Tensor Memory (TMEM) and collective operations,
basically realizing the PIM idea:

Processing-in-Memory Tutorials
https://www.sigarch.org/processing-in-memory-tutorials-experiences-from-past-two-years-and-thoughts-looking-forward/



Have Fun!

Bye

That's bullshit, and alike those talking heads that
sniff their way into talking about many-core jumbo-trons,
the super-scalar is as old as the scalar and Cray and examples alike
the Connection Machine what made all the craze of neural nets
is old-wrapped-as-new.

Fabless chips did it already.

Data centers should pay a 10000% excise on electricity,
wherever it comes from, a natural regulator of inverted economies.

And by ten thousand percent I really mean a ten thousand percent.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

In comp.lang.prolog Ross Finlayson <ross.a.finlayson@gmail.com> wrote:
...

Data centers should pay a 10000% excise on electricity,
wherever it comes from, a natural regulator of inverted economies.
And by ten thousand percent I really mean a ten thousand percent.

And what would a huge surcharge do?
Almost always end up affecting the less powerful end of society
with increased costs to services the AI industry will be doing
more and more of over time.

I started a little data center (exaflops.com) many years ago.
In those distant days people (in fact one was a prof of computer
science) told me you could never make money running a supercomputer.
LOL. :)

I've had many years to watch the trends and a far more efficient
way to solve resource problems in this area is to change the
algorithms. There is vast room for improvement, mostly because
of prevailing attitudes.

I used to do competetion data science as a sideline. Companies
would pay almost any price to get an extra decimal place in
the accuracy of their forecasting processes. But typically
they were trying to supercharge a system that should be scrapped
and re-designed from scratch. One area I'm thinking of is
investment. I had a customer one time -- like many times --
ask to improve a system that predicted the future price of
various stocks. The idea (for them) was to have as accurate a
prediction of what some stock would be worth in a week or a month's
time so that some moron could use the information to decide when
to buy or sell the thing.

I tried to argue the efficient thing was to create a system that
takes the human out of the loop altogether. It doesnt provide info
for someone to decide whether or not to follow the advice --
that is just introducing more noise into the loop and probably
cancels any benefit of adding a couple decimal places of precision.
What you *should* do is make a system that is tuned to robustly
maximize the profit from managing a portfolio.

Of course they wouldnt come at that. You can't suggest taking the
managers out of the loop. :)

Another idea relevant to current AI methods might be to curtail
use of typical neural net algorithms. Many of them try to squeeze
the best performance of some NN during the training phase in
the hope the resulting system will generalize well enough to be useful
on new data. But there's kind-of a law that the harder you train
some system to perform a task well, the less well they can subsuently
perform a more general version of the same thing. It's amusing when
you look at the graphs of NN being trained and then tested that
given a more general problem to solve after being trained to solve
similar problems very very well the poor old NN does worse that it
would have done if it had 0 training in the first place.

It's not like we dont know how to improve this kind of performance.
Try less hard in the training phase or make it "more noisy".
Turns out genetic methods are just the ticket for this.
The training produces less over-fitting and the resulting system
generalizes better than it did before training and more importantly
it takes maybe an order of magnitude crunching to produce a good answer
than the usual over-fit answer.

Anyway. Have to go and feed the cat.
--
Nothing in life is to be feared, it is only to be understood.
Now is the time to understand more, so that we may fear less.
-- M. Curie
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

In comp.lang.prolog R Kym Horsell <kym@sdf.com> wrote:

generalizes better than it did before training and more importantly
it takes maybe an order of magnitude crunching to produce a good answer

/\ less

than the usual over-fit answer.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Rossy Boy is neither Einstein nor Zweistein.
He is not Einstein since Einstein is already dead:

Albert Einstein (1879 - 1955)
https://de.wikipedia.org/wiki/Albert_Einstein

He is also not Zweistein, since he doesn't
understand concepts such as:

- NVIDIA Volta ff. architecture

Also his hands are small, and his breath stinks,
and he lives in the basement of his mother.

Bye

Mild Shock schrieb:

Hi,

Slowly I start understanding numbnuts like
Rossy Boy who don't understand tech, although
they are from UK and not from a 3rd world

country, and also I start understanding morons
like Micro Penis, who are behind a curtain,
and cannot access a lot of tech.

The same holds for SWI Prologs newest campaign
that probably adresses some poor indians that
have neither 5G nor Macs:

1:38:01 The Kyiv keynote disaster
https://www.youtube.com/watch?v=U8goS6B3BbI

Woa! Real time download of Scala, Closure,
etc.. Whats the magic behind that? Some SWI
point of sale, downloading it via its

keyboard and some telephathy module ?

Bye

Mild Shock schrieb:

Hi,

Ride the snake
He's old and his skin is cold
The west is the best
The west is the best
Get here and we'll do the rest
The blue bus is calling us
The blue bus is calling us
Driver, where you taking us?

Apocalypse Now intro: The Doors, The End {1979}
https://www.youtube.com/watch?v=CIrvSJwwJUE

Bye

Hi,

Again I posted everything here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

The repo says, same time when I posted
the link first time:

This repository was archived by the
owner on Jul 9, 2026. It is now read-only.

Now a USENET user, who had already entitled
himself for a couple of irrational accusations

towards my side, is asking this question:

Chris M. Thomasson schrieb, Jul 24, 2026

Show an outline of what you
need you compute shader to do?

Bravo, thats a delay of a wooping 15 days.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Come on Horsy Boy, you can do better. I
no where wrote something about curve
fitting and/or increasing the precision of

float point numbers:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

What makes you think LIPS measures precision?
You should know better as a 50% Prologer.

I explictily wrote here what the goal is:

"shave off some of the TOPS to do Prolog inferencing"

What are TOPS? Its a metric for GPUs:

TOPS stands for rCLTrillions of Operations Per Second.rCY https://www.lenovo.com/us/en/glossary/tops-in-computing/

See for yourself what is behind my post:

11.4 Giga Lips with a Budget Laptop
At the end of 2025 we acquired a couple of AI Laptops , that were still
cheap, since RAM prices had not yet rocketed. The intend was to tap into
the Copilot+ certified hardware, and shave off some of the TOPS to do
Prolog inferencing. Amazingly our -C-WAM can churn 11.4 GIGA LIPS.

GPUs have evolved form lock-step to independent thread scheduling. This
made it possible to port the Hack VM variant, that forms the basis for
our -C-WAM, to WebGPU computer shaders. Using NUM_SHADERS = 4096 we could produce 11.4 Giga Lips on a Ryzen AI 7 350 w/ Radeon 860M.

See also:

Medium Article - 11.4 Giga Lips
https://medium.com/2989/899b0d5c027b

So just get lost with your crazy irrelevant rant.
When I get more LIPS, things run faster, and
I remove digits from the time dimension.

Got it. Or are you too stupid?

Bye

R Kym Horsell schrieb:

In comp.lang.prolog Ross Finlayson <ross.a.finlayson@gmail.com> wrote:
...

Data centers should pay a 10000% excise on electricity,
wherever it comes from, a natural regulator of inverted economies.
And by ten thousand percent I really mean a ten thousand percent.

And what would a huge surcharge do?
Almost always end up affecting the less powerful end of society
with increased costs to services the AI industry will be doing
more and more of over time.

I started a little data center (exaflops.com) many years ago.
In those distant days people (in fact one was a prof of computer
science) told me you could never make money running a supercomputer.
LOL. :)

I've had many years to watch the trends and a far more efficient
way to solve resource problems in this area is to change the
algorithms. There is vast room for improvement, mostly because
of prevailing attitudes.

I used to do competetion data science as a sideline. Companies
would pay almost any price to get an extra decimal place in
the accuracy of their forecasting processes. But typically
they were trying to supercharge a system that should be scrapped
and re-designed from scratch. One area I'm thinking of is
investment. I had a customer one time -- like many times --
ask to improve a system that predicted the future price of
various stocks. The idea (for them) was to have as accurate a
prediction of what some stock would be worth in a week or a month's
time so that some moron could use the information to decide when
to buy or sell the thing.

I tried to argue the efficient thing was to create a system that
takes the human out of the loop altogether. It doesnt provide info
for someone to decide whether or not to follow the advice --
that is just introducing more noise into the loop and probably
cancels any benefit of adding a couple decimal places of precision.
What you *should* do is make a system that is tuned to robustly
maximize the profit from managing a portfolio.

Of course they wouldnt come at that. You can't suggest taking the
managers out of the loop. :)

Another idea relevant to current AI methods might be to curtail
use of typical neural net algorithms. Many of them try to squeeze
the best performance of some NN during the training phase in
the hope the resulting system will generalize well enough to be useful
on new data. But there's kind-of a law that the harder you train
some system to perform a task well, the less well they can subsuently
perform a more general version of the same thing. It's amusing when
you look at the graphs of NN being trained and then tested that
given a more general problem to solve after being trained to solve
similar problems very very well the poor old NN does worse that it
would have done if it had 0 training in the first place.

It's not like we dont know how to improve this kind of performance.
Try less hard in the training phase or make it "more noisy".
Turns out genetic methods are just the ticket for this.
The training produces less over-fitting and the resulting system
generalizes better than it did before training and more importantly
it takes maybe an order of magnitude crunching to produce a good answer
than the usual over-fit answer.

Anyway. Have to go and feed the cat.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Because of the mobile GPU design, the TOPS,
aka rCLTrillions of Operations Per Second.rCY
come with not extremly high power consumption.

Especially the presence of vector and matrix
operations can lower the energy consumption,
since they can avoid redundant memory access.

Its quite a difference between discrete graphic
cards and accelerator iGPUs that are directly
on the silicon chip, and have mobile design.

So basically with newer AI Laptops you get more
performence units for less energy units.

Have Fun!

Bye

Mild Shock schrieb:

Hi,

Come on Horsy Boy, you can do better. I
no where wrote something about curve
fitting and/or increasing the precision of

float point numbers:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

What makes you think LIPS measures precision?
You should know better as a 50% Prologer.

I explictily wrote here what the goal is:

"shave off some of the TOPS to do Prolog inferencing"

What are TOPS? Its a metric for GPUs:

TOPS stands for rCLTrillions of Operations Per Second.rCY https://www.lenovo.com/us/en/glossary/tops-in-computing/

See for yourself what is behind my post:

11.4 Giga Lips with a Budget Laptop
At the end of 2025 we acquired a couple of AI Laptops , that were still cheap, since RAM prices had not yet rocketed. The intend was to tap into
the Copilot+ certified hardware, and shave off some of the TOPS to do
Prolog inferencing. Amazingly our -C-WAM can churn 11.4 GIGA LIPS.

GPUs have evolved form lock-step to independent thread scheduling. This
made it possible to port the Hack VM variant, that forms the basis for
our -C-WAM, to WebGPU computer shaders. Using NUM_SHADERS = 4096 we could produce 11.4 Giga Lips on a Ryzen AI 7 350 w/ Radeon 860M.

See also:

Medium Article - 11.4 Giga Lips
https://medium.com/2989/899b0d5c027b

So just get lost with your crazy irrelevant rant.
When I get more LIPS, things run faster, and
I remove digits from the time dimension.

Got it. Or are you too stupid?

Bye

R Kym Horsell schrieb:

In comp.lang.prolog Ross Finlayson <ross.a.finlayson@gmail.com> wrote:
...

Data centers should pay a 10000% excise on electricity,
wherever it comes from, a natural regulator of inverted economies.
And by ten thousand percent I really mean a ten thousand percent.

And what would a huge surcharge do?
Almost always end up affecting the less powerful end of society
with increased costs to services the AI industry will be doing
more and more of over time.

I started a little data center (exaflops.com) many years ago.
In those distant days people (in fact one was a prof of computer
science) told me you could never make money running a supercomputer.
LOL. :)

I've had many years to watch the trends and a far more efficient
way to solve resource problems in this area is to change the
algorithms. There is vast room for improvement, mostly because
of prevailing attitudes.

I used to do competetion data science as a sideline. Companies
would pay almost any price to get an extra decimal place in
the accuracy of their forecasting processes. But typically
they were trying to supercharge a system that should be scrapped
and re-designed from scratch. One area I'm thinking of is
investment. I had a customer one time -- like many times --
ask to improve a system that predicted the future price of
various stocks. The idea (for them) was to have as accurate a
prediction of what some stock would be worth in a week or a month's
time so that some moron could use the information to decide when
to buy or sell the thing.

I tried to argue the efficient thing was to create a system that
takes the human out of the loop altogether. It doesnt provide info
for someone to decide whether or not to follow the advice --
that is just introducing more noise into the loop and probably
cancels any benefit of adding a couple decimal places of precision.
What you *should* do is make a system that is tuned to robustly
maximize the profit from managing a portfolio.

Of course they wouldnt come at that. You can't suggest taking the
managers out of the loop. :)

Another idea relevant to current AI methods might be to curtail
use of typical neural net algorithms. Many of them try to squeeze
the best performance of some NN during the training-a phase in
the hope the resulting system will generalize well enough to be useful
on new data. But there's kind-of a law that the harder you train
some system to perform a task well, the less well they can subsuently
perform a more general version of the same thing. It's amusing when
you look at the graphs of NN being trained and then tested that
given a more general problem to solve after being trained to solve
similar problems very very well the poor old NN does worse that it
would have done if it had 0 training in the first place.

It's not like we dont know how to improve this kind of performance.
Try less hard in the training phase or make it "more noisy".
Turns out genetic methods are just the ticket for this.
The training produces less over-fitting and the resulting system
generalizes better than it did before training and more importantly
it takes maybe an order of magnitude crunching to produce a good answer
than the usual over-fit answer.

Anyway. Have to go and feed the cat.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Ok, guys lets face it. You are a bunch of
morons. When did I do this post:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Yes on Jul 9, 2026, now we have Jul 27, 2026.
Thats a wooping 18 days meanwhile.
And you still don't get the meaning and

implications of the post. Like you even
don't get what "budget" nowdays means in
terms of performance units and energy units?

And what LIPS means, drawn from TOPS,
in terms of applications? Shame on you guys!
You are a bunch of brainless idiots.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly. https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic


Hi,

Ha Ha, ASML stocks are plunging:

ASML verliert |+ber f|+nf Prozent
nach Meldung |+ber chinesische Lithografie-Pl|nne. https://www.boerse-express.com/news/articles/asml-551-prozent-crash-auf-chinas-duv-plaene-932825

I told you so! in 2027 there will be
smartphones with Kirin AI chips:

Huawei plans new smartphone chips this fall as
rivalry with Nvidia and Apple heats up https://www.cnbc.com/2026/05/25/huawei-chip-logicfolding-semiconductor-nvidia-china.html

Bye

Mild Shock schrieb:

Hi,

How it started:

Captain: Throw the switch, Scotty!
Enterprise: Cloaking Device makes it invisible
Spock: Military secrets are the most fleeting of all.
Kirk Escapes the Romulans - The Enterprise Incident https://www.youtube.com/watch?v=AusAGjwlql8

How its going:

CEO Jensen Huang said the company has rCLlargely
concededrCY ChinarCOs artificial intelligence chip
market to Huawei, as U.S. export restrictions
continue to reshape the global AI semiconductor landscape.

https://www.cnbc.com/2026/05/21/nvidia-jensen-huang-china-ai-chip-market-huawei.html

Bye

P.S.: What does China do?

HuaweirCOs semiconductor chief He Tingbo at the IEEE
ISCAS 2026 conference, Huawei's Tau Scaling Law is a newly
introduced semiconductor design framework that
shifts the industryrCOs optimization focus from
geometric scaling (shrinking physical transistor
sizes) to temporal scaling (compressing signal
propagation delay).
Nvidia Gave Up China - 4 Days Later THIS Happened https://www.youtube.com/watch?v=dLLw-qADKSU

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic


Hi,

Ha Ha, ASML stocks are plunging:

ASML verliert |+ber f|+nf Prozent
nach Meldung |+ber chinesische Lithografie-Pl|nne. https://www.boerse-express.com/news/articles/asml-551-prozent-crash-auf-chinas-duv-plaene-932825

I told you so! in 2027 there will be
smartphones with Kirin AI chips:

Huawei plans new smartphone chips this fall as
rivalry with Nvidia and Apple heats up https://www.cnbc.com/2026/05/25/huawei-chip-logicfolding-semiconductor-nvidia-china.html

Bye

Mild Shock schrieb:

Hi,

How it started:

Captain: Throw the switch, Scotty!
Enterprise: Cloaking Device makes it invisible
Spock: Military secrets are the most fleeting of all.
Kirk Escapes the Romulans - The Enterprise Incident https://www.youtube.com/watch?v=AusAGjwlql8

How its going:

CEO Jensen Huang said the company has rCLlargely
concededrCY ChinarCOs artificial intelligence chip
market to Huawei, as U.S. export restrictions
continue to reshape the global AI semiconductor landscape.

https://www.cnbc.com/2026/05/21/nvidia-jensen-huang-china-ai-chip-market-huawei.html

Bye

P.S.: What does China do?

HuaweirCOs semiconductor chief He Tingbo at the IEEE
ISCAS 2026 conference, Huawei's Tau Scaling Law is a newly
introduced semiconductor design framework that
shifts the industryrCOs optimization focus from
geometric scaling (shrinking physical transistor
sizes) to temporal scaling (compressing signal
propagation delay).
Nvidia Gave Up China - 4 Days Later THIS Happened https://www.youtube.com/watch?v=dLLw-qADKSU

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Why does this Lama have a red pyjama.
Oh, its a baby Lama. Its still in the cradle
and needs some training:

RedPajama-Data-v2
https://github.com/togethercomputer/RedPajama-Data

But then Andrej Karpathy recently showed
GPT-2 training on rented GPUs for less
than 100 USD in less then 2 hours.

So where do these grown up Lamas go.
Well Georgi Gerganov prefered C++/C
when he shouted Llama Llama Red Pyjama.

But you also find WebLLM, wrapping the
underlying C++/C GPU interface via the
W3C standard WebGPU / WGSL, with JavaScript:

In-Browser LLM Inference Engine
https://webllm.mlc.ai/

My experience with WebLLM 6 months
ago on an iPad Pro 2024, still a little early
stage performance and robustness.

But hey hardware of AI mobile iGPUs is
still evolving, and AI laptop, AI smartphones
and AI tablets, will soon feature Chinese

hardware such some new Kirin AI in 2027.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly. https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Usual question:

Why implement both pre-emptive threading

AND cooperative tasks/engines?

I had implemented the ISO proposal in formerly Jekejeke
Prolog, you find the ISO proposal here:

ISO/IEC DTR 13211rCo5:2007
Prolog multi-threading support
https://logtalk.org/plstd/threads.pdf

But the ISO proposal doesn't match modern WebGPU APIs,
where your logical threads can live remotely in a dedicated GPU
in the VRAM there, and where you would have launch

parameters that say: Hey please run 4096 compute
shaders for me, that have independet thread state. Using
cooperative multi-tasking as the orchestrator works well.

Bye

Mild Shock schrieb:

Hi,

Why does this Lama have a red pyjama.
Oh, its a baby Lama. Its still in the cradle
and needs some training:

RedPajama-Data-v2
https://github.com/togethercomputer/RedPajama-Data

But then Andrej Karpathy recently showed
GPT-2 training on rented GPUs for less
than 100 USD in less then 2 hours.

So where do these grown up Lamas go.
Well Georgi Gerganov prefered C++/C
when he shouted Llama Llama Red Pyjama.

But you also find WebLLM, wrapping the
underlying C++/C GPU interface via the
W3C standard WebGPU / WGSL, with JavaScript:

In-Browser LLM Inference Engine
https://webllm.mlc.ai/

My experience with WebLLM 6 months
ago on an iPad Pro 2024, still a little early
stage performance and robustness.

But hey hardware of AI mobile iGPUs is
still evolving, and AI laptop, AI smartphones
and AI tablets, will soon feature Chinese

hardware such some new Kirin AI in 2027.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Mostlikely for high performance computing |a la,
the Actor/Erlang model is dead, they might rely
on MPMC (Multiple Producer, Multiple Consumer)

queue entities separate from the threads. The
ISO Prolog multi-threading support had also such
threads. But besides that was also Actor/Erlang

leaning in practice, like SWI, where threads
have some default queues. So an actor is basically
a Thread and Mailbox conflation. While a MPMC queue

is a kind of separate Mailbox, where multiple
"actors" can read from and write from. A kind of
localized Linda Tuple store.

Which Programming language did adopted the
non-Actor pi-calculus model? Right golang
with its channels.

Bye


Mild Shock schrieb:

Hi,

Usual question:

Why implement both pre-emptive threading

AND cooperative tasks/engines?

I had implemented the ISO proposal in formerly Jekejeke
Prolog, you find the ISO proposal here:

ISO/IEC DTR 13211rCo5:2007
Prolog multi-threading support
https://logtalk.org/plstd/threads.pdf

But the ISO proposal doesn't match modern WebGPU APIs,
where your logical threads can live remotely in a dedicated GPU
in the VRAM there, and where you would have launch

parameters that say: Hey please run 4096 compute
shaders for me, that have independet thread state. Using
cooperative multi-tasking as the orchestrator works well.

Bye

Mild Shock schrieb:

Hi,

Why does this Lama have a red pyjama.
Oh, its a baby Lama. Its still in the cradle
and needs some training:

RedPajama-Data-v2
https://github.com/togethercomputer/RedPajama-Data

But then Andrej Karpathy recently showed
GPT-2 training on rented GPUs for less
than 100 USD in less then 2 hours.

So where do these grown up Lamas go.
Well Georgi Gerganov prefered C++/C
when he shouted Llama Llama Red Pyjama.

But you also find WebLLM, wrapping the
underlying C++/C GPU interface via the
W3C standard WebGPU / WGSL, with JavaScript:

In-Browser LLM Inference Engine
https://webllm.mlc.ai/

My experience with WebLLM 6 months
ago on an iPad Pro 2024, still a little early
stage performance and robustness.

But hey hardware of AI mobile iGPUs is
still evolving, and AI laptop, AI smartphones
and AI tablets, will soon feature Chinese

hardware such some new Kirin AI in 2027.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

You are still chewing on SIMD. LoL

Ross Finlayson schrieb:

Then the idea is that any of those can be found and matched in
one "run", i.e. a stall-less, branch-less, call-less list of less than
a few or less than a few dozens or less than a few hundreds
instructions, the results "findings" in data and corresponding
"matchings" of expressions, that runs in less than one microsecond.

You cannot make the mental translation that if you have:

Ross Finlayson schrieb:

So, the context then is for register state and stack contents, that
the indicators of the above as "positive presence" then is to make
for that the adjustments to the offsets and extents and the shifts
is according to those, otherwise no-ops. Then the idea is that a

As independent logical thread state, that automatically MIMD follows?

Whats the problem to solve then?

Bye

Mild Shock schrieb:

Hi,

Rossy Boy is neither Einstein nor Zweistein.
He is not Einstein since Einstein is already dead:

Albert Einstein (1879 - 1955)
https://de.wikipedia.org/wiki/Albert_Einstein

He is also not Zweistein, since he doesn't
understand concepts such as:

- NVIDIA Volta ff. architecture

Also his hands are small, and his breath stinks,
and he lives in the basement of his mother.

Bye

Mild Shock schrieb:

Hi,

Slowly I start understanding numbnuts like
Rossy Boy who don't understand tech, although
they are from UK and not from a 3rd world

country, and also I start understanding morons
like Micro Penis, who are behind a curtain,
and cannot access a lot of tech.

The same holds for SWI Prologs newest campaign
that probably adresses some poor indians that
have neither 5G nor Macs:

1:38:01 The Kyiv keynote disaster
https://www.youtube.com/watch?v=U8goS6B3BbI

Woa! Real time download of Scala, Closure,
etc.. Whats the magic behind that? Some SWI
point of sale, downloading it via its

keyboard and some telephathy module ?

Bye

Mild Shock schrieb:

Hi,

Ride the snake
He's old and his skin is cold
The west is the best
The west is the best
Get here and we'll do the rest
The blue bus is calling us
The blue bus is calling us
Driver, where you taking us?

Apocalypse Now intro: The Doors, The End {1979}
https://www.youtube.com/watch?v=CIrvSJwwJUE

Bye

Hi,

Again I posted everything here:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

The repo says, same time when I posted
the link first time:

This repository was archived by the
owner on Jul 9, 2026. It is now read-only.

Now a USENET user, who had already entitled
himself for a couple of irrational accusations

towards my side, is asking this question:

Chris M. Thomasson schrieb, Jul 24, 2026

Show an outline of what you
need you compute shader to do?

Bravo, thats a delay of a wooping 15 days.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people started
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Hurry Rossy Boy, the blue bus is waiting.
There is a quite a hyperbole from here:

Tesla S1070 in 2008
700 Watts , 1 Terra Flop
SOLVE TOMORROWrCOS PROBLEMS TODAY https://www.azken.com/download/Tesla_DS_S1070_EU.pdf

To here:

Blackwell GPU in 2026
575 Watts, 104.8 Terra Flops ( RTX 5090 )
From Volta To Blackwell https://newsletter.semianalysis.com/p/nvidia-tensor-core-evolution-from-volta-to-blackwell

But somehow the S1070 had already Massively-
Parallel, Many-Core Architecture, and forms
of MIMD, since it had 960 / 240 = 4 cores.

960 scalar processor cores (240 per GPU).
But possibly more resticted inside work
groups, than later NVIDIA Volta ff

architecture with independent thread state.

Bye

Disclaimer: The above is only a very rough
RTX 5090 spec. Its doesn't say what value
format and what vector/matrics ops were

used. Also energy consumption may vary.

Mild Shock schrieb:

Hi,

You are still chewing on SIMD. LoL

Ross Finlayson schrieb:

Then the idea is that any of those can be found and matched in
one "run", i.e. a stall-less, branch-less, call-less list of less than
a few or less than a few dozens or less than a few hundreds
instructions, the results "findings" in data and corresponding
"matchings" of expressions, that runs in less than one microsecond.

You cannot make the mental translation that if you have:

Ross Finlayson schrieb:

So, the context then is for register state and stack contents, that
the indicators of the above as "positive presence" then is to make
for that the adjustments to the offsets and extents and the shifts
is according to those, otherwise no-ops. Then the idea is that a

As independent logical thread state, that automatically MIMD follows?

Whats the problem to solve then?

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

So what does NUM_SHADERS = 4096 shaders mean here?

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Its only the number of logical threads.

CUDArao TEChNOLOGY UNLOCkS ThE POWER OF TESLA MANY-CORE PROCESSORS
The CUDA C compiler simplifies many-core programming
by enabling code development in a high-level language
and optimizing code to run on systems without knowledge of
how many cores are in the hardware.

CUDA applications automatically take advantage of more
cores or fewer cores in a system, so they can scale from
entry-level notebook GPUs to high end GPUs in technical
workstations and further into racks of GPUs in data
centers. This allows developers to

rCLcode oncerCY and deploy on a range of systems, as well as
scale forward in time as future GPUs deliver more
performance per watt and more cores per processor. The benefit
for software users is the opportunity to boost computing
performance simply by adding GPUs or using their

existing GPUs in new ways.
https://www.azken.com/download/Tesla_DS_S1070_EU.pdf

Bye

Mild Shock schrieb:

Hi,

Hurry Rossy Boy, the blue bus is waiting.
There is a quite a hyperbole from here:

Tesla S1070 in 2008
700 Watts , 1 Terra Flop
SOLVE TOMORROWrCOS PROBLEMS TODAY https://www.azken.com/download/Tesla_DS_S1070_EU.pdf

To here:

Blackwell GPU in 2026
575 Watts, 104.8 Terra Flops ( RTX 5090 )
From Volta To Blackwell https://newsletter.semianalysis.com/p/nvidia-tensor-core-evolution-from-volta-to-blackwell

But somehow the S1070 had already Massively-
Parallel, Many-Core Architecture, and forms
of MIMD, since it had 960 / 240 = 4 cores.

960 scalar processor cores (240 per GPU).
But possibly more resticted inside work
groups, than later NVIDIA Volta ff

architecture with independent thread state.

Bye

Disclaimer: The above is only a very rough
RTX 5090 spec. Its doesn't say what value
format and what vector/matrics ops were

used. Also energy consumption may vary.

Mild Shock schrieb:

Hi,

You are still chewing on SIMD. LoL

Ross Finlayson schrieb:

Then the idea is that any of those can be found and matched in
one "run", i.e. a stall-less, branch-less, call-less list of less than >> -a> a few or less than a few dozens or less than a few hundreds
instructions, the results "findings" in data and corresponding
"matchings" of expressions, that runs in less than one microsecond.

You cannot make the mental translation that if you have:

Ross Finlayson schrieb:

So, the context then is for register state and stack contents, that
the indicators of the above as "positive presence" then is to make
for that the adjustments to the offsets and extents and the shifts
is according to those, otherwise no-ops. Then the idea is that a

As independent logical thread state, that automatically MIMD follows?

Whats the problem to solve then?

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

Because of this parallelism you anyway
need to forget about any arithmetization
of product FSA (finite-state automata).

Just forget it. What modern GPU provide
is a kind of hirarchical viewpoint. You
can have barriers in groups etc..

So you can exercise control over your
mongolian horde of logical threads in
a kind of multilevel schema.

Have Fun!

Bye

Mild Shock schrieb:

Hi,

So what does NUM_SHADERS = 4096 shaders mean here?

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Its only the number of logical threads.

CUDArao TEChNOLOGY UNLOCkS ThE POWER OF TESLA MANY-CORE PROCESSORS
The CUDA C compiler simplifies-a many-core programming
by enabling code development in a high-level language
and optimizing code to run on systems without knowledge of
how many cores are in the hardware.

CUDA applications automatically take advantage of more
cores or fewer cores in a system, so they can scale from
entry-level notebook GPUs to high end GPUs in technical
workstations-a and further into racks of GPUs in data
centers. This allows developers to

rCLcode oncerCY and deploy on a range of systems, as well as
scale forward in time as future GPUs deliver more
performance per watt and more cores per processor. The benefit
for software users is the opportunity to boost computing
performance simply by adding GPUs or using their

existing GPUs in new ways. https://www.azken.com/download/Tesla_DS_S1070_EU.pdf

Bye

Mild Shock schrieb:

Hi,

Hurry Rossy Boy, the blue bus is waiting.
There is a quite a hyperbole from here:

Tesla S1070 in 2008
700 Watts , 1 Terra Flop
SOLVE TOMORROWrCOS PROBLEMS TODAY
https://www.azken.com/download/Tesla_DS_S1070_EU.pdf

To here:

Blackwell GPU in 2026
575 Watts, 104.8 Terra Flops ( RTX 5090 )
-aFrom Volta To Blackwell
https://newsletter.semianalysis.com/p/nvidia-tensor-core-evolution-from-volta-to-blackwell


But somehow the S1070 had already Massively-
Parallel, Many-Core Architecture, and forms
of MIMD, since it had 960 / 240 = 4 cores.

960 scalar processor cores (240 per GPU).
But possibly more resticted inside work
groups, than later NVIDIA Volta ff

architecture with independent thread state.

Bye

Disclaimer: The above is only a very rough
RTX 5090 spec. Its doesn't say what value
format and what vector/matrics ops were

used. Also energy consumption may vary.

Mild Shock schrieb:

Hi,

You are still chewing on SIMD. LoL

Ross Finlayson schrieb:

Then the idea is that any of those can be found and matched in
one "run", i.e. a stall-less, branch-less, call-less list of less

than

a few or less than a few dozens or less than a few hundreds
instructions, the results "findings" in data and corresponding
"matchings" of expressions, that runs in less than one microsecond.

You cannot make the mental translation that if you have:

Ross Finlayson schrieb:

So, the context then is for register state and stack contents, that
the indicators of the above as "positive presence" then is to make
for that the adjustments to the offsets and extents and the shifts
is according to those, otherwise no-ops. Then the idea is that a

As independent logical thread state, that automatically MIMD follows?

Whats the problem to solve then?

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.logic

Hi,

This was archived on Jul 9, 2026:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Still, Jul 29, Rossy Boy halucinates accusations:

Ross Finlayson schrieb:

.. bla bla goto bla bla ..

Stupid gangster: teamsters are a union.

In the trades, not the steals, ....

Woa! Thats now 20 days of brain desease,
and not understanding the meaning and implications.
Even not understand pi-WAM has Hack VM backend.

But its all opensource. Bravo Rossy Boy, you are
champion in brainlessness and lazyness of
a idiot usenet troll.

Bye

Mild Shock schrieb:

Hi,

If any of you guys do not understand what
is meant by or what the implications are:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Well I wouldn't care less. There are two
outcomes for numb nuts:

- Ignoramus: They don't understand it, but
-a they will understand it before they die.

- Ignorabimus: They don't understand it, and
-a will never understand it, and they die.

So who cares, its not my problem, you people
are stupid as fuck, and slow as fuck...

Bye

Mild Shock schrieb:

Hi,

Micro penis brain is in constant hiatus.
He can even not detect a trope.

LoL

Bye

Lane W schrieb:

Mild Shock wrote:

Hi,

My mother is worried that I fucked Lane W.
aka Micro Penis mother 24 hours straight.
She was screaming, basically singing all

the arias from operas that Luciano Pavarotti
usually sings. You Lane W. aka Micro Penis
should have heard it, since you

live in the basement of your mothers house.

No, actually remarkably, I don't. According to google I live 433

miles away from her.

Strike!

See, what i said about you was spot on.

What you said about me was generic and incorrect.

You really suck, man.

--- Synchronet 3.22a-Linux NewsLink 1.2