From Newsgroup: sci.physics.relativity

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
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.physics.relativity

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,

CAS and XADD have no looping, they
are atomic operations, that take some
time but basically have some outcome

with some ACID property and a result
value. What loops is the ADT, the Abstract
Data Type that you implement. Respectively

the client that uses the Abstract Data Type.
In your case you added the loop inside the
Abstract Data Type or lower level aggregate

code of a higher level operation:

Chris M. Thomasson wrote:
void producer(double state) {
uint32_t ver = XADD(&head, 1);
cell& c = cells[ver & (N - 1)];
while (LOAD(&c.ver) != ver) backoff(); /** Looping **/
c.state = state;
STORE(&c.ver, ver + 1);
}
https://groups.google.com/g/lock-free/c/acjQ3-89abE/m/a6-Di0GZsyEJ

In my case I added the loop during the client
usage of the ADT:

From: Mild Shock <janburse@fastmail.fm>
Subject: Source of the benchmark for DmitryVyukov
Date: Tue, 21 Jul 2026 01:44:21 +0200

private static void producer(Queue q) {
for (int i = 0; i < WORK; i++) {
Integer val = Integer.valueOf(i);
while (!enqueue(q, val)) ; /** Looping **/
}
}

Do you see the two loops, in your C code
and in my Java code? They are marked with a
comment /** Looping **/ .

You see them, don't you? But I don't know
exactly what backoff() does. Sometimes loops
are spurious yield loops, required because

an ADT cannot gurantee that every yield
implies a certain condition. This is for
example already found in the intrinsinc

monitor of Java, the wait(). You might consult
Doug Lea about the matter and how idiomatic
Java code looks like dealing with

spurious yields.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
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.physics.relativity

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,

CAS and XADD have no looping, they
are atomic operations, that take some
time but basically have some outcome

with some ACID property and a result
value. What loops is the ADT, the Abstract
Data Type that you implement. Respectively

the client that uses the Abstract Data Type.
In your case you added the loop inside the
Abstract Data Type or lower level aggregate

code of a higher level operation:

Chris M. Thomasson wrote:
void producer(double state) {
-a-a-a-a-a uint32_t ver = XADD(&head, 1);
-a-a-a-a-a cell& c = cells[ver & (N - 1)];
-a-a-a-a-a while (LOAD(&c.ver) != ver) backoff(); /** Looping **/
-a-a-a-a-a c.state = state;
-a-a-a-a-a STORE(&c.ver, ver + 1);
}
https://groups.google.com/g/lock-free/c/acjQ3-89abE/m/a6-Di0GZsyEJ

In my case I added the loop during the client
usage of the ADT:

From: Mild Shock <janburse@fastmail.fm>
Subject: Source of the benchmark for DmitryVyukov
Date: Tue, 21 Jul 2026 01:44:21 +0200

-a-a-a-a-a private static void producer(Queue q) {
-a-a-a-a-a-a-a-a-a for (int i = 0; i < WORK; i++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a Integer val = Integer.valueOf(i);
-a-a-a-a-a-a-a-a-a-a-a-a-a while (!enqueue(q, val)) ; /** Looping **/
-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a }

Do you see the two loops, in your C code
and in my Java code? They are marked with a
comment /** Looping **/ .

You see them, don't you? But I don't know
exactly what backoff() does. Sometimes loops
are spurious yield loops, required because

an ADT cannot gurantee that every yield
implies a certain condition. This is for
example already found in the intrinsinc

monitor of Java, the wait(). You might consult
Doug Lea about the matter and how idiomatic
Java code looks like dealing with

spurious yields.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
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.physics.relativity

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,

CAS and XADD have no looping, they
are atomic operations, that take some
time but basically have some outcome

with some ACID property and a result
value. What loops is the ADT, the Abstract
Data Type that you implement. Respectively

the client that uses the Abstract Data Type.
In your case you added the loop inside the
Abstract Data Type or lower level aggregate

code of a higher level operation:

Chris M. Thomasson wrote:
void producer(double state) {
-a-a-a-a-a uint32_t ver = XADD(&head, 1);
-a-a-a-a-a cell& c = cells[ver & (N - 1)];
-a-a-a-a-a while (LOAD(&c.ver) != ver) backoff(); /** Looping **/
-a-a-a-a-a c.state = state;
-a-a-a-a-a STORE(&c.ver, ver + 1);
}
https://groups.google.com/g/lock-free/c/acjQ3-89abE/m/a6-Di0GZsyEJ

In my case I added the loop during the client
usage of the ADT:

From: Mild Shock <janburse@fastmail.fm>
Subject: Source of the benchmark for DmitryVyukov
Date: Tue, 21 Jul 2026 01:44:21 +0200

-a-a-a-a-a private static void producer(Queue q) {
-a-a-a-a-a-a-a-a-a for (int i = 0; i < WORK; i++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a Integer val = Integer.valueOf(i);
-a-a-a-a-a-a-a-a-a-a-a-a-a while (!enqueue(q, val)) ; /** Looping **/
-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a }

Do you see the two loops, in your C code
and in my Java code? They are marked with a
comment /** Looping **/ .

You see them, don't you? But I don't know
exactly what backoff() does. Sometimes loops
are spurious yield loops, required because

an ADT cannot gurantee that every yield
implies a certain condition. This is for
example already found in the intrinsinc

monitor of Java, the wait(). You might consult
Doug Lea about the matter and how idiomatic
Java code looks like dealing with

spurious yields.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
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.physics.relativity

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,

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,

CAS and XADD have no looping, they
are atomic operations, that take some
time but basically have some outcome

with some ACID property and a result
value. What loops is the ADT, the Abstract
Data Type that you implement. Respectively

the client that uses the Abstract Data Type.
In your case you added the loop inside the
Abstract Data Type or lower level aggregate

code of a higher level operation:

Chris M. Thomasson wrote:
void producer(double state) {
-a-a-a-a-a uint32_t ver = XADD(&head, 1);
-a-a-a-a-a cell& c = cells[ver & (N - 1)];
-a-a-a-a-a while (LOAD(&c.ver) != ver) backoff(); /** Looping **/
-a-a-a-a-a c.state = state;
-a-a-a-a-a STORE(&c.ver, ver + 1);
}
https://groups.google.com/g/lock-free/c/acjQ3-89abE/m/a6-Di0GZsyEJ

In my case I added the loop during the client
usage of the ADT:

From: Mild Shock <janburse@fastmail.fm>
Subject: Source of the benchmark for DmitryVyukov
Date: Tue, 21 Jul 2026 01:44:21 +0200

-a-a-a-a-a private static void producer(Queue q) {
-a-a-a-a-a-a-a-a-a for (int i = 0; i < WORK; i++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a Integer val = Integer.valueOf(i);
-a-a-a-a-a-a-a-a-a-a-a-a-a while (!enqueue(q, val)) ; /** Looping **/ >>> -a>-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a }

Do you see the two loops, in your C code
and in my Java code? They are marked with a
comment /** Looping **/ .

You see them, don't you? But I don't know
exactly what backoff() does. Sometimes loops
are spurious yield loops, required because

an ADT cannot gurantee that every yield
implies a certain condition. This is for
example already found in the intrinsinc

monitor of Java, the wait(). You might consult
Doug Lea about the matter and how idiomatic
Java code looks like dealing with

spurious yields.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
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.physics.relativity

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,

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,

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,

CAS and XADD have no looping, they
are atomic operations, that take some
time but basically have some outcome

with some ACID property and a result
value. What loops is the ADT, the Abstract
Data Type that you implement. Respectively

the client that uses the Abstract Data Type.
In your case you added the loop inside the
Abstract Data Type or lower level aggregate

code of a higher level operation:

Chris M. Thomasson wrote:
void producer(double state) {
-a-a-a-a-a uint32_t ver = XADD(&head, 1);
-a-a-a-a-a cell& c = cells[ver & (N - 1)];
-a-a-a-a-a while (LOAD(&c.ver) != ver) backoff(); /** Looping **/
-a-a-a-a-a c.state = state;
-a-a-a-a-a STORE(&c.ver, ver + 1);
}
https://groups.google.com/g/lock-free/c/acjQ3-89abE/m/a6-Di0GZsyEJ >>>> -a>
In my case I added the loop during the client
usage of the ADT:

From: Mild Shock <janburse@fastmail.fm>
Subject: Source of the benchmark for DmitryVyukov
Date: Tue, 21 Jul 2026 01:44:21 +0200

-a-a-a-a-a private static void producer(Queue q) {
-a-a-a-a-a-a-a-a-a for (int i = 0; i < WORK; i++) {
-a-a-a-a-a-a-a-a-a-a-a-a-a Integer val = Integer.valueOf(i);
-a-a-a-a-a-a-a-a-a-a-a-a-a while (!enqueue(q, val)) ; /** Looping **/ >>>> -a>-a-a-a-a-a-a-a-a-a }
-a-a-a-a-a }

Do you see the two loops, in your C code
and in my Java code? They are marked with a
comment /** Looping **/ .

You see them, don't you? But I don't know
exactly what backoff() does. Sometimes loops
are spurious yield loops, required because

an ADT cannot gurantee that every yield
implies a certain condition. This is for
example already found in the intrinsinc

monitor of Java, the wait(). You might consult
Doug Lea about the matter and how idiomatic
Java code looks like dealing with

spurious yields.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
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.physics.relativity

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 strated
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.physics.relativity

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.physics.relativity

On 07/22/2026 11:43 PM, Mild Shock wrote:

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,

CAS and XADD have no looping, they
are atomic operations, that take some
time but basically have some outcome

with some ACID property and a result
value. What loops is the ADT, the Abstract
Data Type that you implement. Respectively

the client that uses the Abstract Data Type.
In your case you added the loop inside the
Abstract Data Type or lower level aggregate

code of a higher level operation:

Chris M. Thomasson wrote:
void producer(double state) {
uint32_t ver = XADD(&head, 1);
cell& c = cells[ver & (N - 1)];
while (LOAD(&c.ver) != ver) backoff(); /** Looping **/
c.state = state;
STORE(&c.ver, ver + 1);
}
https://groups.google.com/g/lock-free/c/acjQ3-89abE/m/a6-Di0GZsyEJ

In my case I added the loop during the client
usage of the ADT:

From: Mild Shock <janburse@fastmail.fm>
Subject: Source of the benchmark for DmitryVyukov
Date: Tue, 21 Jul 2026 01:44:21 +0200

private static void producer(Queue q) {
for (int i = 0; i < WORK; i++) {
Integer val = Integer.valueOf(i);
while (!enqueue(q, val)) ; /** Looping **/
}
}

Do you see the two loops, in your C code
and in my Java code? They are marked with a
comment /** Looping **/ .

You see them, don't you? But I don't know
exactly what backoff() does. Sometimes loops
are spurious yield loops, required because

an ADT cannot gurantee that every yield
implies a certain condition. This is for
example already found in the intrinsinc

monitor of Java, the wait(). You might consult
Doug Lea about the matter and how idiomatic
Java code looks like dealing with

spurious yields.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
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

Multiplexer/demultiplexer or mux/demux, num-nuts.


Shut Up


If you haven't heard of a mux/demux,
it's the basic foundation of all internetworking.

Like "why would one even need inetd",
or, "sockets" or "packets".


P.S. I've never asked you a question.


Shut Up


--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Mild Shock wrote:

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.

Russia is doing again historical blunder again after Gorbachev. sitting
with war criminal in the battle field is shame. war criminal Yanks never respect diplomacy, never comply any agreement. while US terrorist State continousely supplying arms and intelligence and continuing sanction whats
the point here diplomacy. Russian is suiciding politicaly. its declaration
of weakness. world must be more cautious resistance against American
hegemony. disaster
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

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 strated
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.physics.relativity

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 strated
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.physics.relativity

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 strated
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.physics.relativity

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.
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

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.physics.relativity

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.physics.relativity

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.physics.relativity

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.physics.relativity

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.physics.relativity

Mild Shock wrote:

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

sci.math, I need to warn you that Mild Shock is profoundly insecure. He
is entirely incapable of admitting fault, such as when he totally fucked
up by accusing me of living in my mother's basement. What a joke. I
haven't lived with my parents since 2003. Mild Shock is a total fuckup
and dreadfully misinformed + insecure.

sci.physics.relativity, I need to warn you that Mild Shock is
embarrassing and inferior. He is likely to spend all summer squabbling
about some generic accusations, without a shred of proof, that he will
level in my direction. Mild Shock has no idea who I am or what I
represent, and he is basically floundering, really similar in fashion to
how a stupid asshole does.

I'm living the American Dream.

He is more generic than an AI overview of the typical lowlife Gen X that
it will bring up on every screen in America. Goodness, how formulaic and blas|-.
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Ni,

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!

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 strated
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.physics.relativity

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

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

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:

Ni,

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!

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 strated
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.physics.relativity

Mild Shock wrote:

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

You are one of those cerebral asshats in the first episode of Star Trek.
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

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

Lane W schrieb:

Mild Shock wrote:

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

You are one of those cerebral asshats in the first episode of Star Trek.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Mild Shock wrote:

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

you fucking irrelevant indolent impertinent puerile imbecile. This guy
thinks shit is AI laptops. You are a shame to your country.
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Mild Shock wrote:

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

These tropes of yours would be funnier if they were closer to truth.
That's not even the right ballpark, Mild Shock. If I were an alpaca I
would spit right on your nose & mouth.
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

LoL

Bradford Babkoff schrieb:

Mild Shock wrote:

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

you fucking irrelevant indolent impertinent puerile imbecile. This guy
thinks shit is AI laptops. You are a shame to your country.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

LoL

Lane W schrieb:

Mild Shock wrote:

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

These tropes of yours would be funnier if they were closer to truth.
That's not even the right ballpark, Mild Shock. If I were an alpaca I
would spit right on your nose & mouth.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

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.physics.relativity

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 strated
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.physics.relativity

Mild Shock wrote:

LoL

Lane W schrieb:

Mild Shock wrote:

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

These tropes of yours would be funnier if they were closer to truth.
That's not even the right ballpark, Mild Shock. If I were an alpaca I
would spit right on your nose & mouth.

I don't see how you can dispute that you made two of the same post, one
right after the other. Not in the right ballpark? It's more exact than a geometry formula. That I called you an ass was not so far from the truth either?
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

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,

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

Its all on GitHub , for the 100-th time .
Just RTFM , i.e. study the repo and the
medim article. Just follow this link:

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

Whats wrong with you guys, did the AI boom
suck out all your braincells. I really have
no words for being that stupid and slow.

Bye

In particular the repo contains two versions
of a Hack VM, written in WebGPU / WGSL:

Hack VM: Version 1.0

https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/boot.mjs

Hack VM: Version 2.0

https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example64/boot2.mjs

Version 1.0 is for a single compute shader
expriment. And Version 2.o is for a multi
compute shader experiment.

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 strated
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.physics.relativity

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,

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

Its all on GitHub , for the 100-th time .
Just RTFM , i.e. study the repo and the
medim article. Just follow this link:

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

Whats wrong with you guys, did the AI boom
suck out all your braincells. I really have
no words for being that stupid and slow.

Bye

In particular the repo contains two versions
of a Hack VM, written in WebGPU / WGSL:

Hack VM: Version 1.0

https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/boot.mjs

Hack VM: Version 2.0

https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example64/boot2.mjs

Version 1.0 is for a single compute shader
expriment. And Version 2.o is for a multi
compute shader experiment.

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 strated
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.physics.relativity

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,

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

Its all on GitHub , for the 100-th time .
Just RTFM , i.e. study the repo and the
medim article. Just follow this link:

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

Whats wrong with you guys, did the AI boom
suck out all your braincells. I really have
no words for being that stupid and slow.

Bye

In particular the repo contains two versions
of a Hack VM, written in WebGPU / WGSL:

Hack VM: Version 1.0

https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/boot.mjs

Hack VM: Version 2.0

https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example64/boot2.mjs

Version 1.0 is for a single compute shader
expriment. And Version 2.o is for a multi
compute shader experiment.

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 strated
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.physics.relativity

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,

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

Its all on GitHub , for the 100-th time .
Just RTFM , i.e. study the repo and the
medim article. Just follow this link:

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

Whats wrong with you guys, did the AI boom
suck out all your braincells. I really have
no words for being that stupid and slow.

Bye

In particular the repo contains two versions
of a Hack VM, written in WebGPU / WGSL:

Hack VM: Version 1.0

https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/boot.mjs

Hack VM: Version 2.0

https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example64/boot2.mjs

Version 1.0 is for a single compute shader
expriment. And Version 2.o is for a multi
compute shader experiment.

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 strated
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.physics.relativity

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

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,

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

Its all on GitHub , for the 100-th time .
Just RTFM , i.e. study the repo and the
medim article. Just follow this link:

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

Whats wrong with you guys, did the AI boom
suck out all your braincells. I really have
no words for being that stupid and slow.

Bye

In particular the repo contains two versions
of a Hack VM, written in WebGPU / WGSL:

Hack VM: Version 1.0

https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example63/boot.mjs

Hack VM: Version 2.0

https://github.com/Jean-Luc-Picard-2021/gigabudget/blob/main/course/example64/boot2.mjs

Version 1.0 is for a single compute shader
expriment. And Version 2.o is for a multi
compute shader experiment.

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 strated
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

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.physics.relativity

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 public static void run() {
-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 int Y = (771-A)/3;
-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 int Z = (771-A-B)/2;
-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 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 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 711000000 % D == 0)
-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 }

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.physics.relativity

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 public static void run() {
-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 int Y = (771-A)/3;
-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 int Z = (771-A-B)/2;
-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 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 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 711000000 % D == 0) >>> -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 }

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.physics.relativity

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 public static void run() {
-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 int Y = (771-A)/3;
-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 int Z = (771-A-B)/2;
-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 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 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 711000000 % D == 0) >>>> -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 }

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.physics.relativity

On 07/27/2026 12:14 AM, Mild Shock wrote:

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.

Ah, then the arithmetic is a trick,
and the sock-puppet howler troll drools nonsense.


There are lots of tricks of arithmetic.

Systolic flow machines their ideas are around a long time.

Shut Up


--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

On 07/27/2026 12:51 AM, Mild Shock wrote:

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:

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.

Hopfield + Kohonen and some arithmetic coding,
dirt simple since the '90's, generative programming
and online psychiatrists are around since the '60's,
chat-bots are simple in the scheme of things,
and very widely varied in their implementation,
feedback-directed optimization,
pile arithmetic coding and call it vectors on big data
and say that's a requirement, when really it's just
_bloat_ and a sales-case for _bloat_ and it's bloated
and it _bloats_ thee. Bloater.


--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

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 public static void run() {
-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 int Y = (771-A)/3;
-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 int Z = (771-A-B)/2;
-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 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 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 711000000 % D == 0) >>> -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 }

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.physics.relativity

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 public static void run() {
-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 int Y = (771-A)/3;
-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 int Z = (771-A-B)/2;
-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 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 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 711000000 % D == 0) >>>> -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 }

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.physics.relativity

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 public static void run() {
-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 int Y = (771-A)/3;
-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 int Z = (771-A-B)/2;
-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 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 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 711000000 % D == 0) >>>>> -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 }

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.physics.relativity

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 public static void run() {
-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 int Y = (771-A)/3;
-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 int Z = (771-A-B)/2;
-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 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 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 711000000 % D == 0) >>>>>> -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 }

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.physics.relativity

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 strated
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.physics.relativity

Hi,

I guess Rossy Boys mother was so disappointed
in the 50's that her son didn't become the
next Einstein, physics was the ultimate 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

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):

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

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.physics.relativity

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.physics.relativity

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.physics.relativity

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 strated
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.physics.relativity

Mild Shock wrote:

And what LIPS means, drawn from TOPS,
in terms of applications? Shame on you guys!
You are a bunch of brainless idiots.

post the link, i want to buy one, then kiss my ass and fuck off. Git links
are for fools like you are.
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Hi,

Moron there is no SIMT. As I already wrote:

He is also not Zweistein, since he doesn't
understand concepts such as:

- NVIDIA Volta ff. architecture

But you had the SIMD and MIMD disctinction
alreay in OpenMP (via #pragma omp simd and
#pragma omp parallel):

Flynn's Taxonomy classifies computer architectures
according to how many instruction streams (processes)
and data streams they can process simultaneously,
dividing them into four categories:
SISD, SIMD, MISD, and MIMD. https://www.geeksforgeeks.org/computer-organization-architecture/computer-architecture-flynns-taxonomy/

Its not so difficult to understand what
the NVIDIA Volta ff. architecture means.

Bye

Ross Finlayson schrieb:

They're considered really quite simple,
each of those threads is simple, SIMT.

Ross Finlayson schrieb:

On 06/25/2021 07:54 PM, Archimedes Plutonium wrote:

On Monday, June 21, 2021 at 12:00:21 PM UTC-5, Graham Cooper wrote:

On Tuesday, June 22, 2021 at 2:54:40 AM UTC+10, burs...@gmail.com

wrote:

Try yourself:
misc.prolog.compound.parenthesis.missing

LMAO!

Jan you work too hard. nobody wants theorem provers on prolog

ASIMO tech is going to LISP which will just have a UNIFY routine

but people can LEARN PROLOG if you EFF OFF!



VOTE NOW! BAN JAN

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

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.physics.relativity

Hi,

They are the same:

Performance of the Cray T3D
https://arxiv.org/abs/hep-lat/9509003v1

GPU Backend: Find 0xCAFFEE with -C-WAM
https://medium.com/2989/8890efd3503c

Both Cray T3D as installed at PSC, and
the on chip GPU of my Ryzen AI 7 350
w/ Radeon 860M Laptop for ca. 1000 CHF.

they both have MIMD (Multiple instruction,
multiple data) and 512 PE (Processing Elements).
Quite amazing what happend in 30 years of

Very-large-scale integration (VLSI).

LoL

Bye

Mild Shock schrieb:

Ni,

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!

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 strated
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.physics.relativity

On 08/05/2026 05:23 AM, Mild Shock wrote:

Hi,

They are the same:

Performance of the Cray T3D
https://arxiv.org/abs/hep-lat/9509003v1

GPU Backend: Find 0xCAFFEE with -C-WAM
https://medium.com/2989/8890efd3503c

Both Cray T3D as installed at PSC, and
the on chip GPU of my Ryzen AI 7 350
w/ Radeon 860M Laptop for ca. 1000 CHF.

they both have MIMD (Multiple instruction,
multiple data) and 512 PE (Processing Elements).
Quite amazing what happend in 30 years of

Very-large-scale integration (VLSI).

LoL

Bye

Mild Shock schrieb:

Ni,

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!

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 strated
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

LoL = cretin rictus


"Hugging Face" evokes the face-hugger from Aliens
or the head-crab from Half-Life, either of which
is a parasite that infests then consumes the host,
resulting either xenomorph-host bastards or the
headcrab-zombies. It's like those villains from
Alpha Flight or rather Excalibur comic book #1, they're
evil and consume people then steal their skins. In some sense
it's like the "Soul Eater" from Daredevil or other
usual accounts of the "soul vampire", from the 1980's,
eg the narcissistic, manipulative types, the "Warwolves".

https://marvel.fandom.com/wiki/Warwolves

About as relevant as recycling buzzwords
and cranking-the-hype-machine.


If one's interested in parallelization according to
language, then besides the like of array-languages
which are organized to result products however they're
so done, for example the "embarrassingly parallel",
where MIMD to some types is just a bucket of a worker pool,
there's that "occam" of the "Transputer" which has in
its instruction architecture directives about the
parallel, in small computers vis-a-vis the sort of
super-scalar super-computers, in the 1980's, since
the 1980's.



Leibniz makes a neat account of why there's "doubly-objective
relativity theory", after "relational theory of space" since
"relativity theory" has been around forever while reason was
trying to figure out whether "real space" exists, yet though that
FitzGerald's "space contraction", which is both length-contraction and time-dilation together, is considered the seminal account,
with "real space-contraction", for which there's room in
the formalism and room in the data, rooom in the theory.



Shut Up, I'm Not Listening, Shut Up







--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Hi,

How it started, NVIDIA being cool:

NCCL provides routines such as all-gather,
all-reduce, broadcast, reduce, reduce-scatter,
and point-to-point send and receive. These
routines are optimized to achieve high
bandwidth and low latency over PCIe,
NVIDIA NVLinkrao, and other high-speed
interconnects within a node and over
NVIDIA networking across nodes.
https://developer.nvidia.com/nccl

How its going, vLLM trying to be cool:

[RFC]: Native Weight Syncing APIs
However, there are no standardized methods for
performing online weight syncing. Open source projects
like SkyRL, VeRL, and TRL need to include their
own implementations of the weight syncing
infrastructure, leading to added complexity
for developers seeking to adopt vLLM as their
inference server for post-training workloads. https://github.com/vllm-project/vllm/issues/31848

How much Workers are enough? I guess it depends
on I/O parallelism, CPU Memory parallelism, CPU
Processing parallelism, and now also

GPU Memory parallelism and GPU Processing
parallelism, and last but least you might have
a couple DMAs sitting here and there,

or even invoking a sort of RDMA. Quite amazing!

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
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