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

Groups > sci.physics.relativity > #671475 > unrolled thread

I'm a spinner, I'm a sinner [Dmitry Vyukov for pi-WAM] (Was: Paul Tarau versus Mr. Taskmanager, who would win? [A PDP-11 Humunkulus from 1979])

Back to article view | Back to sci.physics.relativity

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  I'm a spinner, I'm a sinner [Dmitry Vyukov for pi-WAM] (Was: Paul Tarau versus Mr. Taskmanager, who would win? [A PDP-11 Humunkulus from 1979]) Mild Shock <janburse@fastmail.fm> - 2026-07-19 11:53 +0200
    Corr.: Re: I'm a spinner, I'm a sinner [Dmitry Vyukov for pi-WAM] (Was: Paul Tarau versus Mr. Taskmanager, who would win? [A PDP-11 Humunkulus from 1979]) Mild Shock <janburse@fastmail.fm> - 2026-07-19 11:55 +0200
    Re: I'm a spinner, I'm a sinner [Dmitry Vyukov for pi-WAM] (Was: Paul Tarau versus Mr. Taskmanager, who would win? [A PDP-11 Humunkulus from 1979]) "Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> - 2026-07-19 14:04 -0700
      Re: I'm a spinner, I'm a sinner [Dmitry Vyukov for pi-WAM] (Was: Paul Tarau versus Mr. Taskmanager, who would win? [A PDP-11 Humunkulus from 1979]) "Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> - 2026-07-19 14:07 -0700
        Gemini, DeepSeek, OpenAI all know Dmitriy V'jukov (Was: I'm a spinner, I'm a sinner [Dmitry Vyukov for pi-WAM]) Mild Shock <janburse@fastmail.fm> - 2026-07-20 08:31 +0200
          Re: Gemini, DeepSeek, OpenAI all know Dmitriy V'jukov (Was: I'm a spinner, I'm a sinner [Dmitry Vyukov for pi-WAM]) Romelio Balakhonsky <lrel@lao.ru> - 2026-07-20 10:07 +0000
            The Cache Identity Crisis by Micro Penis (Re: Gemini, DeepSeek, OpenAI all know Dmitriy V'jukov) Mild Shock <janburse@fastmail.fm> - 2026-07-20 13:51 +0200
              Just RTFM the RDNA 3.5 specs! [GPU Cache Lines] (Was: The Cache Identity Crisis by Micro Penis) Mild Shock <janburse@fastmail.fm> - 2026-07-20 14:09 +0200
                The large memory tax: ECC RAM (Was: Just RTFM the RDNA 3.5 specs! [GPU Cache Lines]) Mild Shock <janburse@fastmail.fm> - 2026-07-20 14:22 +0200
                Friendly Reminder: GPU 10x more performant than CPU (Re: Just RTFM the RDNA 3.5 specs! [GPU Cache Lines]) Mild Shock <janburse@fastmail.fm> - 2026-07-21 00:40 +0200
                  Breaking the CUDA edge in AI by WebGPU (Re: Friendly Reminder: GPU 10x more performant than CPU) Mild Shock <janburse@fastmail.fm> - 2026-07-21 00:57 +0200
                    Like WebAssembly before it, WebGPU has "escaped" the browser. (Re: Breaking the CUDA edge in AI by WebGPU (Re: Friendly Reminder: GPU 10x more performant than CPU) Mild Shock <janburse@fastmail.fm> - 2026-07-21 01:07 +0200
                    Re: Breaking the CUDA edge in AI by WebGPU (Re: Friendly Reminder: GPU 10x more performant than CPU) Will Bakshandaev <bev@lwesi.ru> - 2026-07-21 14:33 +0000
                      http://localhost:567921/ is a private REST endpoint [Teaching Micro Penis Vilage Idiot] (Was: Breaking the CUDA edge in AI by WebGPU) Mild Shock <janburse@fastmail.fm> - 2026-07-21 22:41 +0200
                        If you are paranoid you can use Falco [Agentic AI] (Re: http://localhost:567921/ is a private REST endpoint) Mild Shock <janburse@fastmail.fm> - 2026-07-21 22:57 +0200
                          What would an EMACs guru say [Windows Recall] (Was: If you are paranoid you can use Falco [Agentic AI]) Mild Shock <janburse@fastmail.fm> - 2026-07-21 23:33 +0200
                        Re: http://localhost:567921/ is a private REST endpoint [Teaching Micro Penis Vilage Idiot] (Was: Breaking the CUDA edge in AI by WebGPU) Hants Baibikov <vi@bi.ru> - 2026-07-21 21:51 +0000
                        Re: http://localhost:567921/ is a private REST endpoint [Teaching Micro Penis Vilage Idiot] (Was: Breaking the CUDA edge in AI by WebGPU) Pascual Talbaev <ps@laalapa.ru> - 2026-07-21 22:02 +0000
                        Decide what you critique tiny winy penis (Was: http://localhost:567921/ is a private REST endpoint) Mild Shock <janburse@fastmail.fm> - 2026-07-22 08:13 +0200
                          How confused is tiny winy penis? (Re: Decide what you critique tiny winy penis) Mild Shock <janburse@fastmail.fm> - 2026-07-22 08:29 +0200
                            Maybe change your hobby, become a dog owner? (Re: How confused is tiny winy penis?) Mild Shock <janburse@fastmail.fm> - 2026-07-22 09:38 +0200
                          Re: Decide what you critique tiny winy penis (Was: http://localhost:567921/ is a private REST endpoint) Audie Balaban <aie@ndabl.ru> - 2026-07-22 08:02 +0000
                            Even dogs know Switzerland != Germany [Syphilis Brain Micro Penis] (Re: Decide what you critique tiny winy penis (Was: http://localhost:567921/ is a private REST endpoint) Mild Shock <janburse@fastmail.fm> - 2026-07-22 11:17 +0200
                              Re: Even dogs know Switzerland != Germany [Syphilis Brain Micro Penis] (Re: Decide what you critique tiny winy penis (Was: http://localhost:567921/ is a private REST endpoint) Roque Bahtinov <aoqhi@hrot.ru> - 2026-07-22 12:04 +0000
                                My Swift Go 16 AI has no IMEI, are you nuts? (Was: Even dogs know Switzerland != Germany [Syphilis Brain Micro Penis]) Mild Shock <janburse@fastmail.fm> - 2026-07-22 14:12 +0200
                                  Same nickname and email, could post faster [5 year old moron] (Re: My Swift Go 16 AI has no IMEI, are you nuts?) Mild Shock <janburse@fastmail.fm> - 2026-07-22 14:24 +0200
                                  Re: My Swift Go 16 AI has no IMEI, are you nuts? (Was: Even dogs know Switzerland != Germany [Syphilis Brain Micro Penis]) Randolf Mukanov <mfroa@unvfo.ru> - 2026-07-22 12:27 +0000
                                    I have nothing to hide, you can find me in search.ch (Was: My Swift Go 16 AI has no IMEI, are you nuts?) Mild Shock <janburse@fastmail.fm> - 2026-07-22 14:31 +0200
                                      Re: I have nothing to hide, you can find me in search.ch (Was: My Swift Go 16 AI has no IMEI, are you nuts?) Keiv Babenchikov <hi@babebek.ru> - 2026-07-22 12:35 +0000
                                        Where did I confirm German via .ch, you are more than nuts! (Was: I have nothing to hide, you can find me in search.ch) Mild Shock <janburse@fastmail.fm> - 2026-07-22 14:46 +0200
                                          Ask a Ukrainian Neighbour to do Detective [CCCP Troll] (Was: Where did I confirm German via .ch, you are more than nuts!) Mild Shock <janburse@fastmail.fm> - 2026-07-22 14:51 +0200
              Re: The Cache Identity Crisis by Micro Penis (Re: Gemini, DeepSeek, OpenAI all know Dmitriy V'jukov) Jeiker Makulov <rmkru@eeeamu.ru> - 2026-07-20 16:18 +0000
                L1,..,Ln caches are located on the CPU AND on the GPU (Was: The Cache Identity Crisis by Micro Penis) Mild Shock <janburse@fastmail.fm> - 2026-07-20 19:33 +0200
                  GPU Cache Hierarchy: Understanding L1, L2, and VRAM (Re: L1,..,Ln caches are located on the CPU AND on the GPU) Mild Shock <janburse@fastmail.fm> - 2026-07-20 19:42 +0200
                    Re: GPU Cache Hierarchy: Understanding L1, L2, and VRAM (Re: L1,..,Ln caches are located on the CPU AND on the GPU) Zackee Mulatov <azauv@omtla.ru> - 2026-07-20 19:09 +0000
                      Well thats good, co-location, onto the same processor die (Was: GPU Cache Hierarchy: Understanding L1, L2, and VRAM) Mild Shock <janburse@fastmail.fm> - 2026-07-20 22:02 +0200
                        Where is micro penis mental error? (Re: Well thats good, co-location, onto the same processor die) Mild Shock <janburse@fastmail.fm> - 2026-07-20 22:07 +0200
                        Re: Well thats good, co-location, onto the same processor die (Was: GPU Cache Hierarchy: Understanding L1, L2, and VRAM) Hermis Molochkov <me@olech.ru> - 2026-07-20 22:27 +0000
          Re: Gemini, DeepSeek, OpenAI all know Dmitriy V'jukov (Was: I'm a spinner, I'm a sinner [Dmitry Vyukov for pi-WAM]) "Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> - 2026-07-20 13:40 -0700
            I didn't find Futex in WebGPU / WGSL (Was: Gemini, DeepSeek, OpenAI all know Dmitriy V'jukov) Mild Shock <janburse@fastmail.fm> - 2026-07-20 23:25 +0200
              Re: I didn't find Futex in WebGPU / WGSL (Was: Gemini, DeepSeek, OpenAI all know Dmitriy V'jukov) "Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> - 2026-07-20 14:32 -0700
                Re: I didn't find Futex in WebGPU / WGSL (Was: Gemini, DeepSeek, OpenAI all know Dmitriy V'jukov) "Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> - 2026-07-20 14:35 -0700
                  There is no imageAtomicAdd in WGSL (Was: I didn't find Futex in WebGPU / WGSL) Mild Shock <janburse@fastmail.fm> - 2026-07-21 00:12 +0200
                    OpenGL is dead. Apple said bye bye / Wayland Compositor (Was: There is no imageAtomicAdd in WGSL) Mild Shock <janburse@fastmail.fm> - 2026-07-21 00:24 +0200
                    Re: There is no imageAtomicAdd in WGSL (Was: I didn't find Futex in WebGPU / WGSL) "Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> - 2026-07-20 16:15 -0700
                      Flogging a Dead Horse, OpenGL is EOL (Was: There is no imageAtomicAdd in WGSL) Mild Shock <janburse@fastmail.fm> - 2026-07-21 01:24 +0200
                        imageAtomicAdd trivial, Dmitry Vyukov requires capacity (Re: Flogging a Dead Horse, OpenGL is EOL (Was: There is no imageAtomicAdd in WGSL) Mild Shock <janburse@fastmail.fm> - 2026-07-21 01:32 +0200
                          capacity = 2^n for some n / systolic system (Was: imageAtomicAdd trivial, Dmitry Vyukov requires capacity) Mild Shock <janburse@fastmail.fm> - 2026-07-21 01:37 +0200
                            Source of the benchmark for DmitryVyukov (Re: capacity = 2^n for some n / systolic system) Mild Shock <janburse@fastmail.fm> - 2026-07-21 01:45 +0200
                          Re: imageAtomicAdd trivial, Dmitry Vyukov requires capacity (Re: Flogging a Dead Horse, OpenGL is EOL (Was: There is no imageAtomicAdd in WGSL) "Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> - 2026-07-20 17:00 -0700
                            I never used OpenGL Version 4.2 and later (Re: imageAtomicAdd trivial, Dmitry Vyukov requires capacity) Mild Shock <janburse@fastmail.fm> - 2026-07-21 08:49 +0200
                              Because of MIMD you have to reassess algorithms (Was: I never used OpenGL Version 4.2 and later) Mild Shock <janburse@fastmail.fm> - 2026-07-21 08:59 +0200
                                Why MIMD is interesting for pi-WAM? Mild Shock <janburse@fastmail.fm> - 2026-07-21 09:16 +0200
                                  Re: Why MIMD is interesting for pi-WAM? Ross Finlayson <ross.a.finlayson@gmail.com> - 2026-07-21 00:32 -0700
                                Re: Because of MIMD you have to reassess algorithms (Was: I never used OpenGL Version 4.2 and later) Ross Finlayson <ross.a.finlayson@gmail.com> - 2026-07-21 00:42 -0700
                                  You still don't understand "budget" [Rossy Boy slower than Micro Penis] (Was: Because of MIMD you have to reassess algorithms) Mild Shock <janburse@fastmail.fm> - 2026-07-21 10:13 +0200
                                    Go on Rossy Boy, ask more stupid questions (Was: You still don't understand "budget" [Rossy Boy slower than Micro Penis]) Mild Shock <janburse@fastmail.fm> - 2026-07-21 10:17 +0200
                                      Need to be Einstein to understand Giga Lips (Was: Go on Rossy Boy, ask more stupid questions) Mild Shock <janburse@fastmail.fm> - 2026-07-21 10:22 +0200
                                        Marketing invents Gucci Bag AI Laptops (Was: Need to be Einstein to understand Giga Lips) Mild Shock <janburse@fastmail.fm> - 2026-07-21 10:43 +0200
                                      Re: Go on Rossy Boy, ask more stupid questions (Was: You still don't understand "budget" [Rossy Boy slower than Micro Penis]) Ross Finlayson <ross.a.finlayson@gmail.com> - 2026-07-21 08:59 -0700
                                        Rossy Boy says I am a crazy frothing lunatic (Was: Go on Rossy Boy, ask more stupid questions) Mild Shock <janburse@fastmail.fm> - 2026-07-21 22:27 +0200
                                          Re: Rossy Boy says I am a crazy frothing lunatic (Was: Go on Rossy Boy, ask more stupid questions) Ross Finlayson <ross.a.finlayson@gmail.com> - 2026-07-21 13:52 -0700

Page 1 of 4  [1] 2 3 4  Next page →

#671475 — I'm a spinner, I'm a sinner [Dmitry Vyukov for pi-WAM] (Was: Paul Tarau versus Mr. Taskmanager, who would win? [A PDP-11 Humunkulus from 1979])

Hi,

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

2 Producer threads, 2 Consumer threads
Each Producer generating 8192 items
the single common channel between
them 128 elements larger.

public class DmitryVyukov
8 ms
public class DougLea
10 ms

On the Ryzen AI 7 350 which has 8 physical
cores with hyperthreading, making it 16
logical cores.

LoL

Bye

See also:

Producer-Consumer Queues by Dmitry Vyukov
https://sites.google.com/site/1024cores/home/lock-free-algorithms/queues

Steve Miller Band - The Joker (Official Music Video)
https://www.youtube.com/watch?v=dV3AziKTBUo

Mild Shock schrieb:
> Hi,
> 
> Ok I was looking at this learning challenge,
> producing vector (y1,y2,y3,y4) from a vector
> (x1,x2,x3,x4), System R can do it via least square?
> 
> | 0 0 0 1 |   | x1 |     | x4 |
> | 0 0 1 0 |   | x2 |  =  | x3 |
> | 0 1 0 0 |   | x3 |     | x2 |
> | 1 0 0 0 |   | x4 |     | x1 |
> 
> How it started:
> 
> "multiplicative RNNs arises naturally from a
> proof-theoretic interpretation of next-token
> prediction as nested intuitionistic implication"
> Paul Tarau - 2026
> https://arxiv.org/abs/2601.19915
> 
> How its going:
> 
> "Dave uses a PDP-11 to train a real Neural
> Network complete with Transformers and
> Attention so you can see them at their most basic."
> Mr. Taskmanager - 2026
> https://www.youtube.com/watch?v=OUE3FSIk46g
> 
> We see Doctor Frankstein in action from
> the Bronze Age of Computing, producing
> a Humunkulus, the progenitor of todays
> 
> Bulgakov Shuriks in the Hyperscale Age!
> 
> Bye
> 
> P.S.: My impression neither cut to the core, that
> this incredible transformer most likely
> produced this deterministic attention:
> 
> | -1 | * | k | + | 5 | = | k' |
> 
> Or differently expressed y_k = x_{5-k}.
> 
> How did the transformer do it? It produced
> a neural network with 1216 parameters, but
> didn't use embeddings or polar encoding
> 
> of positions. But if we strip the noise
> and denoise from the position encoding,
> the denoise is done via softmax. We somehow
> 
> must get the above, right? I still need to
> verify my claim! BTW: The PDP-11 assembly
> from 1979 uses wider example not with n=4
> 
> but with n=8.

[toc] | [next] | [standalone]

#671476 — Corr.: Re: I'm a spinner, I'm a sinner [Dmitry Vyukov for pi-WAM] (Was: Paul Tarau versus Mr. Taskmanager, who would win? [A PDP-11 Humunkulus from 1979])

Corr.:

 > them 128 elements larger.
them 128 elements large.

Mild Shock schrieb:
> Hi,
> 
> I'm a spinner, I'm a sinner
> I spin on CAS loops for my dinner
> Some call it busy-wait, I call it fate
> When the queue is empty, I just rotate
> 
> 2 Producer threads, 2 Consumer threads
> Each Producer generating 8192 items
> the single common channel between
> them 128 elements larger.
> 
> public class DmitryVyukov
> 8 ms
> public class DougLea
> 10 ms
> 
> On the Ryzen AI 7 350 which has 8 physical
> cores with hyperthreading, making it 16
> logical cores.
> 
> LoL
> 
> Bye
> 
> See also:
> 
> Producer-Consumer Queues by Dmitry Vyukov
> https://sites.google.com/site/1024cores/home/lock-free-algorithms/queues
> 
> Steve Miller Band - The Joker (Official Music Video)
> https://www.youtube.com/watch?v=dV3AziKTBUo
> 
> Mild Shock schrieb:
>> Hi,
>>
>> Ok I was looking at this learning challenge,
>> producing vector (y1,y2,y3,y4) from a vector
>> (x1,x2,x3,x4), System R can do it via least square?
>>
>> | 0 0 0 1 |   | x1 |     | x4 |
>> | 0 0 1 0 |   | x2 |  =  | x3 |
>> | 0 1 0 0 |   | x3 |     | x2 |
>> | 1 0 0 0 |   | x4 |     | x1 |
>>
>> How it started:
>>
>> "multiplicative RNNs arises naturally from a
>> proof-theoretic interpretation of next-token
>> prediction as nested intuitionistic implication"
>> Paul Tarau - 2026
>> https://arxiv.org/abs/2601.19915
>>
>> How its going:
>>
>> "Dave uses a PDP-11 to train a real Neural
>> Network complete with Transformers and
>> Attention so you can see them at their most basic."
>> Mr. Taskmanager - 2026
>> https://www.youtube.com/watch?v=OUE3FSIk46g
>>
>> We see Doctor Frankstein in action from
>> the Bronze Age of Computing, producing
>> a Humunkulus, the progenitor of todays
>>
>> Bulgakov Shuriks in the Hyperscale Age!
>>
>> Bye
>>
>> P.S.: My impression neither cut to the core, that
>> this incredible transformer most likely
>> produced this deterministic attention:
>>
>> | -1 | * | k | + | 5 | = | k' |
>>
>> Or differently expressed y_k = x_{5-k}.
>>
>> How did the transformer do it? It produced
>> a neural network with 1216 parameters, but
>> didn't use embeddings or polar encoding
>>
>> of positions. But if we strip the noise
>> and denoise from the position encoding,
>> the denoise is done via softmax. We somehow
>>
>> must get the above, right? I still need to
>> verify my claim! BTW: The PDP-11 assembly
>> from 1979 uses wider example not with n=4
>>
>> but with n=8.
> 

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

#671479

On 7/19/2026 2:53 AM, Mild Shock wrote:
> Hi,
> 
> I'm a spinner, I'm a sinner
> I spin on CAS loops for my dinner
> Some call it busy-wait, I call it fate
> When the queue is empty, I just rotate
> 
> 2 Producer threads, 2 Consumer threads
> Each Producer generating 8192 items
> the single common channel between
> them 128 elements larger.
> 
> public class DmitryVyukov
> 8 ms
> public class DougLea
> 10 ms
> 
> On the Ryzen AI 7 350 which has 8 physical
> cores with hyperthreading, making it 16
> logical cores.
> 
> LoL
> 
> Bye
> 
> See also:
> 
> Producer-Consumer Queues by Dmitry Vyukov
> https://sites.google.com/site/1024cores/home/lock-free-algorithms/queues
> 
> Steve Miller Band - The Joker (Official Music Video)
> https://www.youtube.com/watch?v=dV3AziKTBUo
> 
> Mild Shock schrieb:
>> Hi,
>>
>> Ok I was looking at this learning challenge,
>> producing vector (y1,y2,y3,y4) from a vector
>> (x1,x2,x3,x4), System R can do it via least square?
>>
>> | 0 0 0 1 |   | x1 |     | x4 |
>> | 0 0 1 0 |   | x2 |  =  | x3 |
>> | 0 1 0 0 |   | x3 |     | x2 |
>> | 1 0 0 0 |   | x4 |     | x1 |
>>
>> How it started:
>>
>> "multiplicative RNNs arises naturally from a
>> proof-theoretic interpretation of next-token
>> prediction as nested intuitionistic implication"
>> Paul Tarau - 2026
>> https://arxiv.org/abs/2601.19915
>>
>> How its going:
>>
>> "Dave uses a PDP-11 to train a real Neural
>> Network complete with Transformers and
>> Attention so you can see them at their most basic."
>> Mr. Taskmanager - 2026
>> https://www.youtube.com/watch?v=OUE3FSIk46g
>>
>> We see Doctor Frankstein in action from
>> the Bronze Age of Computing, producing
>> a Humunkulus, the progenitor of todays
>>
>> Bulgakov Shuriks in the Hyperscale Age!
>>
>> Bye
>>
>> P.S.: My impression neither cut to the core, that
>> this incredible transformer most likely
>> produced this deterministic attention:
>>
>> | -1 | * | k | + | 5 | = | k' |
>>
>> Or differently expressed y_k = x_{5-k}.
>>
>> How did the transformer do it? It produced
>> a neural network with 1216 parameters, but
>> didn't use embeddings or polar encoding
>>
>> of positions. But if we strip the noise
>> and denoise from the position encoding,
>> the denoise is done via softmax. We somehow
>>
>> must get the above, right? I still need to
>> verify my claim! BTW: The PDP-11 assembly
>> from 1979 uses wider example not with n=4
>>
>> but with n=8.
> 

I am friends with Dmitry Vyukov from way back, 23+ years ago. Actually, 
I helped him find some bugs in Relacy when it was in pre-alpha way back 
on comp.programming.threads. Also, I created some neat eventcount algos, 
read all:

https://gist.github.com/mratsim/04a29bdd98d6295acda4d0677c4d0041

For starters... ;^)

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

#671480

On 7/19/2026 2:04 PM, Chris M. Thomasson wrote:
> On 7/19/2026 2:53 AM, Mild Shock wrote:
>> Hi,
>>
>> I'm a spinner, I'm a sinner
>> I spin on CAS loops for my dinner
>> Some call it busy-wait, I call it fate
>> When the queue is empty, I just rotate
>>
>> 2 Producer threads, 2 Consumer threads
>> Each Producer generating 8192 items
>> the single common channel between
>> them 128 elements larger.
>>
>> public class DmitryVyukov
>> 8 ms
>> public class DougLea
>> 10 ms
>>
>> On the Ryzen AI 7 350 which has 8 physical
>> cores with hyperthreading, making it 16
>> logical cores.
>>
>> LoL
>>
>> Bye
>>
>> See also:
>>
>> Producer-Consumer Queues by Dmitry Vyukov
>> https://sites.google.com/site/1024cores/home/lock-free-algorithms/queues
>>
>> Steve Miller Band - The Joker (Official Music Video)
>> https://www.youtube.com/watch?v=dV3AziKTBUo
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>> Ok I was looking at this learning challenge,
>>> producing vector (y1,y2,y3,y4) from a vector
>>> (x1,x2,x3,x4), System R can do it via least square?
>>>
>>> | 0 0 0 1 |   | x1 |     | x4 |
>>> | 0 0 1 0 |   | x2 |  =  | x3 |
>>> | 0 1 0 0 |   | x3 |     | x2 |
>>> | 1 0 0 0 |   | x4 |     | x1 |
>>>
>>> How it started:
>>>
>>> "multiplicative RNNs arises naturally from a
>>> proof-theoretic interpretation of next-token
>>> prediction as nested intuitionistic implication"
>>> Paul Tarau - 2026
>>> https://arxiv.org/abs/2601.19915
>>>
>>> How its going:
>>>
>>> "Dave uses a PDP-11 to train a real Neural
>>> Network complete with Transformers and
>>> Attention so you can see them at their most basic."
>>> Mr. Taskmanager - 2026
>>> https://www.youtube.com/watch?v=OUE3FSIk46g
>>>
>>> We see Doctor Frankstein in action from
>>> the Bronze Age of Computing, producing
>>> a Humunkulus, the progenitor of todays
>>>
>>> Bulgakov Shuriks in the Hyperscale Age!
>>>
>>> Bye
>>>
>>> P.S.: My impression neither cut to the core, that
>>> this incredible transformer most likely
>>> produced this deterministic attention:
>>>
>>> | -1 | * | k | + | 5 | = | k' |
>>>
>>> Or differently expressed y_k = x_{5-k}.
>>>
>>> How did the transformer do it? It produced
>>> a neural network with 1216 parameters, but
>>> didn't use embeddings or polar encoding
>>>
>>> of positions. But if we strip the noise
>>> and denoise from the position encoding,
>>> the denoise is done via softmax. We somehow
>>>
>>> must get the above, right? I still need to
>>> verify my claim! BTW: The PDP-11 assembly
>>> from 1979 uses wider example not with n=4
>>>
>>> but with n=8.
>>
> 
> I am friends with Dmitry Vyukov from way back, 23+ years ago. Actually, 
> I helped him find some bugs in Relacy when it was in pre-alpha way back 
> on comp.programming.threads. Also, I created some neat eventcount algos, 
> read all:
> 
> https://gist.github.com/mratsim/04a29bdd98d6295acda4d0677c4d0041
> 
> For starters... ;^)


also, I made a neat alteration to one of Dmitry Vyukov MPMC queues, 
check it out:

https://groups.google.com/g/lock-free/c/acjQ3-89abE/m/a6-Di0GZsyEJ

can you get to the link? Thanks. Read all.

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

#671488 — Gemini, DeepSeek, OpenAI all know Dmitriy V'jukov (Was: I'm a spinner, I'm a sinner [Dmitry Vyukov for pi-WAM])

Hi,

Its actually quite amazing. Gemini, DeepSeek,
OpenAI all know Dmitriy V'jukov. I have asked
the IntelliJ integrated Freeium AI to generate

some code for me, I guess their service uses
by default OpenAI (Codex), and had it reviewed
by Gemini and DeepSeek. These AIs started lecturing

me about lazySet() in Java. But I went with set():

     private static boolean enqueue(Queue q, Object data) {
         int pos = q.enqueuePos.get();
         for (; ; ) {
             int index = pos & q.bufferMask;
             int seq = q.sequences.get(index);
             int dif = seq - pos;
             if (dif == 0) {
                 if (q.enqueuePos.compareAndSet(pos, pos + 1)) {
                     q.data[index] = data;
                     q.sequences.set(index, pos + 1);
                     return true;
                 }
                 pos = q.enqueuePos.get();
             } else if (dif < 0) {
                 return false;
             } else {
                 pos = q.enqueuePos.get();
             }
         }
     }

The above version seems to be more suitable
for my purpose, since it allows polling, it
basically implements offer(). While the

version posted on in the lock free group
by Chris M. Thomasson implements a spin wait
blocking put() already.

But I didn't port it yet to JavaScript or WGSL.
Currently busy with realizing a Worker facade
in JavaScript for a CPU backend that will run

in both the browser and node.js.

Bye

Chris M. Thomasson schrieb:
> On 7/19/2026 2:04 PM, Chris M. Thomasson wrote:
>> On 7/19/2026 2:53 AM, Mild Shock wrote:
>>> Hi,
>>>
>>> I'm a spinner, I'm a sinner
>>> I spin on CAS loops for my dinner
>>> Some call it busy-wait, I call it fate
>>> When the queue is empty, I just rotate
>>>
>>> 2 Producer threads, 2 Consumer threads
>>> Each Producer generating 8192 items
>>> the single common channel between
>>> them 128 elements larger.
>>>
>>> public class DmitryVyukov
>>> 8 ms
>>> public class DougLea
>>> 10 ms
>>>
>>> On the Ryzen AI 7 350 which has 8 physical
>>> cores with hyperthreading, making it 16
>>> logical cores.
>>>
>>> LoL
>>>
>>> Bye
>>>
>>> See also:
>>>
>>> Producer-Consumer Queues by Dmitry Vyukov
>>> https://sites.google.com/site/1024cores/home/lock-free-algorithms/queues
>>>
>>> Steve Miller Band - The Joker (Official Music Video)
>>> https://www.youtube.com/watch?v=dV3AziKTBUo
>>>
>>> Mild Shock schrieb:
>>>> Hi,
>>>>
>>>> Ok I was looking at this learning challenge,
>>>> producing vector (y1,y2,y3,y4) from a vector
>>>> (x1,x2,x3,x4), System R can do it via least square?
>>>>
>>>> | 0 0 0 1 |   | x1 |     | x4 |
>>>> | 0 0 1 0 |   | x2 |  =  | x3 |
>>>> | 0 1 0 0 |   | x3 |     | x2 |
>>>> | 1 0 0 0 |   | x4 |     | x1 |
>>>>
>>>> How it started:
>>>>
>>>> "multiplicative RNNs arises naturally from a
>>>> proof-theoretic interpretation of next-token
>>>> prediction as nested intuitionistic implication"
>>>> Paul Tarau - 2026
>>>> https://arxiv.org/abs/2601.19915
>>>>
>>>> How its going:
>>>>
>>>> "Dave uses a PDP-11 to train a real Neural
>>>> Network complete with Transformers and
>>>> Attention so you can see them at their most basic."
>>>> Mr. Taskmanager - 2026
>>>> https://www.youtube.com/watch?v=OUE3FSIk46g
>>>>
>>>> We see Doctor Frankstein in action from
>>>> the Bronze Age of Computing, producing
>>>> a Humunkulus, the progenitor of todays
>>>>
>>>> Bulgakov Shuriks in the Hyperscale Age!
>>>>
>>>> Bye
>>>>
>>>> P.S.: My impression neither cut to the core, that
>>>> this incredible transformer most likely
>>>> produced this deterministic attention:
>>>>
>>>> | -1 | * | k | + | 5 | = | k' |
>>>>
>>>> Or differently expressed y_k = x_{5-k}.
>>>>
>>>> How did the transformer do it? It produced
>>>> a neural network with 1216 parameters, but
>>>> didn't use embeddings or polar encoding
>>>>
>>>> of positions. But if we strip the noise
>>>> and denoise from the position encoding,
>>>> the denoise is done via softmax. We somehow
>>>>
>>>> must get the above, right? I still need to
>>>> verify my claim! BTW: The PDP-11 assembly
>>>> from 1979 uses wider example not with n=4
>>>>
>>>> but with n=8.
>>>
>>
>> I am friends with Dmitry Vyukov from way back, 23+ years ago. 
>> Actually, I helped him find some bugs in Relacy when it was in 
>> pre-alpha way back on comp.programming.threads. Also, I created some 
>> neat eventcount algos, read all:
>>
>> https://gist.github.com/mratsim/04a29bdd98d6295acda4d0677c4d0041
>>
>> For starters... ;^)
> 
> 
> also, I made a neat alteration to one of Dmitry Vyukov MPMC queues, 
> check it out:
> 
> https://groups.google.com/g/lock-free/c/acjQ3-89abE/m/a6-Di0GZsyEJ
> 
> can you get to the link? Thanks. Read all.

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

#671489 — Re: Gemini, DeepSeek, OpenAI all know Dmitriy V'jukov (Was: I'm a spinner, I'm a sinner [Dmitry Vyukov for pi-WAM])

Mild Shock wrote:

> me about lazySet() in Java. But I went with set():
> 
>      private static boolean enqueue(Queue q, Object data) {
>          int pos = q.enqueuePos.get();
>          for (; ; ) {
>              int index = pos & q.bufferMask; int seq =
>              q.sequences.get(index);
>              int dif = seq - pos;
>              if (dif == 0) {
>                  if (q.enqueuePos.compareAndSet(pos, pos + 1)) {
>                      q.data[index] = data; q.sequences.set(index, pos +
>                      1);
>                      return true;
>                  }
>                  pos = q.enqueuePos.get();
>              } else if (dif < 0) {
>                  return false;
>              } else {
>                  pos = q.enqueuePos.get();
>              }
>          }
>      }
> 
> The above version seems to be more suitable for my purpose, since it
> allows polling, it basically implements offer(). While the

completely nonsense. Not even correct grammatically

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

#671494 — The Cache Identity Crisis by Micro Penis (Re: Gemini, DeepSeek, OpenAI all know Dmitriy V'jukov)

Hi,

 > completely nonsense. Not even correct grammatically

Yes of course, Micro Penis will know. Nothing, LoL
But here we find how village idiot Micro Penis is.
The Cache Identity Crisis by Micro Penis:

Claiming that L1/L2/L3 caches and core
topology are "embedded cpu registers area,
it has nothing to do with the ram." That is
a brilliant piece of hardware fiction.

Caches are SRAM acting as a hierarchical
staging ground for and backed by system
or device RAM, not an alternate dimension
divorced from memory entirely.

And calling cache hierarchies "registers"
is like confusing a warehouse parking lot
with the glove compartment of a single car.
Never heard of the crossbar in AMD GPUs?

LoL

Bye

Obelin Baisaroff schrieb:
 > Ross Finlayson wrote:
 >
 >> access to memory, has that these days with
 >> L1/L2/L3 caches and the proximity and affinity
 >> in the topology of the cores and
 >
 > those are embedded cpu registers area, it has
 > nothing to do with the ram. Works by higher
 > clocks compared, but merely useless in AI.
 >


Romelio Balakhonsky schrieb:
> Mild Shock wrote:
> 
>> me about lazySet() in Java. But I went with set():
>>
>>       private static boolean enqueue(Queue q, Object data) {
>>           int pos = q.enqueuePos.get();
>>           for (; ; ) {
>>               int index = pos & q.bufferMask; int seq =
>>               q.sequences.get(index);
>>               int dif = seq - pos;
>>               if (dif == 0) {
>>                   if (q.enqueuePos.compareAndSet(pos, pos + 1)) {
>>                       q.data[index] = data; q.sequences.set(index, pos +
>>                       1);
>>                       return true;
>>                   }
>>                   pos = q.enqueuePos.get();
>>               } else if (dif < 0) {
>>                   return false;
>>               } else {
>>                   pos = q.enqueuePos.get();
>>               }
>>           }
>>       }
>>
>> The above version seems to be more suitable for my purpose, since it
>> allows polling, it basically implements offer(). While the
> 
> completely nonsense. Not even correct grammatically
> 

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

#671495 — Just RTFM the RDNA 3.5 specs! [GPU Cache Lines] (Was: The Cache Identity Crisis by Micro Penis)

Hi,

If you don't know how GPU caches work.
Just RTFM the RDNA 3.5 specs! They have
some explanations in the docs, what jobs

the GPU cache lines do, in relation to
what instructions:

Abbreviation for ‘Read The Fucking Manual’.
http://www.catb.org/esr/jargon/html/R/RTFM.html

Here some RDNA 4.0 specs (smaller GPUs):

"RDNA4" Instruction Set Architecture
Reference Guide - 7-April-2025
https://docs.amd.com/v/u/en-US/rdna4-instruction-set-architecture

Here some CDNA 4.0 specs (bigger GPUs):

CDNA4 Instruction Set Architecture
Reference Guide - 5-August-2025
https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/instruction-set-architectures/amd-instinct-cdna4-instruction-set-architecture.pdf

Bye

But the difference between CDNA GPUs and
RDNA GPUs seems not to be that big, at least
in 2025:

Commitment to High-Performance
Computing in the Age of AI
https://docs.amd.com/v/u/en-US/precision-at-scale-amd-whitepaper

While the Copilot+ threshold is 45 TFLOPs,
a MI355X, Datacenter, rack-mounted, liquid cooling,
has surely more memory, but only 157 TFLOPS.

And RTX 5090, Desktop, 600W+ TDP, massive cooling,
has also only 100 TFLOPS. So it looks these
Copilot+ AI Laptops are pretty swell, arent they?

Mild Shock schrieb:
> Hi,
> 
>  > completely nonsense. Not even correct grammatically
> 
> Yes of course, Micro Penis will know. Nothing, LoL
> But here we find how village idiot Micro Penis is.
> The Cache Identity Crisis by Micro Penis:
> 
> Claiming that L1/L2/L3 caches and core
> topology are "embedded cpu registers area,
> it has nothing to do with the ram." That is
> a brilliant piece of hardware fiction.
> 
> Caches are SRAM acting as a hierarchical
> staging ground for and backed by system
> or device RAM, not an alternate dimension
> divorced from memory entirely.
> 
> And calling cache hierarchies "registers"
> is like confusing a warehouse parking lot
> with the glove compartment of a single car.
> Never heard of the crossbar in AMD GPUs?
> 
> LoL
> 
> Bye
> 
> Obelin Baisaroff schrieb:
>  > Ross Finlayson wrote:
>  >
>  >> access to memory, has that these days with
>  >> L1/L2/L3 caches and the proximity and affinity
>  >> in the topology of the cores and
>  >
>  > those are embedded cpu registers area, it has
>  > nothing to do with the ram. Works by higher
>  > clocks compared, but merely useless in AI.
>  >
> 
> 
> Romelio Balakhonsky schrieb:
>> Mild Shock wrote:
>>
>>> me about lazySet() in Java. But I went with set():
>>>
>>>       private static boolean enqueue(Queue q, Object data) {
>>>           int pos = q.enqueuePos.get();
>>>           for (; ; ) {
>>>               int index = pos & q.bufferMask; int seq =
>>>               q.sequences.get(index);
>>>               int dif = seq - pos;
>>>               if (dif == 0) {
>>>                   if (q.enqueuePos.compareAndSet(pos, pos + 1)) {
>>>                       q.data[index] = data; q.sequences.set(index, pos +
>>>                       1);
>>>                       return true;
>>>                   }
>>>                   pos = q.enqueuePos.get();
>>>               } else if (dif < 0) {
>>>                   return false;
>>>               } else {
>>>                   pos = q.enqueuePos.get();
>>>               }
>>>           }
>>>       }
>>>
>>> The above version seems to be more suitable for my purpose, since it
>>> allows polling, it basically implements offer(). While the
>>
>> completely nonsense. Not even correct grammatically
>>
> 

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

#671496 — The large memory tax: ECC RAM (Was: Just RTFM the RDNA 3.5 specs! [GPU Cache Lines])

Hi,

in high-stakes environments—whether
it's enterprise servers running mission-
critical transactions or heavy parallel
compute clusters crunching massive datasets.

At scale, a "soft error" (a bit randomly
flipping from 0 to 1) isn't a rare anomaly;
it is a statistical certainty.

ECC requires extra bits per data word
(e.g., a 72-bit bus for 64 bits of data) a
nd a small computational cycle overhead to
compute and check parity bits on every access.

instead of instantly crashing the entire
system, modern processors use data poisoning.
The memory controller or cache logic stamps a
"poison bit" or signature onto that specific

cache line and propagates it along with
the corrupted data.

Bye

Mild Shock schrieb:
> Hi,
> 
> If you don't know how GPU caches work.
> Just RTFM the RDNA 3.5 specs! They have
> some explanations in the docs, what jobs
> 
> the GPU cache lines do, in relation to
> what instructions:
> 
> Abbreviation for ‘Read The Fucking Manual’.
> http://www.catb.org/esr/jargon/html/R/RTFM.html
> 
> Here some RDNA 4.0 specs (smaller GPUs):
> 
> "RDNA4" Instruction Set Architecture
> Reference Guide - 7-April-2025
> https://docs.amd.com/v/u/en-US/rdna4-instruction-set-architecture
> 
> Here some CDNA 4.0 specs (bigger GPUs):
> 
> CDNA4 Instruction Set Architecture
> Reference Guide - 5-August-2025
> https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/instruction-set-architectures/amd-instinct-cdna4-instruction-set-architecture.pdf 
> 
> 
> Bye
> 
> But the difference between CDNA GPUs and
> RDNA GPUs seems not to be that big, at least
> in 2025:
> 
> Commitment to High-Performance
> Computing in the Age of AI
> https://docs.amd.com/v/u/en-US/precision-at-scale-amd-whitepaper
> 
> While the Copilot+ threshold is 45 TFLOPs,
> a MI355X, Datacenter, rack-mounted, liquid cooling,
> has surely more memory, but only 157 TFLOPS.
> 
> And RTX 5090, Desktop, 600W+ TDP, massive cooling,
> has also only 100 TFLOPS. So it looks these
> Copilot+ AI Laptops are pretty swell, arent they?
> 
> Mild Shock schrieb:
>> Hi,
>>
>>  > completely nonsense. Not even correct grammatically
>>
>> Yes of course, Micro Penis will know. Nothing, LoL
>> But here we find how village idiot Micro Penis is.
>> The Cache Identity Crisis by Micro Penis:
>>
>> Claiming that L1/L2/L3 caches and core
>> topology are "embedded cpu registers area,
>> it has nothing to do with the ram." That is
>> a brilliant piece of hardware fiction.
>>
>> Caches are SRAM acting as a hierarchical
>> staging ground for and backed by system
>> or device RAM, not an alternate dimension
>> divorced from memory entirely.
>>
>> And calling cache hierarchies "registers"
>> is like confusing a warehouse parking lot
>> with the glove compartment of a single car.
>> Never heard of the crossbar in AMD GPUs?
>>
>> LoL
>>
>> Bye
>>
>> Obelin Baisaroff schrieb:
>>  > Ross Finlayson wrote:
>>  >
>>  >> access to memory, has that these days with
>>  >> L1/L2/L3 caches and the proximity and affinity
>>  >> in the topology of the cores and
>>  >
>>  > those are embedded cpu registers area, it has
>>  > nothing to do with the ram. Works by higher
>>  > clocks compared, but merely useless in AI.
>>  >
>>
>>
>> Romelio Balakhonsky schrieb:
>>> Mild Shock wrote:
>>>
>>>> me about lazySet() in Java. But I went with set():
>>>>
>>>>       private static boolean enqueue(Queue q, Object data) {
>>>>           int pos = q.enqueuePos.get();
>>>>           for (; ; ) {
>>>>               int index = pos & q.bufferMask; int seq =
>>>>               q.sequences.get(index);
>>>>               int dif = seq - pos;
>>>>               if (dif == 0) {
>>>>                   if (q.enqueuePos.compareAndSet(pos, pos + 1)) {
>>>>                       q.data[index] = data; q.sequences.set(index, 
>>>> pos +
>>>>                       1);
>>>>                       return true;
>>>>                   }
>>>>                   pos = q.enqueuePos.get();
>>>>               } else if (dif < 0) {
>>>>                   return false;
>>>>               } else {
>>>>                   pos = q.enqueuePos.get();
>>>>               }
>>>>           }
>>>>       }
>>>>
>>>> The above version seems to be more suitable for my purpose, since it
>>>> allows polling, it basically implements offer(). While the
>>>
>>> completely nonsense. Not even correct grammatically
>>>
>>
> 

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

#671515 — Friendly Reminder: GPU 10x more performant than CPU (Re: Just RTFM the RDNA 3.5 specs! [GPU Cache Lines])

Hi,

You guys are not paying attention. GPU is 10x
more performant than CPU for certain integerish
payload on the pi-WAM. 11.4 GLips on a GPU is

ca 10x more than 1.7 GLips on a CPU:

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

Parallel π-WAM: 1.7 Giga Lips on a CPU
https://medium.com/2989/8a984e75af44

So GPUs are by way not dead. Especially since
they can be accessed via WebGPU / WGSL. But
the line between GPU and CPU increasingly

blurrs. It already happens for APUs in
that unified memory puts them into the same
RAM space. But the programming means differ.

Also unified memory doesn't mean that
the GPU sees memory the same way as a CPU.
To get a taste how a GPU sees memory:

Memory Model
     Execution Barriers
     Fence and Address Spaces
     Memory Model GFX6-GFX9
     Memory Model GFX90A
     Memory Model GFX942
     Memory Model GFX10-GFX11
     Memory Model GFX12
     Memory Model GFX125x

https://llvm.org/docs/AMDGPUUsage.html

WebGPU / WGSL seems to have good support,
since it can have Vulkan, Direct 12, or Metal
as GPU low level interface. Maybe more

platforms, something NVIDIA I guess. But I
don't know the details, how this is all done.

Bye

Mild Shock schrieb:
> Hi,
> 
> If you don't know how GPU caches work.
> Just RTFM the RDNA 3.5 specs! They have
> some explanations in the docs, what jobs
> 
> the GPU cache lines do, in relation to
> what instructions:
> 
> Abbreviation for ‘Read The Fucking Manual’.
> http://www.catb.org/esr/jargon/html/R/RTFM.html
> 
> Here some RDNA 4.0 specs (smaller GPUs):
> 
> "RDNA4" Instruction Set Architecture
> Reference Guide - 7-April-2025
> https://docs.amd.com/v/u/en-US/rdna4-instruction-set-architecture
> 
> Here some CDNA 4.0 specs (bigger GPUs):
> 
> CDNA4 Instruction Set Architecture
> Reference Guide - 5-August-2025
> https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/instruction-set-architectures/amd-instinct-cdna4-instruction-set-architecture.pdf 
> 
> 
> Bye
> 
> But the difference between CDNA GPUs and
> RDNA GPUs seems not to be that big, at least
> in 2025:
> 
> Commitment to High-Performance
> Computing in the Age of AI
> https://docs.amd.com/v/u/en-US/precision-at-scale-amd-whitepaper
> 
> While the Copilot+ threshold is 45 TFLOPs,
> a MI355X, Datacenter, rack-mounted, liquid cooling,
> has surely more memory, but only 157 TFLOPS.
> 
> And RTX 5090, Desktop, 600W+ TDP, massive cooling,
> has also only 100 TFLOPS. So it looks these
> Copilot+ AI Laptops are pretty swell, arent they?
> 
> Mild Shock schrieb:
>> Hi,
>>
>>  > completely nonsense. Not even correct grammatically
>>
>> Yes of course, Micro Penis will know. Nothing, LoL
>> But here we find how village idiot Micro Penis is.
>> The Cache Identity Crisis by Micro Penis:
>>
>> Claiming that L1/L2/L3 caches and core
>> topology are "embedded cpu registers area,
>> it has nothing to do with the ram." That is
>> a brilliant piece of hardware fiction.
>>
>> Caches are SRAM acting as a hierarchical
>> staging ground for and backed by system
>> or device RAM, not an alternate dimension
>> divorced from memory entirely.
>>
>> And calling cache hierarchies "registers"
>> is like confusing a warehouse parking lot
>> with the glove compartment of a single car.
>> Never heard of the crossbar in AMD GPUs?
>>
>> LoL
>>
>> Bye
>>
>> Obelin Baisaroff schrieb:
>>  > Ross Finlayson wrote:
>>  >
>>  >> access to memory, has that these days with
>>  >> L1/L2/L3 caches and the proximity and affinity
>>  >> in the topology of the cores and
>>  >
>>  > those are embedded cpu registers area, it has
>>  > nothing to do with the ram. Works by higher
>>  > clocks compared, but merely useless in AI.
>>  >
>>
>>
>> Romelio Balakhonsky schrieb:
>>> Mild Shock wrote:
>>>
>>>> me about lazySet() in Java. But I went with set():
>>>>
>>>>       private static boolean enqueue(Queue q, Object data) {
>>>>           int pos = q.enqueuePos.get();
>>>>           for (; ; ) {
>>>>               int index = pos & q.bufferMask; int seq =
>>>>               q.sequences.get(index);
>>>>               int dif = seq - pos;
>>>>               if (dif == 0) {
>>>>                   if (q.enqueuePos.compareAndSet(pos, pos + 1)) {
>>>>                       q.data[index] = data; q.sequences.set(index, 
>>>> pos +
>>>>                       1);
>>>>                       return true;
>>>>                   }
>>>>                   pos = q.enqueuePos.get();
>>>>               } else if (dif < 0) {
>>>>                   return false;
>>>>               } else {
>>>>                   pos = q.enqueuePos.get();
>>>>               }
>>>>           }
>>>>       }
>>>>
>>>> The above version seems to be more suitable for my purpose, since it
>>>> allows polling, it basically implements offer(). While the
>>>
>>> completely nonsense. Not even correct grammatically
>>>
>>
> 

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

#671516 — Breaking the CUDA edge in AI by WebGPU (Re: Friendly Reminder: GPU 10x more performant than CPU)

Hi,

Concerning programmig means of GPUs,
"CUDA edge" refers to the massive competitive
advantage NVIDIA holds in artificial intelligence
due to its proprietary parallel computing platform,
CUDA (Compute Unified Device Architecture).

WebGPU and WGSL (WebGPU Shading Language)
represent a massive shift in AI because they
bring high-performance hardware acceleration
directly to the web browser without relying
on proprietary ecosystems like CUDA.

Developed by the W3C GPU for the Web
Community Group, this standard allows web
applications to interact directly with native
graphics APIs like Microsoft DirectX 12,
Apple Metal, and Vulkan

Bye

Some WebGPU show cases:

Frameworks like WebLLM leverage WebGPU to
run large language models locally inside
browsers like Google Chrome, providing
completely private, offline AI assistants.

Developers use libraries like Hugging Face
Transformers.js v3 to run computer vision,
speech recognition, and natural language
processing tasks directly on consumer
laptops and smartphones.

Mild Shock schrieb:
> Hi,
> 
> You guys are not paying attention. GPU is 10x
> more performant than CPU for certain integerish
> payload on the pi-WAM. 11.4 GLips on a GPU is
> 
> ca 10x more than 1.7 GLips on a CPU:
> 
> 11.4 Giga Lips with a Budget Laptop
> https://github.com/Jean-Luc-Picard-2021/gigabudget
> 
> Parallel π-WAM: 1.7 Giga Lips on a CPU
> https://medium.com/2989/8a984e75af44
> 
> So GPUs are by way not dead. Especially since
> they can be accessed via WebGPU / WGSL. But
> the line between GPU and CPU increasingly
> 
> blurrs. It already happens for APUs in
> that unified memory puts them into the same
> RAM space. But the programming means differ.
> 
> Also unified memory doesn't mean that
> the GPU sees memory the same way as a CPU.
> To get a taste how a GPU sees memory:
> 
> Memory Model
>      Execution Barriers
>      Fence and Address Spaces
>      Memory Model GFX6-GFX9
>      Memory Model GFX90A
>      Memory Model GFX942
>      Memory Model GFX10-GFX11
>      Memory Model GFX12
>      Memory Model GFX125x
> 
> https://llvm.org/docs/AMDGPUUsage.html
> 
> WebGPU / WGSL seems to have good support,
> since it can have Vulkan, Direct 12, or Metal
> as GPU low level interface. Maybe more
> 
> platforms, something NVIDIA I guess. But I
> don't know the details, how this is all done.
> 
> Bye
> 
> Mild Shock schrieb:
>> Hi,
>>
>> If you don't know how GPU caches work.
>> Just RTFM the RDNA 3.5 specs! They have
>> some explanations in the docs, what jobs
>>
>> the GPU cache lines do, in relation to
>> what instructions:
>>
>> Abbreviation for ‘Read The Fucking Manual’.
>> http://www.catb.org/esr/jargon/html/R/RTFM.html
>>
>> Here some RDNA 4.0 specs (smaller GPUs):
>>
>> "RDNA4" Instruction Set Architecture
>> Reference Guide - 7-April-2025
>> https://docs.amd.com/v/u/en-US/rdna4-instruction-set-architecture
>>
>> Here some CDNA 4.0 specs (bigger GPUs):
>>
>> CDNA4 Instruction Set Architecture
>> Reference Guide - 5-August-2025
>> https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/instruction-set-architectures/amd-instinct-cdna4-instruction-set-architecture.pdf 
>>
>>
>> Bye
>>
>> But the difference between CDNA GPUs and
>> RDNA GPUs seems not to be that big, at least
>> in 2025:
>>
>> Commitment to High-Performance
>> Computing in the Age of AI
>> https://docs.amd.com/v/u/en-US/precision-at-scale-amd-whitepaper
>>
>> While the Copilot+ threshold is 45 TFLOPs,
>> a MI355X, Datacenter, rack-mounted, liquid cooling,
>> has surely more memory, but only 157 TFLOPS.
>>
>> And RTX 5090, Desktop, 600W+ TDP, massive cooling,
>> has also only 100 TFLOPS. So it looks these
>> Copilot+ AI Laptops are pretty swell, arent they?
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>>  > completely nonsense. Not even correct grammatically
>>>
>>> Yes of course, Micro Penis will know. Nothing, LoL
>>> But here we find how village idiot Micro Penis is.
>>> The Cache Identity Crisis by Micro Penis:
>>>
>>> Claiming that L1/L2/L3 caches and core
>>> topology are "embedded cpu registers area,
>>> it has nothing to do with the ram." That is
>>> a brilliant piece of hardware fiction.
>>>
>>> Caches are SRAM acting as a hierarchical
>>> staging ground for and backed by system
>>> or device RAM, not an alternate dimension
>>> divorced from memory entirely.
>>>
>>> And calling cache hierarchies "registers"
>>> is like confusing a warehouse parking lot
>>> with the glove compartment of a single car.
>>> Never heard of the crossbar in AMD GPUs?
>>>
>>> LoL
>>>
>>> Bye
>>>
>>> Obelin Baisaroff schrieb:
>>>  > Ross Finlayson wrote:
>>>  >
>>>  >> access to memory, has that these days with
>>>  >> L1/L2/L3 caches and the proximity and affinity
>>>  >> in the topology of the cores and
>>>  >
>>>  > those are embedded cpu registers area, it has
>>>  > nothing to do with the ram. Works by higher
>>>  > clocks compared, but merely useless in AI.
>>>  >
>>>
>>>
>>> Romelio Balakhonsky schrieb:
>>>> Mild Shock wrote:
>>>>
>>>>> me about lazySet() in Java. But I went with set():
>>>>>
>>>>>       private static boolean enqueue(Queue q, Object data) {
>>>>>           int pos = q.enqueuePos.get();
>>>>>           for (; ; ) {
>>>>>               int index = pos & q.bufferMask; int seq =
>>>>>               q.sequences.get(index);
>>>>>               int dif = seq - pos;
>>>>>               if (dif == 0) {
>>>>>                   if (q.enqueuePos.compareAndSet(pos, pos + 1)) {
>>>>>                       q.data[index] = data; q.sequences.set(index, 
>>>>> pos +
>>>>>                       1);
>>>>>                       return true;
>>>>>                   }
>>>>>                   pos = q.enqueuePos.get();
>>>>>               } else if (dif < 0) {
>>>>>                   return false;
>>>>>               } else {
>>>>>                   pos = q.enqueuePos.get();
>>>>>               }
>>>>>           }
>>>>>       }
>>>>>
>>>>> The above version seems to be more suitable for my purpose, since it
>>>>> allows polling, it basically implements offer(). While the
>>>>
>>>> completely nonsense. Not even correct grammatically
>>>>
>>>
>>
> 

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

#671517 — Like WebAssembly before it, WebGPU has "escaped" the browser. (Re: Breaking the CUDA edge in AI by WebGPU (Re: Friendly Reminder: GPU 10x more performant than CPU)

Hi,

I didn't try yet escaping the browser with
WebGPU / WGSL. Only did inside the last 24 hour
another experiment, escaping Workers from the

browser. And found that I can easily use
Workers in node.js. But the story goes
that WebGPU / WGSL could be quite attractive,

across UX and non-UX, i.e. headless:

"Like WebAssembly before it, WebGPU has "escaped"
the browser. WebAssembly began as a browser
technology but quickly grew into a universal runtime
with standalone engines like Wasmtime and Wasmer.

WebGPU appears to follow the same path. With
bindings for Node.js, Deno, C++, and Rust
engines like Bevy [6], developers can already
run WebGPU workloads outside the browser.

This positions WebGPU not just as a graphics
API, but as a long-term portability layer for
GPU compute and rendering across ecosystems."

https://blog.4dpipeline.com/client-side-ai-is-here-how-webgpu-transforms-your-gpu-server-economics

Bye

Mild Shock schrieb:
> Hi,
> 
> Concerning programmig means of GPUs,
> "CUDA edge" refers to the massive competitive
> advantage NVIDIA holds in artificial intelligence
> due to its proprietary parallel computing platform,
> CUDA (Compute Unified Device Architecture).
> 
> WebGPU and WGSL (WebGPU Shading Language)
> represent a massive shift in AI because they
> bring high-performance hardware acceleration
> directly to the web browser without relying
> on proprietary ecosystems like CUDA.
> 
> Developed by the W3C GPU for the Web
> Community Group, this standard allows web
> applications to interact directly with native
> graphics APIs like Microsoft DirectX 12,
> Apple Metal, and Vulkan
> 
> Bye
> 
> Some WebGPU show cases:
> 
> Frameworks like WebLLM leverage WebGPU to
> run large language models locally inside
> browsers like Google Chrome, providing
> completely private, offline AI assistants.
> 
> Developers use libraries like Hugging Face
> Transformers.js v3 to run computer vision,
> speech recognition, and natural language
> processing tasks directly on consumer
> laptops and smartphones.
> 
> Mild Shock schrieb:
>> Hi,
>>
>> You guys are not paying attention. GPU is 10x
>> more performant than CPU for certain integerish
>> payload on the pi-WAM. 11.4 GLips on a GPU is
>>
>> ca 10x more than 1.7 GLips on a CPU:
>>
>> 11.4 Giga Lips with a Budget Laptop
>> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>
>> Parallel π-WAM: 1.7 Giga Lips on a CPU
>> https://medium.com/2989/8a984e75af44
>>
>> So GPUs are by way not dead. Especially since
>> they can be accessed via WebGPU / WGSL. But
>> the line between GPU and CPU increasingly
>>
>> blurrs. It already happens for APUs in
>> that unified memory puts them into the same
>> RAM space. But the programming means differ.
>>
>> Also unified memory doesn't mean that
>> the GPU sees memory the same way as a CPU.
>> To get a taste how a GPU sees memory:
>>
>> Memory Model
>>      Execution Barriers
>>      Fence and Address Spaces
>>      Memory Model GFX6-GFX9
>>      Memory Model GFX90A
>>      Memory Model GFX942
>>      Memory Model GFX10-GFX11
>>      Memory Model GFX12
>>      Memory Model GFX125x
>>
>> https://llvm.org/docs/AMDGPUUsage.html
>>
>> WebGPU / WGSL seems to have good support,
>> since it can have Vulkan, Direct 12, or Metal
>> as GPU low level interface. Maybe more
>>
>> platforms, something NVIDIA I guess. But I
>> don't know the details, how this is all done.
>>
>> Bye
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>> If you don't know how GPU caches work.
>>> Just RTFM the RDNA 3.5 specs! They have
>>> some explanations in the docs, what jobs
>>>
>>> the GPU cache lines do, in relation to
>>> what instructions:
>>>
>>> Abbreviation for ‘Read The Fucking Manual’.
>>> http://www.catb.org/esr/jargon/html/R/RTFM.html
>>>
>>> Here some RDNA 4.0 specs (smaller GPUs):
>>>
>>> "RDNA4" Instruction Set Architecture
>>> Reference Guide - 7-April-2025
>>> https://docs.amd.com/v/u/en-US/rdna4-instruction-set-architecture
>>>
>>> Here some CDNA 4.0 specs (bigger GPUs):
>>>
>>> CDNA4 Instruction Set Architecture
>>> Reference Guide - 5-August-2025
>>> https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/instruction-set-architectures/amd-instinct-cdna4-instruction-set-architecture.pdf 
>>>
>>>
>>> Bye
>>>
>>> But the difference between CDNA GPUs and
>>> RDNA GPUs seems not to be that big, at least
>>> in 2025:
>>>
>>> Commitment to High-Performance
>>> Computing in the Age of AI
>>> https://docs.amd.com/v/u/en-US/precision-at-scale-amd-whitepaper
>>>
>>> While the Copilot+ threshold is 45 TFLOPs,
>>> a MI355X, Datacenter, rack-mounted, liquid cooling,
>>> has surely more memory, but only 157 TFLOPS.
>>>
>>> And RTX 5090, Desktop, 600W+ TDP, massive cooling,
>>> has also only 100 TFLOPS. So it looks these
>>> Copilot+ AI Laptops are pretty swell, arent they?
>>>
>>> Mild Shock schrieb:
>>>> Hi,
>>>>
>>>>  > completely nonsense. Not even correct grammatically
>>>>
>>>> Yes of course, Micro Penis will know. Nothing, LoL
>>>> But here we find how village idiot Micro Penis is.
>>>> The Cache Identity Crisis by Micro Penis:
>>>>
>>>> Claiming that L1/L2/L3 caches and core
>>>> topology are "embedded cpu registers area,
>>>> it has nothing to do with the ram." That is
>>>> a brilliant piece of hardware fiction.
>>>>
>>>> Caches are SRAM acting as a hierarchical
>>>> staging ground for and backed by system
>>>> or device RAM, not an alternate dimension
>>>> divorced from memory entirely.
>>>>
>>>> And calling cache hierarchies "registers"
>>>> is like confusing a warehouse parking lot
>>>> with the glove compartment of a single car.
>>>> Never heard of the crossbar in AMD GPUs?
>>>>
>>>> LoL
>>>>
>>>> Bye
>>>>
>>>> Obelin Baisaroff schrieb:
>>>>  > Ross Finlayson wrote:
>>>>  >
>>>>  >> access to memory, has that these days with
>>>>  >> L1/L2/L3 caches and the proximity and affinity
>>>>  >> in the topology of the cores and
>>>>  >
>>>>  > those are embedded cpu registers area, it has
>>>>  > nothing to do with the ram. Works by higher
>>>>  > clocks compared, but merely useless in AI.
>>>>  >
>>>>
>>>>
>>>> Romelio Balakhonsky schrieb:
>>>>> Mild Shock wrote:
>>>>>
>>>>>> me about lazySet() in Java. But I went with set():
>>>>>>
>>>>>>       private static boolean enqueue(Queue q, Object data) {
>>>>>>           int pos = q.enqueuePos.get();
>>>>>>           for (; ; ) {
>>>>>>               int index = pos & q.bufferMask; int seq =
>>>>>>               q.sequences.get(index);
>>>>>>               int dif = seq - pos;
>>>>>>               if (dif == 0) {
>>>>>>                   if (q.enqueuePos.compareAndSet(pos, pos + 1)) {
>>>>>>                       q.data[index] = data; q.sequences.set(index, 
>>>>>> pos +
>>>>>>                       1);
>>>>>>                       return true;
>>>>>>                   }
>>>>>>                   pos = q.enqueuePos.get();
>>>>>>               } else if (dif < 0) {
>>>>>>                   return false;
>>>>>>               } else {
>>>>>>                   pos = q.enqueuePos.get();
>>>>>>               }
>>>>>>           }
>>>>>>       }
>>>>>>
>>>>>> The above version seems to be more suitable for my purpose, since it
>>>>>> allows polling, it basically implements offer(). While the
>>>>>
>>>>> completely nonsense. Not even correct grammatically
>>>>>
>>>>
>>>
>>
> 

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

#671533 — Re: Breaking the CUDA edge in AI by WebGPU (Re: Friendly Reminder: GPU 10x more performant than CPU)

Mild Shock wrote:

> Frameworks like WebLLM leverage WebGPU to run large language models
> locally inside browsers like Google Chrome, providing completely
> private, offline AI assistants.

since when google chrome private, think again

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

#671536 — http://localhost:567921/ is a private REST endpoint [Teaching Micro Penis Vilage Idiot] (Was: Breaking the CUDA edge in AI by WebGPU)

Hi,

------------------- begin --------------------
Teaching Micro Penis Vilage Idiot
------------------- begin --------------------

You dont have to use WebLLM, respectively
WebGPU / WGSL literally, just read the next
post I did AND use your brains moron:

Like WebAssembly before it, WebGPU has "escaped" the browser
https://blog.4dpipeline.com/client-side-ai-is-here-how-webgpu-transforms-your-gpu-server-economics

http://localhost:567921/ is private you moron.,
or what ever port REST is using. You typically access
an offline AI assistant, via some REST end-point

on your machine. Nothing to do with Google Chrome
browser security. You can make it as private as you want, by
having a firewall and not outward or inward

connection at all, only your REST end-point
on your machine. Or if you want a REST end-point
on a server of yours in the same intranet.

You don't need to use the internet, or put
something on the extranet, or use some sort of
subscription. What you need is access through

the firewall to download the REST software
and the LLM model. Tools like LM Studio and oMLX
offer this download and also install REST endpoint.

------------------- end --------------------
Teaching Micro Penis Vilage Idiot}
------------------- end --------------------

Bye

Will Bakshandaev schrieb:
> Mild Shock wrote:
> 
>> Frameworks like WebLLM leverage WebGPU to run large language models
>> locally inside browsers like Google Chrome, providing completely
>> private, offline AI assistants.
> 
> since when google chrome private, think again
> 

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

#671538 — If you are paranoid you can use Falco [Agentic AI] (Re: http://localhost:567921/ is a private REST endpoint)

Hi,

Currently companies such as Apple, Windows, etc..
are hardning their operating systems, so
that they can provide agentic AI sandboxes.

Problem is an agentic AI, that acts on your
behalf, when not enough supervised, might
do all kind of stuff on its own. So how do you

have harder borders. Besides companies that
write operating systems, there is also a cottage
industry now that adresses this paranoia,

here an example from a former Prologer:

Stop guessing what your coding agent just did
Prempti: Guardrails and Observability for AI Coding Agents.
https://prempti.falco.org/

IntelliJ doesn't have this problem, it shows a
not yet hyper locally commited change, in the editor,
created by the AI, that you can review, and

then hyper locally commit in the editor. Only then
it lands in the file system. But also there it
will be subject to the local history and repository

version system. So the IntelliJ AI is pretty smartly
implemented, and hooks into their editors and newly
introduced hyper change visualization, a feature that

probably codemirror doesn't have yet. Have to double check.

Have Fun!

Bye

Mild Shock schrieb:
> Hi,
> 
> ------------------- begin --------------------
> Teaching Micro Penis Vilage Idiot
> ------------------- begin --------------------
> 
> You dont have to use WebLLM, respectively
> WebGPU / WGSL literally, just read the next
> post I did AND use your brains moron:
> 
> Like WebAssembly before it, WebGPU has "escaped" the browser
> https://blog.4dpipeline.com/client-side-ai-is-here-how-webgpu-transforms-your-gpu-server-economics 
> 
> 
> http://localhost:567921/ is private you moron.,
> or what ever port REST is using. You typically access
> an offline AI assistant, via some REST end-point
> 
> on your machine. Nothing to do with Google Chrome
> browser security. You can make it as private as you want, by
> having a firewall and not outward or inward
> 
> connection at all, only your REST end-point
> on your machine. Or if you want a REST end-point
> on a server of yours in the same intranet.
> 
> You don't need to use the internet, or put
> something on the extranet, or use some sort of
> subscription. What you need is access through
> 
> the firewall to download the REST software
> and the LLM model. Tools like LM Studio and oMLX
> offer this download and also install REST endpoint.
> 
> ------------------- end --------------------
> Teaching Micro Penis Vilage Idiot}
> ------------------- end --------------------
> 
> Bye
> 
> Will Bakshandaev schrieb:
>> Mild Shock wrote:
>>
>>> Frameworks like WebLLM leverage WebGPU to run large language models
>>> locally inside browsers like Google Chrome, providing completely
>>> private, offline AI assistants.
>>
>> since when google chrome private, think again
>>
> 

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

#671539 — What would an EMACs guru say [Windows Recall] (Was: If you are paranoid you can use Falco [Agentic AI])

Hi,

What would an EMACs guru say. Can
EMACs process the equivalent of these
HTML tags:

The <ins> HTML element represents a
range of text that has been added to a document.
https://developer.mozilla.org/en-US/docs/Web/HTML/Reference/Elements/ins

The <del> HTML element represents a range
of text that has been deleted from a document.
https://developer.mozilla.org/en-US/docs/Web/HTML/Reference/Elements/del

This can be used when rendering "track changes"
or source code diff information, for example.
Office word can do it as well, track changes.

In principle somehow, every editor that has undo
and redo, dunno, can Excel show changes? What was
Windows Recall again on an AI Laptop?

> Windows Recall takes a screenshot of a user's 
> desktop every few seconds, then uses on-device 
> large language models to allow a user to 
> retrieve items and information that had 
> previously been on their screen. 
https://en.wikipedia.org/wiki/Windows_Recall

A little bit unstructured, compared to the
HTML tags and quite neurotic approach.

Bye

Mild Shock schrieb:
> Hi,
> 
> Currently companies such as Apple, Windows, etc..
> are hardning their operating systems, so
> that they can provide agentic AI sandboxes.
> 
> Problem is an agentic AI, that acts on your
> behalf, when not enough supervised, might
> do all kind of stuff on its own. So how do you
> 
> have harder borders. Besides companies that
> write operating systems, there is also a cottage
> industry now that adresses this paranoia,
> 
> here an example from a former Prologer:
> 
> Stop guessing what your coding agent just did
> Prempti: Guardrails and Observability for AI Coding Agents.
> https://prempti.falco.org/
> 
> IntelliJ doesn't have this problem, it shows a
> not yet hyper locally commited change, in the editor,
> created by the AI, that you can review, and
> 
> then hyper locally commit in the editor. Only then
> it lands in the file system. But also there it
> will be subject to the local history and repository
> 
> version system. So the IntelliJ AI is pretty smartly
> implemented, and hooks into their editors and newly
> introduced hyper change visualization, a feature that
> 
> probably codemirror doesn't have yet. Have to double check.
> 
> Have Fun!
> 
> Bye
> 
> Mild Shock schrieb:
>> Hi,
>>
>> ------------------- begin --------------------
>> Teaching Micro Penis Vilage Idiot
>> ------------------- begin --------------------
>>
>> You dont have to use WebLLM, respectively
>> WebGPU / WGSL literally, just read the next
>> post I did AND use your brains moron:
>>
>> Like WebAssembly before it, WebGPU has "escaped" the browser
>> https://blog.4dpipeline.com/client-side-ai-is-here-how-webgpu-transforms-your-gpu-server-economics 
>>
>>
>> http://localhost:567921/ is private you moron.,
>> or what ever port REST is using. You typically access
>> an offline AI assistant, via some REST end-point
>>
>> on your machine. Nothing to do with Google Chrome
>> browser security. You can make it as private as you want, by
>> having a firewall and not outward or inward
>>
>> connection at all, only your REST end-point
>> on your machine. Or if you want a REST end-point
>> on a server of yours in the same intranet.
>>
>> You don't need to use the internet, or put
>> something on the extranet, or use some sort of
>> subscription. What you need is access through
>>
>> the firewall to download the REST software
>> and the LLM model. Tools like LM Studio and oMLX
>> offer this download and also install REST endpoint.
>>
>> ------------------- end --------------------
>> Teaching Micro Penis Vilage Idiot}
>> ------------------- end --------------------
>>
>> Bye
>>
>> Will Bakshandaev schrieb:
>>> Mild Shock wrote:
>>>
>>>> Frameworks like WebLLM leverage WebGPU to run large language models
>>>> locally inside browsers like Google Chrome, providing completely
>>>> private, offline AI assistants.
>>>
>>> since when google chrome private, think again
>>>
>>
> 

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

#671540 — Re: http://localhost:567921/ is a private REST endpoint [Teaching Micro Penis Vilage Idiot] (Was: Breaking the CUDA edge in AI by WebGPU)

Mild Shock wrote:

> http://localhost:567921/ is private you moron.,
> or what ever port REST is using. You typically access an offline AI
> assistant, via some REST end-point

yet one more proof this half german inbreed is an imbecile, ports go up to 
16bits/64k only, idiot, you cant have a localhost: whatever wrong number 
you put there. You extreme fucking idiot. 

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

#671541 — Re: http://localhost:567921/ is a private REST endpoint [Teaching Micro Penis Vilage Idiot] (Was: Breaking the CUDA edge in AI by WebGPU)

Mild Shock wrote:

> on your machine. Nothing to do with Google Chrome browser security. You
> can make it as private as you want, by having a firewall and not outward
> or inward

yes, i can see your point, they just want your private cellphone number, 
there rest is private and free, idiot

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

#671542 — Decide what you critique tiny winy penis (Was: http://localhost:567921/ is a private REST endpoint)

Hi,

Maybe you could post some subtantial critique moron?
Instead of gibberish all the time. What does a cellphone
number have to do with a REST endpoint? Nothing!

Its all locally and my laptop has no cellphone number:

Start the REST API server

To start the server, run the following command:
lms server start

Endpoints

GET /api/v0/models
List all loaded and downloaded models
Example request
curl -H "Authorization: Bearer $LM_API_TOKEN" 
http://localhost:1234/api/v0/models
Response format
{
   "object": "list",
   "data": [
     {
       "id": "qwen2-vl-7b-instruct",
       "object": "model",
       "type": "vlm",
       "publisher": "mlx-community",
       "arch": "qwen2_vl"
Etc...
https://lmstudio.ai/docs/developer/rest/endpoints

Bye

BTW: LM Studio recently introduced LM Link,
which provides some VPN. It can be used to
create clients or servers that run models.

It is end-to-end encrypted, and built on top
of custom Tailscale mesh VPNs. This is for
the paranoid, that want to acccess a

LLM from one end of the globe, that sits
on the other end of the globe, and have
no evesdroper or whatever on the

information that is exchanged.

> Mild Shock wrote:
> 
>> http://localhost:567921/ is private you moron.,
>> or what ever port REST is using. You typically 
>> access an offline AI assistant, via some REST end-point
> 
> yet one more proof this half german inbreed 
> is an imbecile, ports go up to 16bits/64k only, 
> idiot, you cant have a localhost: whatever wrong 
> number you put there. You extreme fucking idiot. 

>> on your machine. Nothing to do with Google 
>> Chrome browser security. You can make it as private 
>> as you want, by having a firewall and not outward
>> or inward
> 
> yes, i can see your point, they just want 
> your private cellphone number, 
> there rest is private and free, idiot



Mild Shock schrieb:
> Hi,
> 
> ------------------- begin --------------------
> Teaching Micro Penis Vilage Idiot
> ------------------- begin --------------------
> 
> You dont have to use WebLLM, respectively
> WebGPU / WGSL literally, just read the next
> post I did AND use your brains moron:
> 
> Like WebAssembly before it, WebGPU has "escaped" the browser
> https://blog.4dpipeline.com/client-side-ai-is-here-how-webgpu-transforms-your-gpu-server-economics 
> 
> 
> http://localhost:567921/ is private you moron.,
> or what ever port REST is using. You typically access
> an offline AI assistant, via some REST end-point
> 
> on your machine. Nothing to do with Google Chrome
> browser security. You can make it as private as you want, by
> having a firewall and not outward or inward
> 
> connection at all, only your REST end-point
> on your machine. Or if you want a REST end-point
> on a server of yours in the same intranet.
> 
> You don't need to use the internet, or put
> something on the extranet, or use some sort of
> subscription. What you need is access through
> 
> the firewall to download the REST software
> and the LLM model. Tools like LM Studio and oMLX
> offer this download and also install REST endpoint.
> 
> ------------------- end --------------------
> Teaching Micro Penis Vilage Idiot}
> ------------------- end --------------------
> 
> Bye
> 
> Will Bakshandaev schrieb:
>> Mild Shock wrote:
>>
>>> Frameworks like WebLLM leverage WebGPU to run large language models
>>> locally inside browsers like Google Chrome, providing completely
>>> private, offline AI assistants.
>>
>> since when google chrome private, think again
>>
> 

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

#671544 — How confused is tiny winy penis? (Re: Decide what you critique tiny winy penis)

Hi,

How confused is tiny winy penis?
For the 100th time the budget here:

11.4 Giga Lips with a Budget Laptop
Ryzen AI 7 350 w/ Radeon 860M
https://github.com/Jean-Luc-Picard-2021/gigabudget

is a laptop and not a smartphone. It
has no cellphone number. And w/ means
integrated GPU on the silicon chip,

and not a GPU connected to the mainboard
via some PCI bus. The model is a acer
swift go, I already posted this info:

Swift Go 16 AI SFG16-61-R21J Notebook
https://www.acer.com/ch-de/laptops/swift/swift-go-16-ai-amd/pdp/NX.JCREZ.007

Bye

Mild Shock schrieb:
> Hi,
> 
> Maybe you could post some subtantial critique moron?
> Instead of gibberish all the time. What does a cellphone
> number have to do with a REST endpoint? Nothing!
> 
> Its all locally and my laptop has no cellphone number:
> 
> Start the REST API server
> 
> To start the server, run the following command:
> lms server start
> 
> Endpoints
> 
> GET /api/v0/models
> List all loaded and downloaded models
> Example request
> curl -H "Authorization: Bearer $LM_API_TOKEN" 
> http://localhost:1234/api/v0/models
> Response format
> {
>    "object": "list",
>    "data": [
>      {
>        "id": "qwen2-vl-7b-instruct",
>        "object": "model",
>        "type": "vlm",
>        "publisher": "mlx-community",
>        "arch": "qwen2_vl"
> Etc...
> https://lmstudio.ai/docs/developer/rest/endpoints
> 
> Bye
> 
> BTW: LM Studio recently introduced LM Link,
> which provides some VPN. It can be used to
> create clients or servers that run models.
> 
> It is end-to-end encrypted, and built on top
> of custom Tailscale mesh VPNs. This is for
> the paranoid, that want to acccess a
> 
> LLM from one end of the globe, that sits
> on the other end of the globe, and have
> no evesdroper or whatever on the
> 
> information that is exchanged.
> 
>> Mild Shock wrote:
>>
>>> http://localhost:567921/ is private you moron.,
>>> or what ever port REST is using. You typically access an offline AI 
>>> assistant, via some REST end-point
>>
>> yet one more proof this half german inbreed is an imbecile, ports go 
>> up to 16bits/64k only, idiot, you cant have a localhost: whatever 
>> wrong number you put there. You extreme fucking idiot. 
> 
>>> on your machine. Nothing to do with Google Chrome browser security. 
>>> You can make it as private as you want, by having a firewall and not 
>>> outward
>>> or inward
>>
>> yes, i can see your point, they just want your private cellphone 
>> number, there rest is private and free, idiot
> 
> 
> 
> Mild Shock schrieb:
>> Hi,
>>
>> ------------------- begin --------------------
>> Teaching Micro Penis Vilage Idiot
>> ------------------- begin --------------------
>>
>> You dont have to use WebLLM, respectively
>> WebGPU / WGSL literally, just read the next
>> post I did AND use your brains moron:
>>
>> Like WebAssembly before it, WebGPU has "escaped" the browser
>> https://blog.4dpipeline.com/client-side-ai-is-here-how-webgpu-transforms-your-gpu-server-economics 
>>
>>
>> http://localhost:567921/ is private you moron.,
>> or what ever port REST is using. You typically access
>> an offline AI assistant, via some REST end-point
>>
>> on your machine. Nothing to do with Google Chrome
>> browser security. You can make it as private as you want, by
>> having a firewall and not outward or inward
>>
>> connection at all, only your REST end-point
>> on your machine. Or if you want a REST end-point
>> on a server of yours in the same intranet.
>>
>> You don't need to use the internet, or put
>> something on the extranet, or use some sort of
>> subscription. What you need is access through
>>
>> the firewall to download the REST software
>> and the LLM model. Tools like LM Studio and oMLX
>> offer this download and also install REST endpoint.
>>
>> ------------------- end --------------------
>> Teaching Micro Penis Vilage Idiot}
>> ------------------- end --------------------
>>
>> Bye
>>
>> Will Bakshandaev schrieb:
>>> Mild Shock wrote:
>>>
>>>> Frameworks like WebLLM leverage WebGPU to run large language models
>>>> locally inside browsers like Google Chrome, providing completely
>>>> private, offline AI assistants.
>>>
>>> since when google chrome private, think again
>>>
>>
> 

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

Page 1 of 4  [1] 2 3 4  Next page →

Back to top | Article view | sci.physics.relativity

csiph-web