Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.lang.prolog > #15710 > unrolled thread

The Wuhan Virus that destroyed Python [ggml Manifesto]

Back to article view | Back to comp.lang.prolog

Contents

  The Wuhan Virus that destroyed Python [ggml Manifesto] Mild Shock <janburse@fastmail.fm> - 2026-07-22 21:01 +0200
    Deadlock Exorcism: Switch from Push to Pull [A pi-calculus Specification of Prolog] (Re: The Wuhan Virus that destroyed Python [ggml Manifesto]) Mild Shock <janburse@fastmail.fm> - 2026-07-23 00:25 +0200
      Why do you even need a mpmc queue? [Thunder Kittens] (Re: Deadlock Exorcism: Switch from Push to Pull) Mild Shock <janburse@fastmail.fm> - 2026-07-23 08:45 +0200
        Trivial balancing example for (int i=0; i<global_id; i++) (Re: Why do you even need a mpmc queue? [Thunder Kittens]) Mild Shock <janburse@fastmail.fm> - 2026-07-23 08:55 +0200
          The Pixel Phone AI Experiment Song (Enqueue/dequeue need not be fast and can spinn ["fairness" questions]) Mild Shock <janburse@fastmail.fm> - 2026-07-23 09:19 +0200
            Enqueue/dequeue need not be fast and can spinn ["fairness" questions] (Re: The Pixel Phone AI Experiment Song (Enqueue/dequeue need not be fast and can spinn ["fairness" questions]) Mild Shock <janburse@fastmail.fm> - 2026-07-23 09:23 +0200
        And, where did I talk about rockets? [Hint its about xAI's Grok] (Re: Why do you even need a mpmc queue? [Thunder Kittens]) Mild Shock <janburse@fastmail.fm> - 2026-07-25 01:25 +0200
          Why forget something, that was never on my mind (Re: And, where did I talk about rockets? [Hint its about xAI's Grok]) Mild Shock <janburse@fastmail.fm> - 2026-07-25 09:49 +0200
        Example Mandel Brot rendering [Faster with MIMD] (Was: Why do you even need a mpmc queue? [Thunder Kittens]) Mild Shock <janburse@fastmail.fm> - 2026-07-25 09:56 +0200
    Potential Python Recovery: Free Threading [3.13 release] (Re: The Wuhan Virus that destroyed Python [ggml Manifesto]) Mild Shock <janburse@fastmail.fm> - 2026-07-23 10:21 +0200
    The things XILINX braught to the AMD table (Re: The Wuhan Virus that destroyed Python [ggml Manifesto]) Mild Shock <janburse@fastmail.fm> - 2026-07-23 18:48 +0200
      NVIDIA evacuated its Chinese market [Tau Scaling] (Re: The things XILINX braught to the AMD table) Mild Shock <janburse@fastmail.fm> - 2026-07-23 19:13 +0200
        Micro penis mother sung arias (Re: NVIDIA evacuated its Chinese market [Tau Scaling]) Mild Shock <janburse@fastmail.fm> - 2026-07-24 14:40 +0200
          Micro penis brain is in constant hiatus (Re: Micro penis mother sung arias) Mild Shock <janburse@fastmail.fm> - 2026-07-24 15:27 +0200
            Ignoramus or Ignorabimus: I don't care [(Re: Micro penis brain is in constant hiatus (Re: Micro penis mother sung arias) Mild Shock <janburse@fastmail.fm> - 2026-07-24 15:35 +0200
              You are a moron, brainless putin payed (Re: Ignoramus or Ignorabimus: I don't care) Mild Shock <janburse@fastmail.fm> - 2026-07-24 18:01 +0200
                Yeah keep reading my posts, uninspired fool (Re: You are a moron, brainless putin payed) Mild Shock <janburse@fastmail.fm> - 2026-07-24 19:47 +0200
              Out of the blue accusation span 15 days [Empirical USENET study] (Re: Ignoramus or Ignorabimus: I don't care) Mild Shock <janburse@fastmail.fm> - 2026-07-24 20:27 +0200
    Little Data Center on Your Palm [AI Laptops for 500 USD] (Re: The Wuhan Virus that destroyed Python [ggml Manifesto]) Mild Shock <janburse@fastmail.fm> - 2026-07-24 17:59 +0200
      2008: 4 Blades + Tesla S1070 versus 2026: 1 AI Laptop (Re: Little Data Center on Your Palm [AI Laptops for 500 USD]) Mild Shock <janburse@fastmail.fm> - 2026-07-24 18:15 +0200
    Hurry the blue bus doesnt stop indefinitely (Re: The Wuhan Virus that destroyed Python [ggml Manifesto]) Mild Shock <janburse@fastmail.fm> - 2026-07-24 20:37 +0200
      Not SIMD, a MIMD design for NVIDIA Volta (Re: Hurry the blue bus doesnt stop indefinitely) Mild Shock <janburse@fastmail.fm> - 2026-07-24 20:58 +0200
        Could take 3-4 months find machine / browser (Re: Not SIMD, a MIMD design for NVIDIA Volta) Mild Shock <janburse@fastmail.fm> - 2026-07-24 21:16 +0200
        The Koan of pi-WAM queues [FORTRAN-S] (Re: Not SIMD, a MIMD design for NVIDIA Volta) Mild Shock <janburse@fastmail.fm> - 2026-07-26 19:54 +0200
          The turbo capping of AI Laptops (Was: The Koan of pi-WAM queues [FORTRAN-S]) Mild Shock <janburse@fastmail.fm> - 2026-07-26 20:00 +0200
          Re: The Koan of pi-WAM queues [FORTRAN-S] (Re: Not SIMD, a MIMD design for NVIDIA Volta) Mild Shock <janburse@fastmail.fm> - 2026-07-27 09:16 +0200
          Why forget Bulgarians, never on my mind (Re: The Koan of pi-WAM queues [FORTRAN-S]) Mild Shock <janburse@fastmail.fm> - 2026-07-27 09:16 +0200
            miniTriton CUDA is an alternative to torch variants (Re: Why forget Bulgarians, never on my mind) Mild Shock <janburse@fastmail.fm> - 2026-07-27 09:52 +0200
              Andrej Karpathy original gangster of Budget Laptop (Re: miniTriton CUDA is an alternative to torch variants) Mild Shock <janburse@fastmail.fm> - 2026-07-27 09:54 +0200
            The evolution of hardware and GPT-2 training (Re: Why forget Bulgarians, never on my mind) Mild Shock <janburse@fastmail.fm> - 2026-07-27 10:57 +0200
              How speed up π-WAM with vector operations (Re: The evolution of hardware and GPT-2 training) Mild Shock <janburse@fastmail.fm> - 2026-07-27 11:10 +0200
                AI accelerator extend from GPU to CPU [Zero Copying] (Re: How speed up π-WAM with vector operations) Mild Shock <janburse@fastmail.fm> - 2026-07-27 13:21 +0200
                  The invention of vector and matrix registers [NVIDIA Volta] (Re: AI accelerator extend from GPU to CPU [Zero Copying]) Mild Shock <janburse@fastmail.fm> - 2026-07-27 13:22 +0200
                    Re: The invention of vector and matrix registers [NVIDIA Volta] (Re: AI accelerator extend from GPU to CPU [Zero Copying]) Ross Finlayson <ross.a.finlayson@gmail.com> - 2026-07-27 07:34 -0700
                      Maybe they should have named it NVIDIA Einstein [Rossy Boy Toe Sucking] (Was: The invention of vector and matrix registers [NVIDIA Volta]) Mild Shock <janburse@fastmail.fm> - 2026-07-27 17:12 +0200
                      Re: The invention of vector and matrix registers [NVIDIA Volta] (Re: AI accelerator extend from GPU to CPU [Zero Copying]) R Kym Horsell <kym@sdf.com> - 2026-07-27 15:43 +0000
                        Re: The invention of vector and matrix registers [NVIDIA Volta] (Re: AI accelerator extend from GPU to CPU [Zero Copying]) R Kym Horsell <kymhorsell@gmail.com> - 2026-07-27 15:46 +0000
                        π-WAM is not adding decimals, it is removing decimals (Was: The invention of vector and matrix registers [NVIDIA Volta]) Mild Shock <janburse@fastmail.fm> - 2026-07-27 18:34 +0200
      Potato Computer owner impressed by Ukraine Tech [Rossy Boys Brother?] (Was: Hurry the blue bus doesnt stop indefinitely) Mild Shock <janburse@fastmail.fm> - 2026-07-27 16:56 +0200
        Rossy Boy is neither Einstein nor Zweistein (Was: Potato Computer owner impressed by Ukraine Tech) Mild Shock <janburse@fastmail.fm> - 2026-07-27 18:25 +0200

Page 2 of 2 — ← Prev page 1 [2]

#15736 — Hurry the blue bus doesnt stop indefinitely (Re: The Wuhan Virus that destroyed Python [ggml Manifesto])

Hi,

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

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

Bye

 > Hi,
 >
 > Again I posted everything here:
 >
 >> 11.4 Giga Lips with a Budget Laptop
 >> https://github.com/Jean-Luc-Picard-2021/gigabudget
 >
 > The repo says, same time when I posted
 > the link first time:
 >
 >> This repository was archived by the
 >> owner on Jul 9, 2026. It is now read-only.
 >
 > Now a USENET user, who had already entitled
 > himself for a couple of irrational accusations
 >
 > towards my side, is asking this question:
 >
 > Chris M. Thomasson schrieb, Jul 24, 2026
 >> Show an outline of what you
 >> need you compute shader to do?
 >
 > Bravo, thats a delay of a wooping 15 days.
 >
 > Bye

Mild Shock schrieb:
> Hi,
> 
> Remember when first all local AI was Python
> and PyTorch APIs. And then suddently people started
> using bare metal C/C++ Code. Here is the story:
> 
> How it started:
> 
> GPT-J or GPT-J-6B is an open-source large
> language model (LLM) developed by EleutherAI
> in 2021. As the name suggests, it is a
> generative pre-trained transformer model
> designed to produce human-like text that
> continues from a prompt.
> https://www.eleuther.ai/
> 
> How it was going [Georgi Gerganov]:
> 
> So a few days later comes out the LLaMA, I do
> some calculations and I figure out “Okay, 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?”
> 
> Why I was able to do it so quickly - basically,
> for all that I saw it’s pretty much GPT-J architecture
> with some modifications, like some extra memorization
> layers. It’s 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

[toc] | [prev] | [next] | [standalone]

#15737 — Not SIMD, a MIMD design for NVIDIA Volta (Re: Hurry the blue bus doesnt stop indefinitely)

Hi,

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

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

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

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

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

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

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

not a separate GPU card.

Bye

Mild Shock schrieb:
> Hi,
> 
> Ride the snake
> He's old and his skin is cold
> The west is the best
> The west is the best
> Get here and we'll do the rest
> The blue bus is calling us
> The blue bus is calling us
> Driver, where you taking us?
> 
> Apocalypse Now intro: The Doors, The End {1979}
> https://www.youtube.com/watch?v=CIrvSJwwJUE
> 
> Bye
> 
>  > Hi,
>  >
>  > Again I posted everything here:
>  >
>  >> 11.4 Giga Lips with a Budget Laptop
>  >> https://github.com/Jean-Luc-Picard-2021/gigabudget
>  >
>  > The repo says, same time when I posted
>  > the link first time:
>  >
>  >> This repository was archived by the
>  >> owner on Jul 9, 2026. It is now read-only.
>  >
>  > Now a USENET user, who had already entitled
>  > himself for a couple of irrational accusations
>  >
>  > towards my side, is asking this question:
>  >
>  > Chris M. Thomasson schrieb, Jul 24, 2026
>  >> Show an outline of what you
>  >> need you compute shader to do?
>  >
>  > Bravo, thats a delay of a wooping 15 days.
>  >
>  > Bye
> 
> Mild Shock schrieb:
>> Hi,
>>
>> Remember when first all local AI was Python
>> and PyTorch APIs. And then suddently people started
>> using bare metal C/C++ Code. Here is the story:
>>
>> How it started:
>>
>> GPT-J or GPT-J-6B is an open-source large
>> language model (LLM) developed by EleutherAI
>> in 2021. As the name suggests, it is a
>> generative pre-trained transformer model
>> designed to produce human-like text that
>> continues from a prompt.
>> https://www.eleuther.ai/
>>
>> How it was going [Georgi Gerganov]:
>>
>> So a few days later comes out the LLaMA, I do
>> some calculations and I figure out “Okay, 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?”
>>
>> Why I was able to do it so quickly - basically,
>> for all that I saw it’s pretty much GPT-J architecture
>> with some modifications, like some extra memorization
>> layers. It’s 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
> 

[toc] | [prev] | [next] | [standalone]

#15738 — Could take 3-4 months find machine / browser (Re: Not SIMD, a MIMD design for NVIDIA Volta)

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

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

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

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

Bye

Mild Shock schrieb:
> Hi,
> 
> Its not tested on some Single Instruction/
> Multiple Data (SIMD) GPU. It was only tested on
> AI Laptops with Multiple instruction, Multiple
> 
> Data (GPU) architecture for the scalar registers
> per logical thread. As introduced by NVIDIA Volta
> in around 2017:
> 
>  > the first product was not announced until May 2017
>  > https://en.wikipedia.org/wiki/Volta_%28microarchitecture%29
> 
> Although I wrote the code of Hack VM with SIMD
> in mind, I never tested it on a pure SIMD GPU,
> and I never ported boot.mjs or boot2.mjs to
> 
> WebGL2 / GLSL. I uploaded WebGPU / WGSL. Among the
> tester I had were these AI Laptops, that could all
> run WebGPU / WGSL in a browser:
> 
>  > Intel(R) Core(TM) Ultra 7 258V
>  > AMD Ryzen AI 7 350 w/ Radeon 860M
>  > Apple A18 Pro, Darwin Kernel Version 25.5.0
>  > Snapdragon(R) X - X126100 - Qualcomm(R) Oryon(TM) CPU
> 
> Some AI Laptops had WebGPU / WGSL still behind
> a browser flag, since its relatively new on ARM.
> Also the above AI Laptops have all a iGPU and
> 
> not a separate GPU card.
> 
> Bye
> 
> Mild Shock schrieb:
>> Hi,
>>
>> Ride the snake
>> He's old and his skin is cold
>> The west is the best
>> The west is the best
>> Get here and we'll do the rest
>> The blue bus is calling us
>> The blue bus is calling us
>> Driver, where you taking us?
>>
>> Apocalypse Now intro: The Doors, The End {1979}
>> https://www.youtube.com/watch?v=CIrvSJwwJUE
>>
>> Bye
>>
>>  > Hi,
>>  >
>>  > Again I posted everything here:
>>  >
>>  >> 11.4 Giga Lips with a Budget Laptop
>>  >> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>  >
>>  > The repo says, same time when I posted
>>  > the link first time:
>>  >
>>  >> This repository was archived by the
>>  >> owner on Jul 9, 2026. It is now read-only.
>>  >
>>  > Now a USENET user, who had already entitled
>>  > himself for a couple of irrational accusations
>>  >
>>  > towards my side, is asking this question:
>>  >
>>  > Chris M. Thomasson schrieb, Jul 24, 2026
>>  >> Show an outline of what you
>>  >> need you compute shader to do?
>>  >
>>  > Bravo, thats a delay of a wooping 15 days.
>>  >
>>  > Bye
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>> Remember when first all local AI was Python
>>> and PyTorch APIs. And then suddently people started
>>> using bare metal C/C++ Code. Here is the story:
>>>
>>> How it started:
>>>
>>> GPT-J or GPT-J-6B is an open-source large
>>> language model (LLM) developed by EleutherAI
>>> in 2021. As the name suggests, it is a
>>> generative pre-trained transformer model
>>> designed to produce human-like text that
>>> continues from a prompt.
>>> https://www.eleuther.ai/
>>>
>>> How it was going [Georgi Gerganov]:
>>>
>>> So a few days later comes out the LLaMA, I do
>>> some calculations and I figure out “Okay, 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?”
>>>
>>> Why I was able to do it so quickly - basically,
>>> for all that I saw it’s pretty much GPT-J architecture
>>> with some modifications, like some extra memorization
>>> layers. It’s 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
>>
> 

[toc] | [prev] | [next] | [standalone]

#15743 — The Koan of pi-WAM queues [FORTRAN-S] (Re: Not SIMD, a MIMD design for NVIDIA Volta)

Hi,

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

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

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

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

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

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

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

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

Not great, not terrible!

Bye

Mild Shock schrieb:
> Hi,
> 
> Its not tested on some Single Instruction/
> Multiple Data (SIMD) GPU. It was only tested on
> AI Laptops with Multiple instruction, Multiple
> 
> Data (GPU) architecture for the scalar registers
> per logical thread. As introduced by NVIDIA Volta
> in around 2017:
> 
>  > the first product was not announced until May 2017
>  > https://en.wikipedia.org/wiki/Volta_%28microarchitecture%29
> 
> Although I wrote the code of Hack VM with SIMD
> in mind, I never tested it on a pure SIMD GPU,
> and I never ported boot.mjs or boot2.mjs to
> 
> WebGL2 / GLSL. I uploaded WebGPU / WGSL. Among the
> tester I had were these AI Laptops, that could all
> run WebGPU / WGSL in a browser:
> 
>  > Intel(R) Core(TM) Ultra 7 258V
>  > AMD Ryzen AI 7 350 w/ Radeon 860M
>  > Apple A18 Pro, Darwin Kernel Version 25.5.0
>  > Snapdragon(R) X - X126100 - Qualcomm(R) Oryon(TM) CPU
> 
> Some AI Laptops had WebGPU / WGSL still behind
> a browser flag, since its relatively new on ARM.
> Also the above AI Laptops have all a iGPU and
> 
> not a separate GPU card.
> 
> Bye
> 
> Mild Shock schrieb:
>> Hi,
>>
>> Ride the snake
>> He's old and his skin is cold
>> The west is the best
>> The west is the best
>> Get here and we'll do the rest
>> The blue bus is calling us
>> The blue bus is calling us
>> Driver, where you taking us?
>>
>> Apocalypse Now intro: The Doors, The End {1979}
>> https://www.youtube.com/watch?v=CIrvSJwwJUE
>>
>> Bye
>>
>>  > Hi,
>>  >
>>  > Again I posted everything here:
>>  >
>>  >> 11.4 Giga Lips with a Budget Laptop
>>  >> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>  >
>>  > The repo says, same time when I posted
>>  > the link first time:
>>  >
>>  >> This repository was archived by the
>>  >> owner on Jul 9, 2026. It is now read-only.
>>  >
>>  > Now a USENET user, who had already entitled
>>  > himself for a couple of irrational accusations
>>  >
>>  > towards my side, is asking this question:
>>  >
>>  > Chris M. Thomasson schrieb, Jul 24, 2026
>>  >> Show an outline of what you
>>  >> need you compute shader to do?
>>  >
>>  > Bravo, thats a delay of a wooping 15 days.
>>  >
>>  > Bye
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>> Remember when first all local AI was Python
>>> and PyTorch APIs. And then suddently people started
>>> using bare metal C/C++ Code. Here is the story:
>>>
>>> How it started:
>>>
>>> GPT-J or GPT-J-6B is an open-source large
>>> language model (LLM) developed by EleutherAI
>>> in 2021. As the name suggests, it is a
>>> generative pre-trained transformer model
>>> designed to produce human-like text that
>>> continues from a prompt.
>>> https://www.eleuther.ai/
>>>
>>> How it was going [Georgi Gerganov]:
>>>
>>> So a few days later comes out the LLaMA, I do
>>> some calculations and I figure out “Okay, 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?”
>>>
>>> Why I was able to do it so quickly - basically,
>>> for all that I saw it’s pretty much GPT-J architecture
>>> with some modifications, like some extra memorization
>>> layers. It’s 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
>>
> 

[toc] | [prev] | [next] | [standalone]

#15744 — The turbo capping of AI Laptops (Was: The Koan of pi-WAM queues [FORTRAN-S])

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:
> 
>      public static void run() {
>          for (int A = 1; A < 192; A++) {
>              int Y = (771-A)/3;
>              for (int B = A; B < Y; B++) {
>                  int Z = (771-A-B)/2;
>                  for (int C = B; C < Z; C++) {
>                      int D = 711-A-B-C;
>                      if (A*B*C == 711000000/D &&
>                            711000000 % D == 0)
>      System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
>                  }
>              }
>          }
>      }
> 
> And then compare it to ElevenMulti, doing some
> Work Balancing Scheduler Tetris Game with 8 cores:
> 
> ElevenSingle
> A=120, B=125, C=150, D=316
> 6.628 ms
> 
> ElevenMulti
> A=120, B=125, C=150, D=316
> 1.941 ms
> 
> Not great, not terrible!
> 
> Bye
> 
> Mild Shock schrieb:
>> Hi,
>>
>> Its not tested on some Single Instruction/
>> Multiple Data (SIMD) GPU. It was only tested on
>> AI Laptops with Multiple instruction, Multiple
>>
>> Data (GPU) architecture for the scalar registers
>> per logical thread. As introduced by NVIDIA Volta
>> in around 2017:
>>
>>  > the first product was not announced until May 2017
>>  > https://en.wikipedia.org/wiki/Volta_%28microarchitecture%29
>>
>> Although I wrote the code of Hack VM with SIMD
>> in mind, I never tested it on a pure SIMD GPU,
>> and I never ported boot.mjs or boot2.mjs to
>>
>> WebGL2 / GLSL. I uploaded WebGPU / WGSL. Among the
>> tester I had were these AI Laptops, that could all
>> run WebGPU / WGSL in a browser:
>>
>>  > Intel(R) Core(TM) Ultra 7 258V
>>  > AMD Ryzen AI 7 350 w/ Radeon 860M
>>  > Apple A18 Pro, Darwin Kernel Version 25.5.0
>>  > Snapdragon(R) X - X126100 - Qualcomm(R) Oryon(TM) CPU
>>
>> Some AI Laptops had WebGPU / WGSL still behind
>> a browser flag, since its relatively new on ARM.
>> Also the above AI Laptops have all a iGPU and
>>
>> not a separate GPU card.
>>
>> Bye
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>> Ride the snake
>>> He's old and his skin is cold
>>> The west is the best
>>> The west is the best
>>> Get here and we'll do the rest
>>> The blue bus is calling us
>>> The blue bus is calling us
>>> Driver, where you taking us?
>>>
>>> Apocalypse Now intro: The Doors, The End {1979}
>>> https://www.youtube.com/watch?v=CIrvSJwwJUE
>>>
>>> Bye
>>>
>>>  > Hi,
>>>  >
>>>  > Again I posted everything here:
>>>  >
>>>  >> 11.4 Giga Lips with a Budget Laptop
>>>  >> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>>  >
>>>  > The repo says, same time when I posted
>>>  > the link first time:
>>>  >
>>>  >> This repository was archived by the
>>>  >> owner on Jul 9, 2026. It is now read-only.
>>>  >
>>>  > Now a USENET user, who had already entitled
>>>  > himself for a couple of irrational accusations
>>>  >
>>>  > towards my side, is asking this question:
>>>  >
>>>  > Chris M. Thomasson schrieb, Jul 24, 2026
>>>  >> Show an outline of what you
>>>  >> need you compute shader to do?
>>>  >
>>>  > Bravo, thats a delay of a wooping 15 days.
>>>  >
>>>  > Bye
>>>
>>> Mild Shock schrieb:
>>>> Hi,
>>>>
>>>> Remember when first all local AI was Python
>>>> and PyTorch APIs. And then suddently people started
>>>> using bare metal C/C++ Code. Here is the story:
>>>>
>>>> How it started:
>>>>
>>>> GPT-J or GPT-J-6B is an open-source large
>>>> language model (LLM) developed by EleutherAI
>>>> in 2021. As the name suggests, it is a
>>>> generative pre-trained transformer model
>>>> designed to produce human-like text that
>>>> continues from a prompt.
>>>> https://www.eleuther.ai/
>>>>
>>>> How it was going [Georgi Gerganov]:
>>>>
>>>> So a few days later comes out the LLaMA, I do
>>>> some calculations and I figure out “Okay, 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?”
>>>>
>>>> Why I was able to do it so quickly - basically,
>>>> for all that I saw it’s pretty much GPT-J architecture
>>>> with some modifications, like some extra memorization
>>>> layers. It’s 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
>>>
>>
> 

[toc] | [prev] | [next] | [standalone]

#15745 — Re: The Koan of pi-WAM queues [FORTRAN-S] (Re: Not SIMD, a MIMD design for NVIDIA Volta)

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.
 >
 >

[toc] | [prev] | [next] | [standalone]

#15746 — Why forget Bulgarians, never on my mind (Re: The Koan of pi-WAM queues [FORTRAN-S])

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.
 >
 >

[toc] | [prev] | [next] | [standalone]

#15747 — miniTriton CUDA is an alternative to torch variants (Re: Why forget Bulgarians, never on my mind)

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 — 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 — validating the full pipeline on a
realistic workload. These results demonstrate
that Kimi K3 can build a coherent end-to-end
compiler — from DSL frontend and IR passes to
PTX codegen and runtime — rather than isolated

kernels; its from-scratch Tensor Core path
already rivals Triton’s 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.
>  >
>  >
> 

[toc] | [prev] | [next] | [standalone]

#15748 — Andrej Karpathy original gangster of Budget Laptop (Re: miniTriton CUDA is an alternative to torch variants)

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 — 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 — validating the full pipeline on a
> realistic workload. These results demonstrate
> that Kimi K3 can build a coherent end-to-end
> compiler — from DSL frontend and IR passes to
> PTX codegen and runtime — rather than isolated
> 
> kernels; its from-scratch Tensor Core path
> already rivals Triton’s 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.
>>  >
>>  >
>>
> 

[toc] | [prev] | [next] | [standalone]

#15749 — The evolution of hardware and GPT-2 training (Re: Why forget Bulgarians, never on my mind)

Hi,

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

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

But his nanochat archivement has an
interesting time line:

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

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

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

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

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

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

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

Bye

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


Mild Shock schrieb:
> Hi,
> 
> Whats this "forget" trope of glue sniffing
> Rossy Boy with his herpes blisters?
> 
>  > Bulgarians, that's some real Boris and Natasha crap,
>  > forget Hungarians and Bulgarians.
> 
> Why should I forget Bulgarians,
> they are never on my mind. Do you
> see me doing ggml stuff?
> 
> I only hypothesized that it is
> over for Python as the machine
> learning language or AI inferencing
> 
> locally on AI laptops language, and
> made the ggml case, so I already forgot
> about them. Which might give you a glimps,
> 
> why WebGPU was used for this here:
> 
> 11.4 Giga Lips with a Budget Laptop
> https://github.com/Jean-Luc-Picard-2021/gigabudget
> 
> Is an interesting choice. Even
> github has some Languages statistics,
> giving an account what I used:
> 
> HTML 67.5% JavaScript 23.1% CSS 9.4%
> 
> Have Fun!
> 
> Bye
> 
> P.S.: The example below is not p-adics,
> you complete imbecil moron. Its just:
> 
> 7-11 cubic Solution by Pritchard & Gries
> https://www.cs.cornell.edu/gries/TechReports/83-574.pdf
> 
> Ross Finlayson schrieb:
>  > On 07/26/2026 10:52 AM, Mild Shock wrote:
>  >> Hi,
>  >>
>  >> You see it all boils down to find your inner peace
>  >> by an immaculate inception of some queue datatype.
>  >>
>  >> KOAN/Fortran-S was an early 1990s research programming
>  >> system for distributed-memory multiprocessors . Developed
>  >> at ENS Lyon in the early 1990s . Often listed alongside
>  >> other historical parallel programming efforts.
>  >>
>  >> The Message Passing: The research explicitly
>  >> compared the SVM approach against message passing
>  >> on the same hardware . The finding was that SVM
>  >> could achieve good performance without the low-level
>  >>
>  >> complexity of managing explicit messages, though
>  >> the best results often came from a hybrid approach (sic!)
>  >> Here is an interesting baseline, from Java,
>  >> a class ElevenSingle that only does:
>  >>
>  >>      public static void run() {
>  >>          for (int A = 1; A < 192; A++) {
>  >>              int Y = (771-A)/3;
>  >>              for (int B = A; B < Y; B++) {
>  >>                  int Z = (771-A-B)/2;
>  >>                  for (int C = B; C < Z; C++) {
>  >>                      int D = 711-A-B-C;
>  >>                      if (A*B*C == 711000000/D &&
>  >>                            711000000 % D == 0)
>  >>      System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
>  >>                  }
>  >>              }
>  >>          }
>  >>      }
>  >>
>  >> And then compare it to ElevenMulti, doing some
>  >> Work Balancing Scheduler Tetris Game with 8 cores:
>  >>
>  >> ElevenSingle
>  >> A=120, B=125, C=150, D=316
>  >> 6.628 ms
>  >>
>  >> ElevenMulti
>  >> A=120, B=125, C=150, D=316
>  >> 1.941 ms
>  >>
>  >> Not great, not terrible!
>  >>
>  >> Bye
>  >
>  > Oh, that's just "tricks of p-adic arithmetic".
>  >
>  > Like other sock-puppet howler trolls, when confronted
>  > with its base incredulity, it will descend to its
>  > lower levers of the pathos variety.
>  >
>  > You might be happier learning about Julia trees and
>  > raster ops, instead of shilling yet another Ramanujan
>  > series without saying how it's made.
>  >
>  > Bulgarians, that's some real Boris and Natasha crap,
>  > forget Hungarians and Bulgarians.
>  >
>  >
> 

[toc] | [prev] | [next] | [standalone]

#15750 — How speed up π-WAM with vector operations (Re: The evolution of hardware and GPT-2 training)

Hi,

One could critisize that my π-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 π-WAM, take this compilation of the Prolog
goal between(0,1023,X), Y is X*2+3:

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

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

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

Have Fun!

Bye

Mild Shock schrieb:
> Hi,
> 
> While Huggingfaces hired GG in 2026,
> AK was hired by Anthropic in 2026:
> 
> Andrej Karpathy (born 23 October 1986[3])
> is a Slovak-Canadian AI researcher, who
> co-founded and formerly worked at OpenAI
> In 2026 he joined Anthropic as part of
> the pretraining team.
> https://en.wikipedia.org/wiki/Andrej_Karpathy
> 
> But his nanochat archivement has an
> interesting time line:
> 
> 168 hours , Original OpenAI GPT-2 checkpoint, 2019
> 3 hours , d24 baseline, slightly overtrained, Jan 29 2026
> 1 1/2 hour, autoresearch round 2, Mar 14 2026
> The best ChatGPT that $100 can buy.
> https://github.com/karpathy/nanochat
> 
> But what hardware was the enabler. What is the
> NVIDIA H100 GPU even. Well the thingy is surely not
> a Budget Laptop, performance pretty much
> 
> dependence on data elememt size, the H100 NVL
> version (*), and when using tensor operations,
> and not only scalar operations:
> 
> 8-bit towards 3000 tera flops
> 16-bit towards 1500 tera flops
> 32-bit towards 900 tera flops
> 
> Cool! I guess this experiment would tap into 60
> tera flops, since it only uses scalar operations so far:
> 
> 11.4 Giga Lips with a Budget Laptop
> https://github.com/Jean-Luc-Picard-2021/gigabudget
> 
> You could perform it by migration the web application
> using WebGPU into a node.js standalone application
> using the dawn library for GPU access.
> 
> Bye
> 
> (*) 
> https://resources.nvidia.com/en-us-hopper-architecture/nvidia-tensor-core-gpu-datasheet 
> 
> 
> 
> Mild Shock schrieb:
>> Hi,
>>
>> Whats this "forget" trope of glue sniffing
>> Rossy Boy with his herpes blisters?
>>
>>  > Bulgarians, that's some real Boris and Natasha crap,
>>  > forget Hungarians and Bulgarians.
>>
>> Why should I forget Bulgarians,
>> they are never on my mind. Do you
>> see me doing ggml stuff?
>>
>> I only hypothesized that it is
>> over for Python as the machine
>> learning language or AI inferencing
>>
>> locally on AI laptops language, and
>> made the ggml case, so I already forgot
>> about them. Which might give you a glimps,
>>
>> why WebGPU was used for this here:
>>
>> 11.4 Giga Lips with a Budget Laptop
>> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>
>> Is an interesting choice. Even
>> github has some Languages statistics,
>> giving an account what I used:
>>
>> HTML 67.5% JavaScript 23.1% CSS 9.4%
>>
>> Have Fun!
>>
>> Bye
>>
>> P.S.: The example below is not p-adics,
>> you complete imbecil moron. Its just:
>>
>> 7-11 cubic Solution by Pritchard & Gries
>> https://www.cs.cornell.edu/gries/TechReports/83-574.pdf
>>
>> Ross Finlayson schrieb:
>>  > On 07/26/2026 10:52 AM, Mild Shock wrote:
>>  >> Hi,
>>  >>
>>  >> You see it all boils down to find your inner peace
>>  >> by an immaculate inception of some queue datatype.
>>  >>
>>  >> KOAN/Fortran-S was an early 1990s research programming
>>  >> system for distributed-memory multiprocessors . Developed
>>  >> at ENS Lyon in the early 1990s . Often listed alongside
>>  >> other historical parallel programming efforts.
>>  >>
>>  >> The Message Passing: The research explicitly
>>  >> compared the SVM approach against message passing
>>  >> on the same hardware . The finding was that SVM
>>  >> could achieve good performance without the low-level
>>  >>
>>  >> complexity of managing explicit messages, though
>>  >> the best results often came from a hybrid approach (sic!)
>>  >> Here is an interesting baseline, from Java,
>>  >> a class ElevenSingle that only does:
>>  >>
>>  >>      public static void run() {
>>  >>          for (int A = 1; A < 192; A++) {
>>  >>              int Y = (771-A)/3;
>>  >>              for (int B = A; B < Y; B++) {
>>  >>                  int Z = (771-A-B)/2;
>>  >>                  for (int C = B; C < Z; C++) {
>>  >>                      int D = 711-A-B-C;
>>  >>                      if (A*B*C == 711000000/D &&
>>  >>                            711000000 % D == 0)
>>  >>      System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
>>  >>                  }
>>  >>              }
>>  >>          }
>>  >>      }
>>  >>
>>  >> And then compare it to ElevenMulti, doing some
>>  >> Work Balancing Scheduler Tetris Game with 8 cores:
>>  >>
>>  >> ElevenSingle
>>  >> A=120, B=125, C=150, D=316
>>  >> 6.628 ms
>>  >>
>>  >> ElevenMulti
>>  >> A=120, B=125, C=150, D=316
>>  >> 1.941 ms
>>  >>
>>  >> Not great, not terrible!
>>  >>
>>  >> Bye
>>  >
>>  > Oh, that's just "tricks of p-adic arithmetic".
>>  >
>>  > Like other sock-puppet howler trolls, when confronted
>>  > with its base incredulity, it will descend to its
>>  > lower levers of the pathos variety.
>>  >
>>  > You might be happier learning about Julia trees and
>>  > raster ops, instead of shilling yet another Ramanujan
>>  > series without saying how it's made.
>>  >
>>  > Bulgarians, that's some real Boris and Natasha crap,
>>  > forget Hungarians and Bulgarians.
>>  >
>>  >
>>
> 

[toc] | [prev] | [next] | [standalone]

#15754 — AI accelerator extend from GPU to CPU [Zero Copying] (Re: How speed up π-WAM with vector operations)

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 π-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 π-WAM, take this compilation of the Prolog
> goal between(0,1023,X), Y is X*2+3:
> 
> int X;
> int Y;
> for (X=0; X < 1024; X++) {
>      Y=X*2+3;
>      [...]
> }
> 
> With vector operations, and vectors of size
> 32 one could do:
> 
> int X1;
> int[] X = new int[32];
> int X3;
> int[] Y = new int[32];
> for (X1 = 0; X1 < 1024 / 32; X1++) {
>      for (int X2 = 0; X2 < 32; X2++)
>         X[X2] = X1*32+X2;
>      vec_mul_add(X, 2, 3, Y);
>      [..]
> }
> 
> Have Fun!
> 
> Bye
> 
> Mild Shock schrieb:
>> Hi,
>>
>> While Huggingfaces hired GG in 2026,
>> AK was hired by Anthropic in 2026:
>>
>> Andrej Karpathy (born 23 October 1986[3])
>> is a Slovak-Canadian AI researcher, who
>> co-founded and formerly worked at OpenAI
>> In 2026 he joined Anthropic as part of
>> the pretraining team.
>> https://en.wikipedia.org/wiki/Andrej_Karpathy
>>
>> But his nanochat archivement has an
>> interesting time line:
>>
>> 168 hours , Original OpenAI GPT-2 checkpoint, 2019
>> 3 hours , d24 baseline, slightly overtrained, Jan 29 2026
>> 1 1/2 hour, autoresearch round 2, Mar 14 2026
>> The best ChatGPT that $100 can buy.
>> https://github.com/karpathy/nanochat
>>
>> But what hardware was the enabler. What is the
>> NVIDIA H100 GPU even. Well the thingy is surely not
>> a Budget Laptop, performance pretty much
>>
>> dependence on data elememt size, the H100 NVL
>> version (*), and when using tensor operations,
>> and not only scalar operations:
>>
>> 8-bit towards 3000 tera flops
>> 16-bit towards 1500 tera flops
>> 32-bit towards 900 tera flops
>>
>> Cool! I guess this experiment would tap into 60
>> tera flops, since it only uses scalar operations so far:
>>
>> 11.4 Giga Lips with a Budget Laptop
>> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>
>> You could perform it by migration the web application
>> using WebGPU into a node.js standalone application
>> using the dawn library for GPU access.
>>
>> Bye
>>
>> (*) 
>> https://resources.nvidia.com/en-us-hopper-architecture/nvidia-tensor-core-gpu-datasheet 
>>
>>
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>> Whats this "forget" trope of glue sniffing
>>> Rossy Boy with his herpes blisters?
>>>
>>>  > Bulgarians, that's some real Boris and Natasha crap,
>>>  > forget Hungarians and Bulgarians.
>>>
>>> Why should I forget Bulgarians,
>>> they are never on my mind. Do you
>>> see me doing ggml stuff?
>>>
>>> I only hypothesized that it is
>>> over for Python as the machine
>>> learning language or AI inferencing
>>>
>>> locally on AI laptops language, and
>>> made the ggml case, so I already forgot
>>> about them. Which might give you a glimps,
>>>
>>> why WebGPU was used for this here:
>>>
>>> 11.4 Giga Lips with a Budget Laptop
>>> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>>
>>> Is an interesting choice. Even
>>> github has some Languages statistics,
>>> giving an account what I used:
>>>
>>> HTML 67.5% JavaScript 23.1% CSS 9.4%
>>>
>>> Have Fun!
>>>
>>> Bye
>>>
>>> P.S.: The example below is not p-adics,
>>> you complete imbecil moron. Its just:
>>>
>>> 7-11 cubic Solution by Pritchard & Gries
>>> https://www.cs.cornell.edu/gries/TechReports/83-574.pdf
>>>
>>> Ross Finlayson schrieb:
>>>  > On 07/26/2026 10:52 AM, Mild Shock wrote:
>>>  >> Hi,
>>>  >>
>>>  >> You see it all boils down to find your inner peace
>>>  >> by an immaculate inception of some queue datatype.
>>>  >>
>>>  >> KOAN/Fortran-S was an early 1990s research programming
>>>  >> system for distributed-memory multiprocessors . Developed
>>>  >> at ENS Lyon in the early 1990s . Often listed alongside
>>>  >> other historical parallel programming efforts.
>>>  >>
>>>  >> The Message Passing: The research explicitly
>>>  >> compared the SVM approach against message passing
>>>  >> on the same hardware . The finding was that SVM
>>>  >> could achieve good performance without the low-level
>>>  >>
>>>  >> complexity of managing explicit messages, though
>>>  >> the best results often came from a hybrid approach (sic!)
>>>  >> Here is an interesting baseline, from Java,
>>>  >> a class ElevenSingle that only does:
>>>  >>
>>>  >>      public static void run() {
>>>  >>          for (int A = 1; A < 192; A++) {
>>>  >>              int Y = (771-A)/3;
>>>  >>              for (int B = A; B < Y; B++) {
>>>  >>                  int Z = (771-A-B)/2;
>>>  >>                  for (int C = B; C < Z; C++) {
>>>  >>                      int D = 711-A-B-C;
>>>  >>                      if (A*B*C == 711000000/D &&
>>>  >>                            711000000 % D == 0)
>>>  >>      System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
>>>  >>                  }
>>>  >>              }
>>>  >>          }
>>>  >>      }
>>>  >>
>>>  >> And then compare it to ElevenMulti, doing some
>>>  >> Work Balancing Scheduler Tetris Game with 8 cores:
>>>  >>
>>>  >> ElevenSingle
>>>  >> A=120, B=125, C=150, D=316
>>>  >> 6.628 ms
>>>  >>
>>>  >> ElevenMulti
>>>  >> A=120, B=125, C=150, D=316
>>>  >> 1.941 ms
>>>  >>
>>>  >> Not great, not terrible!
>>>  >>
>>>  >> Bye
>>>  >
>>>  > Oh, that's just "tricks of p-adic arithmetic".
>>>  >
>>>  > Like other sock-puppet howler trolls, when confronted
>>>  > with its base incredulity, it will descend to its
>>>  > lower levers of the pathos variety.
>>>  >
>>>  > You might be happier learning about Julia trees and
>>>  > raster ops, instead of shilling yet another Ramanujan
>>>  > series without saying how it's made.
>>>  >
>>>  > Bulgarians, that's some real Boris and Natasha crap,
>>>  > forget Hungarians and Bulgarians.
>>>  >
>>>  >
>>>
>>
> 

[toc] | [prev] | [next] | [standalone]

#15755 — The invention of vector and matrix registers [NVIDIA Volta] (Re: AI accelerator extend from GPU to CPU [Zero Copying])

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 π-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 π-WAM, take this compilation of the Prolog
>> goal between(0,1023,X), Y is X*2+3:
>>
>> int X;
>> int Y;
>> for (X=0; X < 1024; X++) {
>>      Y=X*2+3;
>>      [...]
>> }
>>
>> With vector operations, and vectors of size
>> 32 one could do:
>>
>> int X1;
>> int[] X = new int[32];
>> int X3;
>> int[] Y = new int[32];
>> for (X1 = 0; X1 < 1024 / 32; X1++) {
>>      for (int X2 = 0; X2 < 32; X2++)
>>         X[X2] = X1*32+X2;
>>      vec_mul_add(X, 2, 3, Y);
>>      [..]
>> }
>>
>> Have Fun!
>>
>> Bye
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>> While Huggingfaces hired GG in 2026,
>>> AK was hired by Anthropic in 2026:
>>>
>>> Andrej Karpathy (born 23 October 1986[3])
>>> is a Slovak-Canadian AI researcher, who
>>> co-founded and formerly worked at OpenAI
>>> In 2026 he joined Anthropic as part of
>>> the pretraining team.
>>> https://en.wikipedia.org/wiki/Andrej_Karpathy
>>>
>>> But his nanochat archivement has an
>>> interesting time line:
>>>
>>> 168 hours , Original OpenAI GPT-2 checkpoint, 2019
>>> 3 hours , d24 baseline, slightly overtrained, Jan 29 2026
>>> 1 1/2 hour, autoresearch round 2, Mar 14 2026
>>> The best ChatGPT that $100 can buy.
>>> https://github.com/karpathy/nanochat
>>>
>>> But what hardware was the enabler. What is the
>>> NVIDIA H100 GPU even. Well the thingy is surely not
>>> a Budget Laptop, performance pretty much
>>>
>>> dependence on data elememt size, the H100 NVL
>>> version (*), and when using tensor operations,
>>> and not only scalar operations:
>>>
>>> 8-bit towards 3000 tera flops
>>> 16-bit towards 1500 tera flops
>>> 32-bit towards 900 tera flops
>>>
>>> Cool! I guess this experiment would tap into 60
>>> tera flops, since it only uses scalar operations so far:
>>>
>>> 11.4 Giga Lips with a Budget Laptop
>>> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>>
>>> You could perform it by migration the web application
>>> using WebGPU into a node.js standalone application
>>> using the dawn library for GPU access.
>>>
>>> Bye
>>>
>>> (*) 
>>> https://resources.nvidia.com/en-us-hopper-architecture/nvidia-tensor-core-gpu-datasheet 
>>>
>>>
>>>
>>> Mild Shock schrieb:
>>>> Hi,
>>>>
>>>> Whats this "forget" trope of glue sniffing
>>>> Rossy Boy with his herpes blisters?
>>>>
>>>>  > Bulgarians, that's some real Boris and Natasha crap,
>>>>  > forget Hungarians and Bulgarians.
>>>>
>>>> Why should I forget Bulgarians,
>>>> they are never on my mind. Do you
>>>> see me doing ggml stuff?
>>>>
>>>> I only hypothesized that it is
>>>> over for Python as the machine
>>>> learning language or AI inferencing
>>>>
>>>> locally on AI laptops language, and
>>>> made the ggml case, so I already forgot
>>>> about them. Which might give you a glimps,
>>>>
>>>> why WebGPU was used for this here:
>>>>
>>>> 11.4 Giga Lips with a Budget Laptop
>>>> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>>>
>>>> Is an interesting choice. Even
>>>> github has some Languages statistics,
>>>> giving an account what I used:
>>>>
>>>> HTML 67.5% JavaScript 23.1% CSS 9.4%
>>>>
>>>> Have Fun!
>>>>
>>>> Bye
>>>>
>>>> P.S.: The example below is not p-adics,
>>>> you complete imbecil moron. Its just:
>>>>
>>>> 7-11 cubic Solution by Pritchard & Gries
>>>> https://www.cs.cornell.edu/gries/TechReports/83-574.pdf
>>>>
>>>> Ross Finlayson schrieb:
>>>>  > On 07/26/2026 10:52 AM, Mild Shock wrote:
>>>>  >> Hi,
>>>>  >>
>>>>  >> You see it all boils down to find your inner peace
>>>>  >> by an immaculate inception of some queue datatype.
>>>>  >>
>>>>  >> KOAN/Fortran-S was an early 1990s research programming
>>>>  >> system for distributed-memory multiprocessors . Developed
>>>>  >> at ENS Lyon in the early 1990s . Often listed alongside
>>>>  >> other historical parallel programming efforts.
>>>>  >>
>>>>  >> The Message Passing: The research explicitly
>>>>  >> compared the SVM approach against message passing
>>>>  >> on the same hardware . The finding was that SVM
>>>>  >> could achieve good performance without the low-level
>>>>  >>
>>>>  >> complexity of managing explicit messages, though
>>>>  >> the best results often came from a hybrid approach (sic!)
>>>>  >> Here is an interesting baseline, from Java,
>>>>  >> a class ElevenSingle that only does:
>>>>  >>
>>>>  >>      public static void run() {
>>>>  >>          for (int A = 1; A < 192; A++) {
>>>>  >>              int Y = (771-A)/3;
>>>>  >>              for (int B = A; B < Y; B++) {
>>>>  >>                  int Z = (771-A-B)/2;
>>>>  >>                  for (int C = B; C < Z; C++) {
>>>>  >>                      int D = 711-A-B-C;
>>>>  >>                      if (A*B*C == 711000000/D &&
>>>>  >>                            711000000 % D == 0)
>>>>  >>      System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
>>>>  >>                  }
>>>>  >>              }
>>>>  >>          }
>>>>  >>      }
>>>>  >>
>>>>  >> And then compare it to ElevenMulti, doing some
>>>>  >> Work Balancing Scheduler Tetris Game with 8 cores:
>>>>  >>
>>>>  >> ElevenSingle
>>>>  >> A=120, B=125, C=150, D=316
>>>>  >> 6.628 ms
>>>>  >>
>>>>  >> ElevenMulti
>>>>  >> A=120, B=125, C=150, D=316
>>>>  >> 1.941 ms
>>>>  >>
>>>>  >> Not great, not terrible!
>>>>  >>
>>>>  >> Bye
>>>>  >
>>>>  > Oh, that's just "tricks of p-adic arithmetic".
>>>>  >
>>>>  > Like other sock-puppet howler trolls, when confronted
>>>>  > with its base incredulity, it will descend to its
>>>>  > lower levers of the pathos variety.
>>>>  >
>>>>  > You might be happier learning about Julia trees and
>>>>  > raster ops, instead of shilling yet another Ramanujan
>>>>  > series without saying how it's made.
>>>>  >
>>>>  > Bulgarians, that's some real Boris and Natasha crap,
>>>>  > forget Hungarians and Bulgarians.
>>>>  >
>>>>  >
>>>>
>>>
>>
> 

[toc] | [prev] | [next] | [standalone]

#15757 — Re: The invention of vector and matrix registers [NVIDIA Volta] (Re: AI accelerator extend from GPU to CPU [Zero Copying])

On 07/27/2026 04:22 AM, Mild Shock wrote:
> Hi,
>
> But the example gives also way to vector
> and matrix registers. The int[] X and
> int[] Y could be also held in vector
>
> registers. Compilers can also optimize
> away int[] Y, and use a inline modification,
> in case X isn't used later, then playing
>
> the role of Y:
>
> vec_mul_add(X, 2, 3, X)
>
> Vector and matrix registers in modern GPUs
> emerged from distinct architectural milestones:
> vector-like register files developed with
> early programmable 3D vertex/pixel pipelines
>
> in the late 1990s to early 2000s. While
> dedicated multi-dimensional matrix registers
> (Tensor Cores/Matrix Cores) were invented by
> NVIDIA in 2017, starting with the Tesla
>
> V100 (Volta microarchitecture):
>
>  From Volta To Blackwell
> https://newsletter.semianalysis.com/p/nvidia-tensor-core-evolution-from-volta-to-blackwell
>
>
> You see the scheduling of tensure core occupation
> scheduling in the above article, including memory
> and register flow, following the section:
>
> MMA Instruction Overview
>
> It went through a couple of generations, leading
> to Tensor Memory (TMEM) and collective operations,
> basically realizing the PIM idea:
>
> Processing-in-Memory Tutorials
> https://www.sigarch.org/processing-in-memory-tutorials-experiences-from-past-two-years-and-thoughts-looking-forward/
>
>
> Have Fun!
>
> Bye
>
>
> Mild Shock schrieb:
>> Hi,
>>
>> The nice thing about AI accelerators, pioneered
>> maybe by Apple Silicon and their unified memory.
>> The AMD APU model can be extended so that
>>
>> vector and matrix operations become uniformly
>> available for GPU and CPU. With unified memory
>> already a vector operation such as:
>>
>> vec_mul_add(X, 2, 3, Y)
>>
>> Only needs the X and Y address. But I havent
>> got my head around yet how this is all organized.
>> Maybe a GPU has still its own GEMM cores,
>>
>> but you find Apple Silicon C++/C source code,
>> that taps into vector and matrix operations
>> by Zero Copying. The Copying is left to the DMA
>>
>> of the vector or matrix operation. And moderated
>> by the various caches. Leading to the slogan, that
>> multiple floating point operations become zero cost:
>>
>> Some teaching can be found here
>> https://www.hpc-ch.org/category/topics/course-workshop/
>>
>> Bye
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>> One could critisize that my π-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 π-WAM, take this compilation of the Prolog
>>> goal between(0,1023,X), Y is X*2+3:
>>>
>>> int X;
>>> int Y;
>>> for (X=0; X < 1024; X++) {
>>>      Y=X*2+3;
>>>      [...]
>>> }
>>>
>>> With vector operations, and vectors of size
>>> 32 one could do:
>>>
>>> int X1;
>>> int[] X = new int[32];
>>> int X3;
>>> int[] Y = new int[32];
>>> for (X1 = 0; X1 < 1024 / 32; X1++) {
>>>      for (int X2 = 0; X2 < 32; X2++)
>>>         X[X2] = X1*32+X2;
>>>      vec_mul_add(X, 2, 3, Y);
>>>      [..]
>>> }
>>>
>>> Have Fun!
>>>
>>> Bye
>>>
>>> Mild Shock schrieb:
>>>> Hi,
>>>>
>>>> While Huggingfaces hired GG in 2026,
>>>> AK was hired by Anthropic in 2026:
>>>>
>>>> Andrej Karpathy (born 23 October 1986[3])
>>>> is a Slovak-Canadian AI researcher, who
>>>> co-founded and formerly worked at OpenAI
>>>> In 2026 he joined Anthropic as part of
>>>> the pretraining team.
>>>> https://en.wikipedia.org/wiki/Andrej_Karpathy
>>>>
>>>> But his nanochat archivement has an
>>>> interesting time line:
>>>>
>>>> 168 hours , Original OpenAI GPT-2 checkpoint, 2019
>>>> 3 hours , d24 baseline, slightly overtrained, Jan 29 2026
>>>> 1 1/2 hour, autoresearch round 2, Mar 14 2026
>>>> The best ChatGPT that $100 can buy.
>>>> https://github.com/karpathy/nanochat
>>>>
>>>> But what hardware was the enabler. What is the
>>>> NVIDIA H100 GPU even. Well the thingy is surely not
>>>> a Budget Laptop, performance pretty much
>>>>
>>>> dependence on data elememt size, the H100 NVL
>>>> version (*), and when using tensor operations,
>>>> and not only scalar operations:
>>>>
>>>> 8-bit towards 3000 tera flops
>>>> 16-bit towards 1500 tera flops
>>>> 32-bit towards 900 tera flops
>>>>
>>>> Cool! I guess this experiment would tap into 60
>>>> tera flops, since it only uses scalar operations so far:
>>>>
>>>> 11.4 Giga Lips with a Budget Laptop
>>>> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>>>
>>>> You could perform it by migration the web application
>>>> using WebGPU into a node.js standalone application
>>>> using the dawn library for GPU access.
>>>>
>>>> Bye
>>>>
>>>> (*)
>>>> https://resources.nvidia.com/en-us-hopper-architecture/nvidia-tensor-core-gpu-datasheet
>>>>
>>>>
>>>>
>>>> Mild Shock schrieb:
>>>>> Hi,
>>>>>
>>>>> Whats this "forget" trope of glue sniffing
>>>>> Rossy Boy with his herpes blisters?
>>>>>
>>>>>  > Bulgarians, that's some real Boris and Natasha crap,
>>>>>  > forget Hungarians and Bulgarians.
>>>>>
>>>>> Why should I forget Bulgarians,
>>>>> they are never on my mind. Do you
>>>>> see me doing ggml stuff?
>>>>>
>>>>> I only hypothesized that it is
>>>>> over for Python as the machine
>>>>> learning language or AI inferencing
>>>>>
>>>>> locally on AI laptops language, and
>>>>> made the ggml case, so I already forgot
>>>>> about them. Which might give you a glimps,
>>>>>
>>>>> why WebGPU was used for this here:
>>>>>
>>>>> 11.4 Giga Lips with a Budget Laptop
>>>>> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>>>>
>>>>> Is an interesting choice. Even
>>>>> github has some Languages statistics,
>>>>> giving an account what I used:
>>>>>
>>>>> HTML 67.5% JavaScript 23.1% CSS 9.4%
>>>>>
>>>>> Have Fun!
>>>>>
>>>>> Bye
>>>>>
>>>>> P.S.: The example below is not p-adics,
>>>>> you complete imbecil moron. Its just:
>>>>>
>>>>> 7-11 cubic Solution by Pritchard & Gries
>>>>> https://www.cs.cornell.edu/gries/TechReports/83-574.pdf
>>>>>
>>>>> Ross Finlayson schrieb:
>>>>>  > On 07/26/2026 10:52 AM, Mild Shock wrote:
>>>>>  >> Hi,
>>>>>  >>
>>>>>  >> You see it all boils down to find your inner peace
>>>>>  >> by an immaculate inception of some queue datatype.
>>>>>  >>
>>>>>  >> KOAN/Fortran-S was an early 1990s research programming
>>>>>  >> system for distributed-memory multiprocessors . Developed
>>>>>  >> at ENS Lyon in the early 1990s . Often listed alongside
>>>>>  >> other historical parallel programming efforts.
>>>>>  >>
>>>>>  >> The Message Passing: The research explicitly
>>>>>  >> compared the SVM approach against message passing
>>>>>  >> on the same hardware . The finding was that SVM
>>>>>  >> could achieve good performance without the low-level
>>>>>  >>
>>>>>  >> complexity of managing explicit messages, though
>>>>>  >> the best results often came from a hybrid approach (sic!)
>>>>>  >> Here is an interesting baseline, from Java,
>>>>>  >> a class ElevenSingle that only does:
>>>>>  >>
>>>>>  >>      public static void run() {
>>>>>  >>          for (int A = 1; A < 192; A++) {
>>>>>  >>              int Y = (771-A)/3;
>>>>>  >>              for (int B = A; B < Y; B++) {
>>>>>  >>                  int Z = (771-A-B)/2;
>>>>>  >>                  for (int C = B; C < Z; C++) {
>>>>>  >>                      int D = 711-A-B-C;
>>>>>  >>                      if (A*B*C == 711000000/D &&
>>>>>  >>                            711000000 % D == 0)
>>>>>  >>      System.out.println("A="+A+", B="+B+", C="+C+", D="+D);
>>>>>  >>                  }
>>>>>  >>              }
>>>>>  >>          }
>>>>>  >>      }
>>>>>  >>
>>>>>  >> And then compare it to ElevenMulti, doing some
>>>>>  >> Work Balancing Scheduler Tetris Game with 8 cores:
>>>>>  >>
>>>>>  >> ElevenSingle
>>>>>  >> A=120, B=125, C=150, D=316
>>>>>  >> 6.628 ms
>>>>>  >>
>>>>>  >> ElevenMulti
>>>>>  >> A=120, B=125, C=150, D=316
>>>>>  >> 1.941 ms
>>>>>  >>
>>>>>  >> Not great, not terrible!
>>>>>  >>
>>>>>  >> Bye
>>>>>  >
>>>>>  > Oh, that's just "tricks of p-adic arithmetic".
>>>>>  >
>>>>>  > Like other sock-puppet howler trolls, when confronted
>>>>>  > with its base incredulity, it will descend to its
>>>>>  > lower levers of the pathos variety.
>>>>>  >
>>>>>  > You might be happier learning about Julia trees and
>>>>>  > raster ops, instead of shilling yet another Ramanujan
>>>>>  > series without saying how it's made.
>>>>>  >
>>>>>  > Bulgarians, that's some real Boris and Natasha crap,
>>>>>  > forget Hungarians and Bulgarians.
>>>>>  >
>>>>>  >
>>>>>
>>>>
>>>
>>
>

That's bullshit, and alike those talking heads that
sniff their way into talking about many-core jumbo-trons,
the super-scalar is as old as the scalar and Cray and examples alike
the Connection Machine what made all the craze of neural nets
is old-wrapped-as-new.

Fabless chips did it already.


Data centers should pay a 10000% excise on electricity,
wherever it comes from, a natural regulator of inverted economies.

And by ten thousand percent I really mean a ten thousand percent.

[toc] | [prev] | [next] | [standalone]

#15760 — Maybe they should have named it NVIDIA Einstein [Rossy Boy Toe Sucking] (Was: The invention of vector and matrix registers [NVIDIA Volta])

Hi,

I guess Rossy Boys mother was so disappointed
in the 50's that is son didn't become the
next Einstain, physics was the ultinate idol,

so that Rossy Boy was left rotting in the
basement. But Rossy Boys indoctrination was
not spurious, he now is conditioned on

Einstein. Maybe NVIDIA should have named
its Tesla V100 card NVIDIA Einstein. You would
then see Rossy Boy toe sucking the graphic

card, in his pyjamas in the basement.

Bye

Ross Finlayson schrieb:
> On 07/27/2026 04:22 AM, Mild Shock wrote:
>> Hi,
>>
>> But the example gives also way to vector
>> and matrix registers. The int[] X and
>> int[] Y could be also held in vector
>>
>> registers. Compilers can also optimize
>> away int[] Y, and use a inline modification,
>> in case X isn't used later, then playing
>>
>> the role of Y:
>>
>> vec_mul_add(X, 2, 3, X)
>>
>> Vector and matrix registers in modern GPUs
>> emerged from distinct architectural milestones:
>> vector-like register files developed with
>> early programmable 3D vertex/pixel pipelines
>>
>> in the late 1990s to early 2000s. While
>> dedicated multi-dimensional matrix registers
>> (Tensor Cores/Matrix Cores) were invented by
>> NVIDIA in 2017, starting with the Tesla
>>
>> V100 (Volta microarchitecture):
>>
>>  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.
> 
> 

[toc] | [prev] | [next] | [standalone]

#15763 — Re: The invention of vector and matrix registers [NVIDIA Volta] (Re: AI accelerator extend from GPU to CPU [Zero Copying])

In comp.lang.prolog Ross Finlayson <ross.a.finlayson@gmail.com> wrote:
...
> Data centers should pay a 10000% excise on electricity,
> wherever it comes from, a natural regulator of inverted economies.
> And by ten thousand percent I really mean a ten thousand percent.

And what would a huge surcharge do?
Almost always end up affecting the less powerful end of society
with increased costs to services the AI industry will be doing
more and more of over time.

I started a little data center (exaflops.com) many years ago.
In those distant days people (in fact one was a prof of computer
science) told me you could never make money running a supercomputer. 
LOL. :)

I've had many years to watch the trends and a far more efficient
way to solve resource problems in this area is to change the
algorithms. There is vast room for improvement, mostly because
of prevailing attitudes.

I used to do competetion data science as a sideline. Companies
would pay almost any price to get an extra decimal place in
the accuracy of their forecasting processes. But typically
they were trying to supercharge a system that should be scrapped
and re-designed from scratch. One area I'm thinking of is
investment. I had a customer one time -- like many times --
ask to improve a system that predicted the future price of
various stocks. The idea (for them) was to have as accurate a
prediction of what some stock would be worth in a week or a month's
time so that some moron could use the information to decide when
to buy or sell the thing.

I tried to argue the efficient thing was to create a system that
takes the human out of the loop altogether. It doesnt provide info
for someone to decide whether or not to follow the advice --
that is just introducing more noise into the loop and probably
cancels any benefit of adding a couple decimal places of precision.
What you *should* do is make a system that is tuned to robustly
maximize the profit from managing a portfolio.

Of course they wouldnt come at that. You can't suggest taking the
managers out of the loop. :)

Another idea relevant to current AI methods might be to curtail
use of typical neural net algorithms. Many of them try to squeeze
the best performance of some NN during the training  phase in
the hope the resulting system will generalize well enough to be useful
on new data. But there's kind-of a law that the harder you train
some system to perform a task well, the less well they can subsuently
perform a more general version of the same thing. It's amusing when
you look at the graphs of NN being trained and then tested that
given a more general problem to solve after being trained to solve
similar problems very very well the poor old NN does worse that it
would have done if it had 0 training in the first place.

It's not like we dont know how to improve this kind of performance.
Try less hard in the training phase or make it "more noisy".
Turns out genetic methods are just the ticket for this.
The training produces less over-fitting and the resulting system
generalizes better than it did before training and more importantly
it takes maybe an order of magnitude crunching to produce a good answer
than the usual over-fit answer.

Anyway. Have to go and feed the cat.

-- 
Nothing in life is to be feared, it is only to be understood. 
Now is the time to understand more, so that we may fear less.
-- M. Curie

[toc] | [prev] | [next] | [standalone]

#15764 — Re: The invention of vector and matrix registers [NVIDIA Volta] (Re: AI accelerator extend from GPU to CPU [Zero Copying])

In comp.lang.prolog R Kym Horsell <kym@sdf.com> wrote:
> generalizes better than it did before training and more importantly
> it takes maybe an order of magnitude crunching to produce a good answer
                                      /\ less
> than the usual over-fit answer.

[toc] | [prev] | [next] | [standalone]

#15766 — π-WAM is not adding decimals, it is removing decimals (Was: The invention of vector and matrix registers [NVIDIA Volta])

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 “Trillions of Operations Per Second.”
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 π-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 π-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.
> 

[toc] | [prev] | [next] | [standalone]

#15758 — Potato Computer owner impressed by Ukraine Tech [Rossy Boys Brother?] (Was: Hurry the blue bus doesnt stop indefinitely)

Hi,

Slowly I start understanding numbnuts like
Rossy Boy who don't understand tech, although
they are from UK and not from a 3rd world

country, and also I start understanding morons
like Micro Penis, who are behind a curtain,
and cannot access a lot of tech.

The same holds for SWI Prologs newest campaign
that probably adresses some poor indians that
have neither 5G nor Macs:

1:38:01 The Kyiv keynote disaster
https://www.youtube.com/watch?v=U8goS6B3BbI

Woa! Real time download of Scala, Closure,
etc.. Whats the magic behind that? Some SWI
point of sale, downloading it via its

keyboard and some telephathy module ?

Bye

Mild Shock schrieb:
> Hi,
> 
> Ride the snake
> He's old and his skin is cold
> The west is the best
> The west is the best
> Get here and we'll do the rest
> The blue bus is calling us
> The blue bus is calling us
> Driver, where you taking us?
> 
> Apocalypse Now intro: The Doors, The End {1979}
> https://www.youtube.com/watch?v=CIrvSJwwJUE
> 
> Bye
> 
>  > Hi,
>  >
>  > Again I posted everything here:
>  >
>  >> 11.4 Giga Lips with a Budget Laptop
>  >> https://github.com/Jean-Luc-Picard-2021/gigabudget
>  >
>  > The repo says, same time when I posted
>  > the link first time:
>  >
>  >> This repository was archived by the
>  >> owner on Jul 9, 2026. It is now read-only.
>  >
>  > Now a USENET user, who had already entitled
>  > himself for a couple of irrational accusations
>  >
>  > towards my side, is asking this question:
>  >
>  > Chris M. Thomasson schrieb, Jul 24, 2026
>  >> Show an outline of what you
>  >> need you compute shader to do?
>  >
>  > Bravo, thats a delay of a wooping 15 days.
>  >
>  > Bye
> 
> Mild Shock schrieb:
>> Hi,
>>
>> Remember when first all local AI was Python
>> and PyTorch APIs. And then suddently people started
>> using bare metal C/C++ Code. Here is the story:
>>
>> How it started:
>>
>> GPT-J or GPT-J-6B is an open-source large
>> language model (LLM) developed by EleutherAI
>> in 2021. As the name suggests, it is a
>> generative pre-trained transformer model
>> designed to produce human-like text that
>> continues from a prompt.
>> https://www.eleuther.ai/
>>
>> How it was going [Georgi Gerganov]:
>>
>> So a few days later comes out the LLaMA, I do
>> some calculations and I figure out “Okay, 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?”
>>
>> Why I was able to do it so quickly - basically,
>> for all that I saw it’s pretty much GPT-J architecture
>> with some modifications, like some extra memorization
>> layers. It’s 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
> 

[toc] | [prev] | [next] | [standalone]

#15765 — Rossy Boy is neither Einstein nor Zweistein (Was: Potato Computer owner impressed by Ukraine Tech)

Hi,

Rossy Boy is neither Einstein nor Zweistein.
He is not Einstein since Einstein is already dead:

Albert Einstein (1879 - 1955)
https://de.wikipedia.org/wiki/Albert_Einstein

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

- NVIDIA Volta ff. architecture

Also his hands are small, and his breath stinks,
and he lives in the basement of his mother.

Bye

Mild Shock schrieb:
> Hi,
> 
> Slowly I start understanding numbnuts like
> Rossy Boy who don't understand tech, although
> they are from UK and not from a 3rd world
> 
> country, and also I start understanding morons
> like Micro Penis, who are behind a curtain,
> and cannot access a lot of tech.
> 
> The same holds for SWI Prologs newest campaign
> that probably adresses some poor indians that
> have neither 5G nor Macs:
> 
> 1:38:01 The Kyiv keynote disaster
> https://www.youtube.com/watch?v=U8goS6B3BbI
> 
> Woa! Real time download of Scala, Closure,
> etc.. Whats the magic behind that? Some SWI
> point of sale, downloading it via its
> 
> keyboard and some telephathy module ?
> 
> Bye
> 
> Mild Shock schrieb:
>> Hi,
>>
>> Ride the snake
>> He's old and his skin is cold
>> The west is the best
>> The west is the best
>> Get here and we'll do the rest
>> The blue bus is calling us
>> The blue bus is calling us
>> Driver, where you taking us?
>>
>> Apocalypse Now intro: The Doors, The End {1979}
>> https://www.youtube.com/watch?v=CIrvSJwwJUE
>>
>> Bye
>>
>>  > Hi,
>>  >
>>  > Again I posted everything here:
>>  >
>>  >> 11.4 Giga Lips with a Budget Laptop
>>  >> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>  >
>>  > The repo says, same time when I posted
>>  > the link first time:
>>  >
>>  >> This repository was archived by the
>>  >> owner on Jul 9, 2026. It is now read-only.
>>  >
>>  > Now a USENET user, who had already entitled
>>  > himself for a couple of irrational accusations
>>  >
>>  > towards my side, is asking this question:
>>  >
>>  > Chris M. Thomasson schrieb, Jul 24, 2026
>>  >> Show an outline of what you
>>  >> need you compute shader to do?
>>  >
>>  > Bravo, thats a delay of a wooping 15 days.
>>  >
>>  > Bye
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>> Remember when first all local AI was Python
>>> and PyTorch APIs. And then suddently people started
>>> using bare metal C/C++ Code. Here is the story:
>>>
>>> How it started:
>>>
>>> GPT-J or GPT-J-6B is an open-source large
>>> language model (LLM) developed by EleutherAI
>>> in 2021. As the name suggests, it is a
>>> generative pre-trained transformer model
>>> designed to produce human-like text that
>>> continues from a prompt.
>>> https://www.eleuther.ai/
>>>
>>> How it was going [Georgi Gerganov]:
>>>
>>> So a few days later comes out the LLaMA, I do
>>> some calculations and I figure out “Okay, 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?”
>>>
>>> Why I was able to do it so quickly - basically,
>>> for all that I saw it’s pretty much GPT-J architecture
>>> with some modifications, like some extra memorization
>>> layers. It’s 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
>>
> 

[toc] | [prev] | [standalone]

Page 2 of 2 — ← Prev page 1 [2]

Back to top | Article view | comp.lang.prolog

csiph-web