500 CHF. So we went already form AI Laptop
Hi,
Looking at the floor plan of a NPU:
Getting peak TOPS on a Ryzen AI 7 350 NPU https://destevez.net/2026/05/getting-peak-tops-on-a-ryzen-ai-7-350-npu/
It seems to me comms between tiles takes
at least Manhattan Distance or L1 Norm time,
if there is no comms congestion
But how does a packet travel? This way:
+----E
|
|
S
Or this way, from start S to end E:
˙˙ +-E
˙ +
˙+
S
And what does the chip do if there is
traffic congestion? Some papers are
here, possibly an old problem giving
that processor "cubes" are nothing new.
But a "cube" would be 3D and not 2D.
This paper is old from 2007 or so:
Routing Algorithms for 2D NoC Architectures http://cva.stanford.edu/classes/ee382c/research/2DRouting.pdf
Bye
Mild Shock schrieb:
Hi,
Tablets and phone are more annoying to
use with WebGPU. The usual browsers don't
have a Chrome DevTools panel integrated,
so that one could do JavaScript Debugging
directly on the device. Instead one has to
use a desktop machine, and connect the
device via UBS-C , and start a Chrome
Browser there . And then start a Chrome
DevTools panel alone, that is pair with
the device, via UBS-C cable. So this way
I already see where it crashes on the
tablets and phone:
await output.mapAsync(GPUMapMode.READ)
Unhandled Promise Rejection: OperationError
The above is the error that one can re-produce
already here with this test:
11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget
Not sure what exactly happens. Maybe
a form of timeout or device lost, that the
primitive HTML / JavaScript doesn't handle
gracefully yet. Maybe redimensioning the
test, so that it consumes less time would
help. Who knows? Will see. For production
use of a GPU integration I have to anyway
provide work slicing it seems.
Bye
Mild Shock schrieb:
Hi,
He uses FIFO, and DMA and Noc:
Getting peak TOPS on a Ryzen AI 7 350 NPU
https://destevez.net/2026/05/getting-peak-tops-on-a-ryzen-ai-7-350-npu/
But lets say whether its FIFO or FILO
isn't so important his used cases are,
what is now found in my library(furryhaze)
for GPU, namely the very basic:
/**
˙˙* test_gpu_comp_start(W, K): internal only
˙˙* The predicate succeeds. As a side effect it
˙˙* starts the ă-WAM W with K warps.
˙˙*/
function test_gpu_comp_start(args)
/**
˙˙* test_gpu_comp_join(W, P): internal only
˙˙* The predicate succeeds in P with a new promise
˙˙* that waits for the ă-WAM W to finish.
˙˙*/
function test_gpu_comp_join(args)
A GPU interface, via the command processor
for example of WebGPU, does the above
synchronization for you.
In the NPU example he does everything
low level, with Python IRON an stuff:
"Since the main way to achieve synchronization
within the IRON framework is by doing data
movement with object FIFOs, I?m sending a
dummy uint32 value as some sort of
synchronization token.
Waiting for all the kernels to finish is
trickier. The object FIFOs support a join
pattern in which an object FIFO consumes an
object from each of multiple object FIFOs,
concatenates these objects and produces the
concatenated object as a result.
Etc.."
Getting peak TOPS on a Ryzen AI 7 350 NPU
https://destevez.net/2026/05/getting-peak-tops-on-a-ryzen-ai-7-350-npu/
So Daniel Est‚vez Scientific & Technical
Amateur Radio, gives a nice glimpse into an
NPU, I have not yet publicitly released
my library(furryhaze), since its still in
testing. Maybe take another week or so,
still I have ironed out all corners,
for example the new gpu_comp_start and
gpu_comp_join works fine on may desktop
AI laptops, but I have still a bug on
my iPad AI tablet, on the Redmi AI phone,
also chokes on a test case.
Bye
Mild Shock schrieb:
Hi,
This was archived on Jul 9, 2026:
11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget
Still, Jul 29, Rossy Boy halucinates accusations:
Ross Finlayson schrieb:
.. bla bla goto bla bla ..
Stupid gangster:˙ teamsters are a union.
In the trades, not the steals, ....
Woa! Thats now 20 days of brain desease,
and not understanding the meaning and implications.
Even not understand pi-WAM has Hack VM backend.
But its all opensource. Bravo Rossy Boy, you are
champion in brainlessness and lazyness of
a idiot usenet troll.
Bye
Johann 'Myrkraverk' Oskarsson schrieb:
On 28/07/2026 2:43 AM, Ross Finlayson wrote:
Hello, here I'll post some design notes and a panel discussion
with some
chat-bots about making some sense of the "vector-wide scalar word" >>>>>> and "character machines", on commodity hardware about ubiquitous
operations.
It's considered at least tangentially relevant to comp.lang.c and
comp.lang.c++ because for example text is ubiquitous and the targets >>>>>> would be low-level, while the higher-level languages would have a
same sort of patternry, and for example that libc and cstdlib are
standard, and as with regards to POSIX and Unicode and so on.
Please feel free to excuse or ignore, or comment as freely.
Thanks for reading.
Are you generating all of your code via LLMs?˙ Rest assured,
the LLM generated code will have subtle and sometimes not so subtle
bugs.
Happy bughunting!
Hi,
Ok, that was quick. While my AI Laptops were
around > 1000 CHF. The Mac Neo was around
500 CHF. So we went already form AI Laptop
to AI Notebook in a few months. Now Aduino
is playing pionier, having a CPU with a NPU
on board, some Qualcomm thingy.
New Arduino Ventuno Q: Better than Raspberry Pi? https://www.youtube.com/watch?v=qQS_xWsU00I
The specs:
Octa-core ARM Cortex CPU
Adreno GPU and Hexagon AI processor (up to 40 TOPS)
16GB LPDDR5 RAM and 64GB eMMC storage
M.2 slot for NVMe SSD expansion
The cost is around 300 CHF.
Bye
Hi,
I miss an ISOMICRO profile of Web Prolog, a profile
that can run on small embedded devices, and only
single threaded. Like Python can do for example.
I deleted my previous post, since it drifted into
high performance computing. It was a reaction of
mine, to these results and how they were viewed.
Web Prolog result:
100,000 4.749 s
I get this here:
/* 100'000 iterations */
?- between(1,3,_), time(ping_pong), fail; true.
% Zeit 97.020 ms, Benutzer 2 %, Lips 735 k
% Zeit 95.040 ms, Benutzer 1 %, Lips 1057 k
% Zeit 98.740 ms, Benutzer 1 %, Lips 974 k
true.
But the results have a few drawbacks. They use a highly
specialized ă-WAM Prolog subset and a highly specialized
Hack VM backend. Also the ping pong code was optimized.
So I guess this high performance view is too specifiec
for the actor model. So to get a more general comparison,
I tried something else. I used a Python implemented Prolog
and a Python asyncio.Future implemented one element
channels, the later equals SWI-Prolog queues with max_size=1.
Finally I used the classical ping pong. Now with PyPy as the Python
runtime the results are, 6x times faster than the shared database
on a SWI-Prolog server provided by Torbj”rn Lager. Difficult
to judge maybe my machine is just 6x times faster? One could
install PyPy, download Dogelog Player and run it on the server:
?- between(1,3,_), time(ping_pong(100000)), fail; true.
% Time 812.000 ms, User 54 %, Lips 5977 k
% Time 703.000 ms, User 53 %, Lips 6915 k
% Time 766.000 ms, User 64 %, Lips 5339 k
true.
But this makes me ask, where would one see using for
example SWI-Prolog Engines for the actor model, so that it
becomes competitive to asyncio.Future? Any idea how to do it?
I guess asyncio.Future only uses a micro queue or something.
This would give the ISOMICRO profile of Web Prolog, a profile
that can run on small embedded devices single threaded.
The opposite of high preformance computing (HPC).
Bye
See also:
https://trinity.elfenbenstornet.se/
P.S.: Here the source code, first what was used for validation:
˙classic ping pong with channels and with logging
And the validation output:
˙log of running N=3
And what was used for benchmarking:
˙classic ping pong with channels and without logging
Mild Shock schrieb:
Hi,
Ok, that was quick. While my AI Laptops were
around > 1000 CHF. The Mac Neo was around
500 CHF. So we went already form AI Laptop
to AI Notebook in a few months. Now Aduino
is playing pionier, having a CPU with a NPU
on board, some Qualcomm thingy.
New Arduino Ventuno Q: Better than Raspberry Pi?
https://www.youtube.com/watch?v=qQS_xWsU00I
The specs:
Octa-core ARM Cortex CPU
Adreno GPU and Hexagon AI processor (up to 40 TOPS)
16GB LPDDR5 RAM and 64GB eMMC storage
M.2 slot for NVMe SSD expansion
The cost is around 300 CHF.
Bye
Hi,
I really wonder what use cases Web Prolog trinity would
have. Would Web Prolog trinity reach in its scope into
the domain of AI chat bots inside a web page? Ok, google
has hijacked the term ?declarative?, when an AI chatbot
assistant helps fill out a HTML form. And it uses the term
?imperative? when the AI chatbot calls JavaScript routines.
Could be related to the ACTOR model who knows. Although
the hijacking is not optimal, I like the thinking in levels, that are
related to states, just like Torbj”rn Lager exercises in his 400
pages, although google does maybe a bender, when it calls
a form submit, and hence a HTTP POST, declarative.
Bye
See also:
WebMCP
https://developer.chrome.com/docs/ai/webmcp
Here the newly arrived browser integration:
Le Petit Bistro https://googlechromelabs.github.io/webmcp-tools/demos/french-bistro
image
Mild Shock schrieb:
Hi,
I miss an ISOMICRO profile of Web Prolog, a profile
that can run on small embedded devices, and only
single threaded. Like Python can do for example.
I deleted my previous post, since it drifted into
high performance computing. It was a reaction of
mine, to these results and how they were viewed.
Web Prolog result:
100,000 4.749 s
I get this here:
/* 100'000 iterations */
?- between(1,3,_), time(ping_pong), fail; true.
% Zeit 97.020 ms, Benutzer 2 %, Lips 735 k
% Zeit 95.040 ms, Benutzer 1 %, Lips 1057 k
% Zeit 98.740 ms, Benutzer 1 %, Lips 974 k
true.
But the results have a few drawbacks. They use a highly
specialized ă-WAM Prolog subset and a highly specialized
Hack VM backend. Also the ping pong code was optimized.
So I guess this high performance view is too specifiec
for the actor model. So to get a more general comparison,
I tried something else. I used a Python implemented Prolog
and a Python asyncio.Future implemented one element
channels, the later equals SWI-Prolog queues with max_size=1.
Finally I used the classical ping pong. Now with PyPy as the Python
runtime the results are, 6x times faster than the shared database
on a SWI-Prolog server provided by Torbj”rn Lager. Difficult
to judge maybe my machine is just 6x times faster? One could
install PyPy, download Dogelog Player and run it on the server:
?- between(1,3,_), time(ping_pong(100000)), fail; true.
% Time 812.000 ms, User 54 %, Lips 5977 k
% Time 703.000 ms, User 53 %, Lips 6915 k
% Time 766.000 ms, User 64 %, Lips 5339 k
true.
But this makes me ask, where would one see using for
example SWI-Prolog Engines for the actor model, so that it
becomes competitive to asyncio.Future? Any idea how to do it?
I guess asyncio.Future only uses a micro queue or something.
This would give the ISOMICRO profile of Web Prolog, a profile
that can run on small embedded devices single threaded.
The opposite of high preformance computing (HPC).
Bye
See also:
https://trinity.elfenbenstornet.se/
P.S.: Here the source code, first what was used for validation:
˙˙classic ping pong with channels and with logging
And the validation output:
˙˙log of running N=3
And what was used for benchmarking:
˙˙classic ping pong with channels and without logging
Mild Shock schrieb:
Hi,
Ok, that was quick. While my AI Laptops were
around > 1000 CHF. The Mac Neo was around
500 CHF. So we went already form AI Laptop
to AI Notebook in a few months. Now Aduino
is playing pionier, having a CPU with a NPU
on board, some Qualcomm thingy.
New Arduino Ventuno Q: Better than Raspberry Pi?
https://www.youtube.com/watch?v=qQS_xWsU00I
The specs:
Octa-core ARM Cortex CPU
Adreno GPU and Hexagon AI processor (up to 40 TOPS)
16GB LPDDR5 RAM and 64GB eMMC storage
M.2 slot for NVMe SSD expansion
The cost is around 300 CHF.
Bye
Hi,
What if a Prolog system can draw enough processing power,
by tapping into the GPU of budget laptop that shows the
web page? And run locally in a ServiceWorker. The WebMCP
use case Le Petit Bistro is already such that it also has a
ServiceWorker for the Gemini stub. But the envisioned variant of a
ISOMICRO model would neither need a Gemini API token,
nor would it need WebSockets or HTTP for communication. It
would all be WebMCP inside the conglomerate of website and
workers. The small language model is loaded into the ServiceWorker
and run on the GPU. Here my contribution to this idea:
11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget
Work in progress cannot demonstrate a Prolog assistant yet.
Bye
Mild Shock schrieb:
Hi,
I really wonder what use cases Web Prolog trinity would
have. Would Web Prolog trinity reach in its scope into
the domain of AI chat bots inside a web page? Ok, google
has hijacked the term ?declarative?, when an AI chatbot
assistant helps fill out a HTML form. And it uses the term
?imperative? when the AI chatbot calls JavaScript routines.
Could be related to the ACTOR model who knows. Although
the hijacking is not optimal, I like the thinking in levels, that are
related to states, just like Torbj”rn Lager exercises in his 400
pages, although google does maybe a bender, when it calls
a form submit, and hence a HTTP POST, declarative.
Bye
See also:
WebMCP
https://developer.chrome.com/docs/ai/webmcp
Here the newly arrived browser integration:
Le Petit Bistro
https://googlechromelabs.github.io/webmcp-tools/demos/french-bistro
image
Mild Shock schrieb:
Hi,
I miss an ISOMICRO profile of Web Prolog, a profile
that can run on small embedded devices, and only
single threaded. Like Python can do for example.
I deleted my previous post, since it drifted into
high performance computing. It was a reaction of
mine, to these results and how they were viewed.
Web Prolog result:
100,000 4.749 s
I get this here:
/* 100'000 iterations */
?- between(1,3,_), time(ping_pong), fail; true.
% Zeit 97.020 ms, Benutzer 2 %, Lips 735 k
% Zeit 95.040 ms, Benutzer 1 %, Lips 1057 k
% Zeit 98.740 ms, Benutzer 1 %, Lips 974 k
true.
But the results have a few drawbacks. They use a highly
specialized ă-WAM Prolog subset and a highly specialized
Hack VM backend. Also the ping pong code was optimized.
So I guess this high performance view is too specifiec
for the actor model. So to get a more general comparison,
I tried something else. I used a Python implemented Prolog
and a Python asyncio.Future implemented one element
channels, the later equals SWI-Prolog queues with max_size=1.
Finally I used the classical ping pong. Now with PyPy as the Python
runtime the results are, 6x times faster than the shared database
on a SWI-Prolog server provided by Torbj”rn Lager. Difficult
to judge maybe my machine is just 6x times faster? One could
install PyPy, download Dogelog Player and run it on the server:
?- between(1,3,_), time(ping_pong(100000)), fail; true.
% Time 812.000 ms, User 54 %, Lips 5977 k
% Time 703.000 ms, User 53 %, Lips 6915 k
% Time 766.000 ms, User 64 %, Lips 5339 k
true.
But this makes me ask, where would one see using for
example SWI-Prolog Engines for the actor model, so that it
becomes competitive to asyncio.Future? Any idea how to do it?
I guess asyncio.Future only uses a micro queue or something.
This would give the ISOMICRO profile of Web Prolog, a profile
that can run on small embedded devices single threaded.
The opposite of high preformance computing (HPC).
Bye
See also:
https://trinity.elfenbenstornet.se/
P.S.: Here the source code, first what was used for validation:
˙˙classic ping pong with channels and with logging
And the validation output:
˙˙log of running N=3
And what was used for benchmarking:
˙˙classic ping pong with channels and without logging
Mild Shock schrieb:
Hi,
Ok, that was quick. While my AI Laptops were
around > 1000 CHF. The Mac Neo was around
500 CHF. So we went already form AI Laptop
to AI Notebook in a few months. Now Aduino
is playing pionier, having a CPU with a NPU
on board, some Qualcomm thingy.
New Arduino Ventuno Q: Better than Raspberry Pi?
https://www.youtube.com/watch?v=qQS_xWsU00I
The specs:
Octa-core ARM Cortex CPU
Adreno GPU and Hexagon AI processor (up to 40 TOPS)
16GB LPDDR5 RAM and 64GB eMMC storage
M.2 slot for NVMe SSD expansion
The cost is around 300 CHF.
Bye
| Sysop: | Tetrazocine |
|---|---|
| Location: | Melbourne, VIC, Australia |
| Users: | 9 |
| Nodes: | 8 (0 / 8) |
| Uptime: | 245:21:00 |
| Calls: | 220 |
| Files: | 21,513 |
| Messages: | 83,782 |