From Newsgroup: sci.physics.relativity
Hi,
On could believe the AI boom is a kind of
Charles Darvin Galapagos Island Evolution
Trick of repurposing FFT hardware.
But this is of course not true, HPC, high
performance computing, has already defined
level 3 ops years ago.
But look at this rabit hole of Ryzen AI 7 350
NPU design, which is a stripped down Xilinx,
stripped of exotic FFT features:
Getting peak TOPS on a Ryzen AI 7 350 NPU
https://destevez.net/2026/05/getting-peak-tops-on-a-ryzen-ai-7-350-npu/
But the core feature, very long instruction
word (VLIW) engines, with hardware accelerated
GEMMs, scattered in grids of ASIC tiles,
connected by DMA and NoC, is even not very
specific to AMD, you find it also in Snapdragon /
Qualcomm SoCs for AI Laptops.
Bye
P.S.: My brain playing tricks, why should I
name a library(ironpaw) ? From the same
article above. Maybe WebNN is easier to use?
"mlir-aie contains a Python framework called
IRON that generates LLVM MLIR code representing
a workload that runs on the NPU, including the
code that runs on each compute tile processor
and the configuration of DMAs and other hardware.
Kernels for the compute tile processor can be
written in C++ and compiled either with the
open-source llvm-aie Peano compiler, which is
a fork of LLVM that adds support for the Xilinx
AI engine processors, or with the closed-source
Xilinx CHESS compiler, which is included in Vitis.
In simple cases the kernels can also be directly
written in Python with IRON."
Getting peak TOPS on a Ryzen AI 7 350 NPU
https://destevez.net/2026/05/getting-peak-tops-on-a-ryzen-ai-7-350-npu/
Mild Shock schrieb:
Hi,
Mostlikely for high performance computing |a la,
the Actor/Erlang model is dead, they might rely
on MPMC (Multiple Producer, Multiple Consumer)
queue entities separate from the threads. The
ISO Prolog multi-threading support had also such
threads. But besides that was also Actor/Erlang
leaning in practice, like SWI, where threads
have some default queues. So an actor is basically
a Thread and Mailbox conflation. While a MPMC queue
is a kind of separate Mailbox, where multiple
"actors" can read from and write from. A kind of
localized Linda Tuple store.
Which Programming language did adopted the
non-Actor pi-calculus model? Right golang
with its channels.
Bye
Mild Shock schrieb:
Hi,
Usual question:
Why implement both pre-emptive threading
AND cooperative tasks/engines?
I had implemented the ISO proposal in formerly Jekejeke
Prolog, you find the ISO proposal here:
ISO/IEC DTR 13211rCo5:2007
Prolog multi-threading support
https://logtalk.org/plstd/threads.pdf
But the ISO proposal doesn't match modern WebGPU APIs,
where your logical threads can live remotely in a dedicated GPU
in the VRAM there, and where you would have launch
parameters that say: Hey please run 4096 compute
shaders for me, that have independet thread state. Using
cooperative multi-tasking as the orchestrator works well.
Bye
Mild Shock schrieb:
Hi,
Why does this Lama have a red pyjama.
Oh, its a baby Lama. Its still in the cradle
and needs some training:
RedPajama-Data-v2
https://github.com/togethercomputer/RedPajama-Data
But then Andrej Karpathy recently showed
GPT-2 training on rented GPUs for less
than 100 USD in less then 2 hours.
So where do these grown up Lamas go.
Well Georgi Gerganov prefered C++/C
when he shouted Llama Llama Red Pyjama.
But you also find WebLLM, wrapping the
underlying C++/C GPU interface via the
W3C standard WebGPU / WGSL, with JavaScript:
In-Browser LLM Inference Engine
https://webllm.mlc.ai/
My experience with WebLLM 6 months
ago on an iPad Pro 2024, still a little early
stage performance and robustness.
But hey hardware of AI mobile iGPUs is
still evolving, and AI laptop, AI smartphones
and AI tablets, will soon feature Chinese
hardware such some new Kirin AI in 2027.
Bye
Mild Shock schrieb:
Hi,
Remember when first all local AI was Python
and PyTorch APIs. And then suddently people 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