From Newsgroup: comp.lang.prolog

Hi,

Because GenX and later suffers from:

Somehow methods and tools to realize
efficient DCGs in Prolog are missing. Most
DCG attempts that one sees succumb

to some declarative nonsense, creating
exponentially many spurious choice points,
you find rarely somebody mastering the Art.

Modern programmers fancy nothing else than
throwing a set of foreign library to their
Prolog system project. This is best seen in WebPL:

LALRPOP MIT/Apache-2.0 Generate the parser https://github.com/w-henderson/WebPL/blob/main/dissertation.pdf

So there is no aim at creating a self hosting
Prolog system. There is a deep distrust in
DCGs. But why build a Prolog system that will

possibly ultimately have DCG, when you distrust
in DCGs? The second problem of GenX and later
is probably they don't know how to bootstrap

a Prolog system B via another Prolog system A.

Bye

P.S.: The result of using a Parser Tool are
often frustrating on the following levels:
- No operator table
- Directives are fixed
- Introducong DCGs need rebuid

Scryer Prolog has Operator Table, but mostlikely
used a Parser Tool some time in the project,
or programming templates borrow from Parser Tools.

Probably the worst recent example building a
Prolog system, which would have a reference for
the Parsing in Rust itself. So we have 2025

and there is not a single self hosting Prolog
yet, while all other programming languages such
as Java, golang, etc.. are self hosting.

Mild Shock schrieb:

Hi,

WebPL is already outdated I guess. It doesn't
show the versions of the other Prolog systems
it is using. While I had these results for

the primes example in the WebPL playground:

/* Trealla Prolog WASM */
(23568.9ms)

When I run the example here:

https://php.energy/trealla.html

I get better results:

/* trealla-js 0.27.1 */

?- time(test).
% Time elapsed 9.907s, 11263917 Inferences, 1.137 MLips

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Maybe it is with Prolog like with Dinosaurs,
when they got extinct by a meteor crash.
All that survived were some small rodents
as the story goes. Their advantage:
- Small Size
- Burrowing Behavior
- Omnivorous Diet
- Reproductive Speed

Now Scryer Prolog uses a Shift-Reduce under
the hood, the small rodent. But it might possibly
shove their end-users heavy Tabled DCG into
the face? So that this left recursion can be solved:

expr --> expr + factor

"constraint programming" got already killed
when ILOG was bought by IBM in 2008. ILOG's
optimization solver, CPLEX, has its roots in
the CHIP (Constraint Handling in Prolog)

language, 1985 at the European Computer-Industry
Research Centre (ECRC), initially using a Prolog
language interface. So its even not a Fench product.
By the time ILOG became a commercial powerhouse,

Prolog largely disappeared from their product
codebases. There was a Transition to C++ for
Performance and Industry Adoption. I have the
gut feeling that Tabled DCG is similarly dead,

especially in the light of large languages models (LLM).
But I cannot point the figure yet perfectly
at the issues. Currently exploring the sad problem
domain of this mostlikely dead horse.

A problem could be the overkill of "Logic Grammars",
that do not tolerate incorrect texts and that cannot
be applied so easy partially. Mostlikely one has
to scrutinize the assumptions behind Tabled DCG, and

review again the possibly options beyond the beaten paths.

Bye

Mild Shock schrieb:

Hi,

Because GenX and later suffers from:

Somehow methods and tools to realize
efficient DCGs in Prolog are missing. Most
DCG attempts that one sees succumb

to some declarative nonsense, creating
exponentially many spurious choice points,
you find rarely somebody mastering the Art.

Modern programmers fancy nothing else than
throwing a set of foreign library to their
Prolog system project. This is best seen in WebPL:

LALRPOP MIT/Apache-2.0 Generate the parser https://github.com/w-henderson/WebPL/blob/main/dissertation.pdf

So there is no aim at creating a self hosting
Prolog system. There is a deep distrust in
DCGs. But why build a Prolog system that will

possibly ultimately have DCG, when you distrust
in DCGs? The second problem of GenX and later
is probably they don't know how to bootstrap

a Prolog system B via another Prolog system A.

Bye

P.S.: The result of using a Parser Tool are
often frustrating on the following levels:
- No operator table
- Directives are fixed
- Introducong DCGs need rebuid

Scryer Prolog has Operator Table, but mostlikely
used a Parser Tool some time in the project,
or programming templates borrow from Parser Tools.

Probably the worst recent example building a
Prolog system, which would have a reference for
the Parsing in Rust itself. So we have 2025

and there is not a single self hosting Prolog
yet, while all other programming languages such
as Java, golang, etc.. are self hosting.

Mild Shock schrieb:

Hi,

WebPL is already outdated I guess. It doesn't
show the versions of the other Prolog systems
it is using. While I had these results for

the primes example in the WebPL playground:

/* Trealla Prolog WASM */
(23568.9ms)

When I run the example here:

https://php.energy/trealla.html

I get better results:

/* trealla-js 0.27.1 */

?- time(test).
% Time elapsed 9.907s, 11263917 Inferences, 1.137 MLips

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

I spent some time thinking about my primes.pl
test. And came to the conclusion that it
mainly tests the Prolog ALU. Things like

integer successor or integer modulo. Then
I found that Java has Math.floorMod() which
I wasn't using yet. And peng results are better:

/* Dogelog Player 2.1.2 for Java, today */
?- time(test).
% Zeit 286 ms, GC 1 ms, Lips 26302430, Uhr 15.10.2025 02:31
true.

Maybe the Java backend picks a CPU instruction
for Math.floorMod() instead of executing the
longer code sequence that is needed to correct

rem/2 into mod/2. Who knows. I also reorganized
the code a little bit, and eliminated an extra
method call in all arithmetic functions, by

inlining the arithmetic function body in the
evaluable predicate definition code. Comparison
to old measurements and some measurements of

other Prolog systems:

/* Dogelog Player 2.1.2 for Java, weeks ago */
?- time(test).
% Zeit 378 ms, GC 1 ms, Lips 19900780, Uhr 28.08.2025 17:44
true.

/* SWI-Prolog 9.0.4 */
?- time(test).
% 7,506,639 inferences, 0.363 CPU in 0.362 seconds
(100% CPU, 20693560 Lips)
true.

/* Scryer Prolog 0.9.4-639 */
?- time(test).
% CPU time: 0.365s, 7_517_613 inferences
true.

/* Trealla Prolog 2.82.23-3 */
?- time(test).
% Time elapsed 0.868s, 11263917 Inferences, 12.983 MLips
true.

Bye

P.S.: The code uses the hated mathematical mod/2,
and not the cheaper rem/2 that CPUs usually have:

test :-
len(L, 1000),
primes(L, _).

primes([], 1).
primes([J|L], J) :-
primes(L, I),
K is I+1,
search(L, K, J).

search(L, I, J) :-
mem(X, L),
I mod X =:= 0, !,
K is I+1,
search(L, K, J).
search(_, I, I).

mem(X, [X|_]).
mem(X, [_|Y]) :-
mem(X, Y).

len([], 0) :- !.
len([_|L], N) :-
N > 0,
M is N-1,
len(L, M).

Mild Shock schrieb:

Hi,

WebPL is already outdated I guess. It doesn't
show the versions of the other Prolog systems
it is using. While I had these results for

the primes example in the WebPL playground:

/* Trealla Prolog WASM */
(23568.9ms)

When I run the example here:

https://php.energy/trealla.html

I get better results:

/* trealla-js 0.27.1 */

?- time(test).
% Time elapsed 9.907s, 11263917 Inferences, 1.137 MLips

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

The change from 378 ms to 286 ms is around 25-30%
is insane. But I did both tests on a novel AI CPU.
To be precise on a AMD Ryzen AI 7 350.

But somehow I picked up rumors that AI CPUs now
might do Neural Network Branch Prediction. The
idea seems to exist in hardware at least since (2012):

Machine learning and artificial intelligence are
the current hype (again). In their new Ryzen
processors, AMD advertises the Neural Net
Prediction. It turns out this is was already
used in their older (2012) Piledriver architecture
used for example in the AMD A10-4600M. It is also
present in recent Samsung processors such as the
one powering the Galaxy S7. What is it really? https://chasethedevil.github.io/post/the_neural_network_in_your_cpu/

It can be done with Convoluted Neural Networks (CNN):

BranchNet: A Convolutional Neural Network to
Predict Hard-To-Predict Branches
To this end, Tarsa et al. proposed using convolutional
neural networks (CNNs) that are trained at
compiletime to accurately predict branches that
TAGE cannot. Given enough profiling coverage, CNNs
learn input-independent branch correlations. https://microarch.org/micro53/papers/738300a118.pdf

Interstingly the above shows cases a PGO based
Machine Learning for Branch Predictors. No clue
how they construct the CPU, that they can feed

it with offline constructed neural neutworks for
their own execution. Maybe an optimizer uses it?
But I guess a more modern solutions would not only

use CNN, but also an Attention Mechanism.

Bye

Mild Shock schrieb:

Hi,

I spent some time thinking about my primes.pl
test. And came to the conclusion that it
mainly tests the Prolog ALU. Things like

integer successor or integer modulo. Then
I found that Java has Math.floorMod() which
I wasn't using yet. And peng results are better:

/* Dogelog Player 2.1.2 for Java, today */
?- time(test).
% Zeit 286 ms, GC 1 ms, Lips 26302430, Uhr 15.10.2025 02:31
true.

Maybe the Java backend picks a CPU instruction
for Math.floorMod() instead of executing the
longer code sequence that is needed to correct

rem/2 into mod/2. Who knows. I also reorganized
the code a little bit, and eliminated an extra
method call in all arithmetic functions, by

inlining the arithmetic function body in the
evaluable predicate definition code. Comparison
to old measurements and some measurements of

other Prolog systems:

/* Dogelog Player 2.1.2 for Java, weeks ago */
?- time(test).
% Zeit 378 ms, GC 1 ms, Lips 19900780, Uhr 28.08.2025 17:44
true.

/* SWI-Prolog 9.0.4 */
?- time(test).
% 7,506,639 inferences, 0.363 CPU in 0.362 seconds
(100% CPU, 20693560 Lips)
true.

/* Scryer Prolog 0.9.4-639 */
?- time(test).
% CPU time: 0.365s, 7_517_613 inferences
true.

/* Trealla Prolog 2.82.23-3 */
?- time(test).
% Time elapsed 0.868s, 11263917 Inferences, 12.983 MLips
true.

Bye

P.S.: The code uses the hated mathematical mod/2,
and not the cheaper rem/2 that CPUs usually have:

test :-
   len(L, 1000),
   primes(L, _).

primes([], 1).
primes([J|L], J) :-
   primes(L, I),
   K is I+1,
   search(L, K, J).

search(L, I, J) :-
   mem(X, L),
   I mod X =:= 0, !,
   K is I+1,
   search(L, K, J).
search(_, I, I).

mem(X, [X|_]).
mem(X, [_|Y]) :-
   mem(X, Y).

len([], 0) :- !.
len([_|L], N) :-
   N > 0,
   M is N-1,
   len(L, M).

Mild Shock schrieb:

Hi,

WebPL is already outdated I guess. It doesn't
show the versions of the other Prolog systems
it is using. While I had these results for

the primes example in the WebPL playground:

/* Trealla Prolog WASM */
(23568.9ms)

When I run the example here:

https://php.energy/trealla.html

I get better results:

/* trealla-js 0.27.1 */

?- time(test).
% Time elapsed 9.907s, 11263917 Inferences, 1.137 MLips

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

It seems I am having problems pacing with
all the new fancy toys. Wasn't able to really
benchmark my NPU from a Desktop AI machine,

picked the wrong driver. Need to try again.
What worked was benchmarking Mobile AI machines.
I just grabbed Geekbench AI and some devices:

USA Fab, M4:

sANN hANN qANN
iPad CPU 4848 7947 6353
iPad GPU 9752 11383 10051
iPad NPU 4873 36544 *51634*

China Fab, Snapdragon:

sANN hANN qANN
Redmi CPU 1044 950 1723
Redmi GPU 480 905 737
Redmi NNAPI 205 205 469
Redmi QNN 226 226 *10221*

Speed-Up via NPU is factor 10x. See the column
qANN which means quantizised artificial neural
networks, when NPU or QNN is picked.

The mobile AI NPUs are optimized using
mimimal amounts of energy, and minimal amounts
of space squeezing (distilling) everything

into INT8 and INT4.

Bye

Mild Shock schrieb:

Hi,

The change from 378 ms to 286 ms is around 25-30%
is insane. But I did both tests on a novel AI CPU.
To be precise on a AMD Ryzen AI 7 350.

But somehow I picked up rumors that AI CPUs now
might do Neural Network Branch Prediction. The
idea seems to exist in hardware at least since (2012):

Machine learning and artificial intelligence are
the current hype (again). In their new Ryzen
processors, AMD advertises the Neural Net
Prediction. It turns out this is was already
used in their older (2012) Piledriver architecture
used for example in the AMD A10-4600M. It is also
present in recent Samsung processors such as the
one powering the Galaxy S7. What is it really? https://chasethedevil.github.io/post/the_neural_network_in_your_cpu/

It can be done with Convoluted Neural Networks (CNN):

BranchNet: A Convolutional Neural Network to
Predict Hard-To-Predict Branches
To this end, Tarsa et al. proposed using convolutional
neural networks (CNNs) that are trained at
compiletime to accurately predict branches that
TAGE cannot. Given enough profiling coverage, CNNs
learn input-independent branch correlations. https://microarch.org/micro53/papers/738300a118.pdf

Interstingly the above shows cases a PGO based
Machine Learning for Branch Predictors. No clue
how they construct the CPU, that they can feed

it with offline constructed neural neutworks for
their own execution. Maybe an optimizer uses it?
But I guess a more modern  solutions would not only

use CNN, but also an Attention Mechanism.

Bye

Mild Shock schrieb:

Hi,

I spent some time thinking about my primes.pl
test. And came to the conclusion that it
mainly tests the Prolog ALU. Things like

integer successor or integer modulo. Then
I found that Java has Math.floorMod() which
I wasn't using yet. And peng results are better:

/* Dogelog Player 2.1.2 for Java, today */
?- time(test).
% Zeit 286 ms, GC 1 ms, Lips 26302430, Uhr 15.10.2025 02:31
true.

Maybe the Java backend picks a CPU instruction
for Math.floorMod() instead of executing the
longer code sequence that is needed to correct

rem/2 into mod/2. Who knows. I also reorganized
the code a little bit, and eliminated an extra
method call in all arithmetic functions, by

inlining the arithmetic function body in the
evaluable predicate definition code. Comparison
to old measurements and some measurements of

other Prolog systems:

/* Dogelog Player 2.1.2 for Java, weeks ago */
?- time(test).
% Zeit 378 ms, GC 1 ms, Lips 19900780, Uhr 28.08.2025 17:44
true.

/* SWI-Prolog 9.0.4 */
?- time(test).
% 7,506,639 inferences, 0.363 CPU in 0.362 seconds
(100% CPU, 20693560 Lips)
true.

/* Scryer Prolog 0.9.4-639 */
?- time(test).
% CPU time: 0.365s, 7_517_613 inferences
true.

/* Trealla Prolog 2.82.23-3 */
?- time(test).
% Time elapsed 0.868s, 11263917 Inferences, 12.983 MLips
true.

Bye

P.S.: The code uses the hated mathematical mod/2,
and not the cheaper rem/2 that CPUs usually have:

test :-
    len(L, 1000),
    primes(L, _).

primes([], 1).
primes([J|L], J) :-
    primes(L, I),
    K is I+1,
    search(L, K, J).

search(L, I, J) :-
    mem(X, L),
    I mod X =:= 0, !,
    K is I+1,
    search(L, K, J).
search(_, I, I).

mem(X, [X|_]).
mem(X, [_|Y]) :-
    mem(X, Y).

len([], 0) :- !.
len([_|L], N) :-
    N > 0,
    M is N-1,
    len(L, M).

Mild Shock schrieb:

Hi,

WebPL is already outdated I guess. It doesn't
show the versions of the other Prolog systems
it is using. While I had these results for

the primes example in the WebPL playground:

/* Trealla Prolog WASM */
(23568.9ms)

When I run the example here:

https://php.energy/trealla.html

I get better results:

/* trealla-js 0.27.1 */

?- time(test).
% Time elapsed 9.907s, 11263917 Inferences, 1.137 MLips

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

But not only Mobie AI and Desktop AI are making
a broader imprint now. We might also experience
Workstation AI, with a 3'000.- USD price tag:

You Can't Buy This... Yet! The NVIDIA GB10 from Dell
The New Superchip that Terrifies the Cloud! https://www.youtube.com/watch?v=x1qViw4xyVo

So whats going on? I was asking Phind, which is
driven by a 70B model tailored towards developers:

Q: Is there an AI inflection point right now ,
with NPUs in mobile, desktop and workstation

A: Evidence of the Inflection Point

- Mobile Leadership
NPUs originated in smartphones
Now becoming ubiquitous across all device types
Enabling sophisticated AI features at consumer price points

- Desktop Revolution
Major manufacturers implementing NPUs across product lines
Apple's Neural Engine integrated into M-series chips
Qualcomm, Intel, and AMD incorporating AI accelerators

- Workstation Transformation
Professional-grade NPUs in mobile workstations
Demonstrated superior performance for AI-specific tasks
Enabling local processing of previously cloud-dependent workloads

https://www.phind.com/search/cmgs1s6jv00023h67g5z2aaa0

Bye

Mild Shock schrieb:

Hi,

It seems I am having problems pacing with
all the new fancy toys. Wasn't able to really
benchmark my NPU from a Desktop AI machine,

picked the wrong driver. Need to try again.
What worked was benchmarking Mobile AI machines.
I just grabbed Geekbench AI and some devices:

USA Fab, M4:

    sANN    hANN    qANN
iPad CPU    4848    7947    6353
iPad GPU    9752    11383    10051
iPad NPU    4873    36544    *51634*

China Fab, Snapdragon:

    sANN    hANN    qANN
Redmi CPU    1044    950    1723
Redmi GPU    480    905    737
Redmi NNAPI    205    205    469
Redmi QNN    226    226    *10221*

Speed-Up via NPU is factor 10x. See the column
qANN which means quantizised artificial neural
networks, when NPU or QNN is picked.

The mobile AI NPUs are optimized using
mimimal amounts of energy, and minimal amounts
of space squeezing (distilling) everything

into INT8 and INT4.

Bye

Mild Shock schrieb:

Hi,

The change from 378 ms to 286 ms is around 25-30%
is insane. But I did both tests on a novel AI CPU.
To be precise on a AMD Ryzen AI 7 350.

But somehow I picked up rumors that AI CPUs now
might do Neural Network Branch Prediction. The
idea seems to exist in hardware at least since (2012):

Machine learning and artificial intelligence are
the current hype (again). In their new Ryzen
processors, AMD advertises the Neural Net
Prediction. It turns out this is was already
used in their older (2012) Piledriver architecture
used for example in the AMD A10-4600M. It is also
present in recent Samsung processors such as the
one powering the Galaxy S7. What is it really?
https://chasethedevil.github.io/post/the_neural_network_in_your_cpu/

It can be done with Convoluted Neural Networks (CNN):

BranchNet: A Convolutional Neural Network to
Predict Hard-To-Predict Branches
To this end, Tarsa et al. proposed using convolutional
neural networks (CNNs) that are trained at
compiletime to accurately predict branches that
TAGE cannot. Given enough profiling coverage, CNNs
learn input-independent branch correlations.
https://microarch.org/micro53/papers/738300a118.pdf

Interstingly the above shows cases a PGO based
Machine Learning for Branch Predictors. No clue
how they construct the CPU, that they can feed

it with offline constructed neural neutworks for
their own execution. Maybe an optimizer uses it?
But I guess a more modern  solutions would not only

use CNN, but also an Attention Mechanism.

Bye

Mild Shock schrieb:

Hi,

I spent some time thinking about my primes.pl
test. And came to the conclusion that it
mainly tests the Prolog ALU. Things like

integer successor or integer modulo. Then
I found that Java has Math.floorMod() which
I wasn't using yet. And peng results are better:

/* Dogelog Player 2.1.2 for Java, today */
?- time(test).
% Zeit 286 ms, GC 1 ms, Lips 26302430, Uhr 15.10.2025 02:31
true.

Maybe the Java backend picks a CPU instruction
for Math.floorMod() instead of executing the
longer code sequence that is needed to correct

rem/2 into mod/2. Who knows. I also reorganized
the code a little bit, and eliminated an extra
method call in all arithmetic functions, by

inlining the arithmetic function body in the
evaluable predicate definition code. Comparison
to old measurements and some measurements of

other Prolog systems:

/* Dogelog Player 2.1.2 for Java, weeks ago */
?- time(test).
% Zeit 378 ms, GC 1 ms, Lips 19900780, Uhr 28.08.2025 17:44
true.

/* SWI-Prolog 9.0.4 */
?- time(test).
% 7,506,639 inferences, 0.363 CPU in 0.362 seconds
(100% CPU, 20693560 Lips)
true.

/* Scryer Prolog 0.9.4-639 */
?- time(test).
% CPU time: 0.365s, 7_517_613 inferences
true.

/* Trealla Prolog 2.82.23-3 */
?- time(test).
% Time elapsed 0.868s, 11263917 Inferences, 12.983 MLips
true.

Bye

P.S.: The code uses the hated mathematical mod/2,
and not the cheaper rem/2 that CPUs usually have:

test :-
    len(L, 1000),
    primes(L, _).

primes([], 1).
primes([J|L], J) :-
    primes(L, I),
    K is I+1,
    search(L, K, J).

search(L, I, J) :-
    mem(X, L),
    I mod X =:= 0, !,
    K is I+1,
    search(L, K, J).
search(_, I, I).

mem(X, [X|_]).
mem(X, [_|Y]) :-
    mem(X, Y).

len([], 0) :- !.
len([_|L], N) :-
    N > 0,
    M is N-1,
    len(L, M).

Mild Shock schrieb:

Hi,

WebPL is already outdated I guess. It doesn't
show the versions of the other Prolog systems
it is using. While I had these results for

the primes example in the WebPL playground:

/* Trealla Prolog WASM */
(23568.9ms)

When I run the example here:

https://php.energy/trealla.html

I get better results:

/* trealla-js 0.27.1 */

?- time(test).
% Time elapsed 9.907s, 11263917 Inferences, 1.137 MLips

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Thinks are definitively accelerating. I really would
like to use an AI that knows about all the News of today.
This bloody cut date is so annoying.

Further indicative that AI is accelerating:

In August 2025, Sam Altman dropped a bombshell:

*months, not years: Rushing GPT-6*
In August 2025, Sam Altman dropped a bombshell:
GPT-6 is already in development and coming sooner
than you think. Not in two years, but
potentially in months.
https://www.youtube.com/watch?v=44mJb5sKji0

Karpathy, coined vibe coding, released in October 2025:

*nanochat: The best ChatGPT that $100 can buy*
This repo is a full-stack implementation of an
LLM like ChatGPT in a single, clean, minimal,
hackable, dependency-lite codebase. nanochat is
designed to run on a single 8XH100 node via
scripts like speedrun.sh, that run the
entire pipeline start to end.
https://github.com/karpathy/nanochat

Bye

Mild Shock schrieb:

Hi,

But not only Mobie AI and Desktop AI are making
a broader imprint now. We might also experience
Workstation AI, with a 3'000.- USD price tag:

You Can't Buy This... Yet! The NVIDIA GB10 from Dell
The New Superchip that Terrifies the Cloud! https://www.youtube.com/watch?v=x1qViw4xyVo

So whats going on? I was asking Phind, which is
driven by a 70B model tailored towards developers:

Q: Is there an AI inflection point right now ,
   with NPUs in mobile, desktop and workstation

A: Evidence of the Inflection Point

- Mobile Leadership
  NPUs originated in smartphones
  Now becoming ubiquitous across all device types
  Enabling sophisticated AI features at consumer price points

- Desktop Revolution
  Major manufacturers implementing NPUs across product lines
  Apple's Neural Engine integrated into M-series chips
  Qualcomm, Intel, and AMD incorporating AI accelerators

- Workstation Transformation
  Professional-grade NPUs in mobile workstations
  Demonstrated superior performance for AI-specific tasks
  Enabling local processing of previously cloud-dependent workloads

https://www.phind.com/search/cmgs1s6jv00023h67g5z2aaa0

Bye

Mild Shock schrieb:

Hi,

It seems I am having problems pacing with
all the new fancy toys. Wasn't able to really
benchmark my NPU from a Desktop AI machine,

picked the wrong driver. Need to try again.
What worked was benchmarking Mobile AI machines.
I just grabbed Geekbench AI and some devices:

USA Fab, M4:

     sANN    hANN    qANN
iPad CPU    4848    7947    6353
iPad GPU    9752    11383    10051
iPad NPU    4873    36544    *51634*

China Fab, Snapdragon:

     sANN    hANN    qANN
Redmi CPU    1044    950    1723
Redmi GPU    480    905    737
Redmi NNAPI    205    205    469
Redmi QNN    226    226    *10221*

Speed-Up via NPU is factor 10x. See the column
qANN which means quantizised artificial neural
networks, when NPU or QNN is picked.

The mobile AI NPUs are optimized using
mimimal amounts of energy, and minimal amounts
of space squeezing (distilling) everything

into INT8 and INT4.

Bye

Mild Shock schrieb:

Hi,

The change from 378 ms to 286 ms is around 25-30%
is insane. But I did both tests on a novel AI CPU.
To be precise on a AMD Ryzen AI 7 350.

But somehow I picked up rumors that AI CPUs now
might do Neural Network Branch Prediction. The
idea seems to exist in hardware at least since (2012):

Machine learning and artificial intelligence are
the current hype (again). In their new Ryzen
processors, AMD advertises the Neural Net
Prediction. It turns out this is was already
used in their older (2012) Piledriver architecture
used for example in the AMD A10-4600M. It is also
present in recent Samsung processors such as the
one powering the Galaxy S7. What is it really?
https://chasethedevil.github.io/post/the_neural_network_in_your_cpu/

It can be done with Convoluted Neural Networks (CNN):

BranchNet: A Convolutional Neural Network to
Predict Hard-To-Predict Branches
To this end, Tarsa et al. proposed using convolutional
neural networks (CNNs) that are trained at
compiletime to accurately predict branches that
TAGE cannot. Given enough profiling coverage, CNNs
learn input-independent branch correlations.
https://microarch.org/micro53/papers/738300a118.pdf

Interstingly the above shows cases a PGO based
Machine Learning for Branch Predictors. No clue
how they construct the CPU, that they can feed

it with offline constructed neural neutworks for
their own execution. Maybe an optimizer uses it?
But I guess a more modern  solutions would not only

use CNN, but also an Attention Mechanism.

Bye

Mild Shock schrieb:

Hi,

I spent some time thinking about my primes.pl
test. And came to the conclusion that it
mainly tests the Prolog ALU. Things like

integer successor or integer modulo. Then
I found that Java has Math.floorMod() which
I wasn't using yet. And peng results are better:

/* Dogelog Player 2.1.2 for Java, today */
?- time(test).
% Zeit 286 ms, GC 1 ms, Lips 26302430, Uhr 15.10.2025 02:31
true.

Maybe the Java backend picks a CPU instruction
for Math.floorMod() instead of executing the
longer code sequence that is needed to correct

rem/2 into mod/2. Who knows. I also reorganized
the code a little bit, and eliminated an extra
method call in all arithmetic functions, by

inlining the arithmetic function body in the
evaluable predicate definition code. Comparison
to old measurements and some measurements of

other Prolog systems:

/* Dogelog Player 2.1.2 for Java, weeks ago */
?- time(test).
% Zeit 378 ms, GC 1 ms, Lips 19900780, Uhr 28.08.2025 17:44
true.

/* SWI-Prolog 9.0.4 */
?- time(test).
% 7,506,639 inferences, 0.363 CPU in 0.362 seconds
(100% CPU, 20693560 Lips)
true.

/* Scryer Prolog 0.9.4-639 */
?- time(test).
% CPU time: 0.365s, 7_517_613 inferences
true.

/* Trealla Prolog 2.82.23-3 */
?- time(test).
% Time elapsed 0.868s, 11263917 Inferences, 12.983 MLips
true.

Bye

P.S.: The code uses the hated mathematical mod/2,
and not the cheaper rem/2 that CPUs usually have:

test :-
    len(L, 1000),
    primes(L, _).

primes([], 1).
primes([J|L], J) :-
    primes(L, I),
    K is I+1,
    search(L, K, J).

search(L, I, J) :-
    mem(X, L),
    I mod X =:= 0, !,
    K is I+1,
    search(L, K, J).
search(_, I, I).

mem(X, [X|_]).
mem(X, [_|Y]) :-
    mem(X, Y).

len([], 0) :- !.
len([_|L], N) :-
    N > 0,
    M is N-1,
    len(L, M).

Mild Shock schrieb:

Hi,

WebPL is already outdated I guess. It doesn't
show the versions of the other Prolog systems
it is using. While I had these results for

the primes example in the WebPL playground:

/* Trealla Prolog WASM */
(23568.9ms)

When I run the example here:

https://php.energy/trealla.html

I get better results:

/* trealla-js 0.27.1 */

?- time(test).
% Time elapsed 9.907s, 11263917 Inferences, 1.137 MLips

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Give Julio Di Egidio the bloody money. He is
craving for 300 USD so that he can buy the
ISO Prolog core standard. Just imagine he would want

to build a MiniMind. Just lets put some more
prespective on the current costs:

This open-source project aims to train a super-small
language model MiniMind with only 3 RMB cost and
2 hours, starting completely from scratch. The
MiniMind series is extremely lightweight, with the
smallest version being 1/7000 the size of GPT-3,
making it possible to train quickly on even the
most ordinary personal GPUs. https://github.com/jingyaogong/minimind/blob/master/README_en.md

ChatGPT tells me that most of the numbers
are correct when you rent a GPU by the hour.
But what about a 100% ownership of a GPU for

a year. I find this might cost 12'000 USD.
One has to separate platforms for execution from
those platforms for training:

GEX44: for AI inference
Nvidia RTX™ 4000, 184 EUR / month

GEX130: for AI training
NVIDIA RTX™ 6000, 813 EUR / month https://www.hetzner.com/dedicated-rootserver/matrix-gpu/

Bye

Mild Shock schrieb:

Hi,

Thinks are definitively accelerating. I really would
like to use an AI that knows about all the News of today.
This bloody cut date is so annoying.

Further indicative that AI is accelerating:

In August 2025, Sam Altman dropped a bombshell:

*months, not years: Rushing GPT-6*
In August 2025, Sam Altman dropped a bombshell:
GPT-6 is already in development and coming sooner
than you think. Not in two years, but
potentially in months.
https://www.youtube.com/watch?v=44mJb5sKji0

Karpathy, coined vibe coding, released in October 2025:

*nanochat: The best ChatGPT that $100 can buy*
This repo is a full-stack implementation of an
LLM like ChatGPT in a single, clean, minimal,
hackable, dependency-lite codebase. nanochat is
designed to run on a single 8XH100 node via
scripts like speedrun.sh, that run the
entire pipeline start to end.
https://github.com/karpathy/nanochat

Bye

Mild Shock schrieb:

Hi,

But not only Mobie AI and Desktop AI are making
a broader imprint now. We might also experience
Workstation AI, with a 3'000.- USD price tag:

You Can't Buy This... Yet! The NVIDIA GB10 from Dell
The New Superchip that Terrifies the Cloud!
https://www.youtube.com/watch?v=x1qViw4xyVo

So whats going on? I was asking Phind, which is
driven by a 70B model tailored towards developers:

Q: Is there an AI inflection point right now ,
    with NPUs in mobile, desktop and workstation

A: Evidence of the Inflection Point

- Mobile Leadership
   NPUs originated in smartphones
   Now becoming ubiquitous across all device types
   Enabling sophisticated AI features at consumer price points

- Desktop Revolution
   Major manufacturers implementing NPUs across product lines
   Apple's Neural Engine integrated into M-series chips
   Qualcomm, Intel, and AMD incorporating AI accelerators

- Workstation Transformation
   Professional-grade NPUs in mobile workstations
   Demonstrated superior performance for AI-specific tasks
   Enabling local processing of previously cloud-dependent workloads

https://www.phind.com/search/cmgs1s6jv00023h67g5z2aaa0

Bye

Mild Shock schrieb:

Hi,

It seems I am having problems pacing with
all the new fancy toys. Wasn't able to really
benchmark my NPU from a Desktop AI machine,

picked the wrong driver. Need to try again.
What worked was benchmarking Mobile AI machines.
I just grabbed Geekbench AI and some devices:

USA Fab, M4:

     sANN    hANN    qANN
iPad CPU    4848    7947    6353
iPad GPU    9752    11383    10051
iPad NPU    4873    36544    *51634*

China Fab, Snapdragon:

     sANN    hANN    qANN
Redmi CPU    1044    950    1723
Redmi GPU    480    905    737
Redmi NNAPI    205    205    469
Redmi QNN    226    226    *10221*

Speed-Up via NPU is factor 10x. See the column
qANN which means quantizised artificial neural
networks, when NPU or QNN is picked.

The mobile AI NPUs are optimized using
mimimal amounts of energy, and minimal amounts
of space squeezing (distilling) everything

into INT8 and INT4.

Bye

Mild Shock schrieb:

Hi,

The change from 378 ms to 286 ms is around 25-30%
is insane. But I did both tests on a novel AI CPU.
To be precise on a AMD Ryzen AI 7 350.

But somehow I picked up rumors that AI CPUs now
might do Neural Network Branch Prediction. The
idea seems to exist in hardware at least since (2012):

Machine learning and artificial intelligence are
the current hype (again). In their new Ryzen
processors, AMD advertises the Neural Net
Prediction. It turns out this is was already
used in their older (2012) Piledriver architecture
used for example in the AMD A10-4600M. It is also
present in recent Samsung processors such as the
one powering the Galaxy S7. What is it really?
https://chasethedevil.github.io/post/the_neural_network_in_your_cpu/

It can be done with Convoluted Neural Networks (CNN):

BranchNet: A Convolutional Neural Network to
Predict Hard-To-Predict Branches
To this end, Tarsa et al. proposed using convolutional
neural networks (CNNs) that are trained at
compiletime to accurately predict branches that
TAGE cannot. Given enough profiling coverage, CNNs
learn input-independent branch correlations.
https://microarch.org/micro53/papers/738300a118.pdf

Interstingly the above shows cases a PGO based
Machine Learning for Branch Predictors. No clue
how they construct the CPU, that they can feed

it with offline constructed neural neutworks for
their own execution. Maybe an optimizer uses it?
But I guess a more modern  solutions would not only

use CNN, but also an Attention Mechanism.

Bye

Mild Shock schrieb:

Hi,

I spent some time thinking about my primes.pl
test. And came to the conclusion that it
mainly tests the Prolog ALU. Things like

integer successor or integer modulo. Then
I found that Java has Math.floorMod() which
I wasn't using yet. And peng results are better:

/* Dogelog Player 2.1.2 for Java, today */
?- time(test).
% Zeit 286 ms, GC 1 ms, Lips 26302430, Uhr 15.10.2025 02:31
true.

Maybe the Java backend picks a CPU instruction
for Math.floorMod() instead of executing the
longer code sequence that is needed to correct

rem/2 into mod/2. Who knows. I also reorganized
the code a little bit, and eliminated an extra
method call in all arithmetic functions, by

inlining the arithmetic function body in the
evaluable predicate definition code. Comparison
to old measurements and some measurements of

other Prolog systems:

/* Dogelog Player 2.1.2 for Java, weeks ago */
?- time(test).
% Zeit 378 ms, GC 1 ms, Lips 19900780, Uhr 28.08.2025 17:44
true.

/* SWI-Prolog 9.0.4 */
?- time(test).
% 7,506,639 inferences, 0.363 CPU in 0.362 seconds
(100% CPU, 20693560 Lips)
true.

/* Scryer Prolog 0.9.4-639 */
?- time(test).
% CPU time: 0.365s, 7_517_613 inferences
true.

/* Trealla Prolog 2.82.23-3 */
?- time(test).
% Time elapsed 0.868s, 11263917 Inferences, 12.983 MLips
true.

Bye

P.S.: The code uses the hated mathematical mod/2,
and not the cheaper rem/2 that CPUs usually have:

test :-
    len(L, 1000),
    primes(L, _).

primes([], 1).
primes([J|L], J) :-
    primes(L, I),
    K is I+1,
    search(L, K, J).

search(L, I, J) :-
    mem(X, L),
    I mod X =:= 0, !,
    K is I+1,
    search(L, K, J).
search(_, I, I).

mem(X, [X|_]).
mem(X, [_|Y]) :-
    mem(X, Y).

len([], 0) :- !.
len([_|L], N) :-
    N > 0,
    M is N-1,
    len(L, M).

Mild Shock schrieb:

Hi,

WebPL is already outdated I guess. It doesn't
show the versions of the other Prolog systems
it is using. While I had these results for

the primes example in the WebPL playground:

/* Trealla Prolog WASM */
(23568.9ms)

When I run the example here:

https://php.energy/trealla.html

I get better results:

/* trealla-js 0.27.1 */

?- time(test).
% Time elapsed 9.907s, 11263917 Inferences, 1.137 MLips

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

This is probably what my AI Laptop can do. The
demo shows AMD Ryzen AI 7 340 micro. I have a
AMD Ryzen AI 7 350 laptop.

100% Powered by AMD Ryzen™ AI NPU https://www.youtube.com/watch?v=0t8ijUPg4A0

Only I am too stupid / too lazy to dig up the right
drivers and install FastFlowLM. But one sees in
the video how the NPU gets 30% - 60% occupied,

and it does the transcription of a YouTube video,
into text (via Whisper-large-v3-turbo from OpenAI).
The demo then switches to summarize mode

(via GPT-OSS-20B from OpenAI). And boom the,
NPU goes to 100%!

Bye

P.S.: I didn't know about the OpenAI and AMD partnership,
also buying an AMD AI Laptop was not motivated by
this development. Not sure whether its really a big thing:

AMD and OpenAI announce partnership https://openai.com/index/openai-amd-strategic-partnership/

It might be a good thing for end users like me, if
Edge uses cases like the above become more common.
But they still depend on models trained not on the Edge,

but in a Data Center. The AMD Instinct product line directly
competes with Nvidia's Tesla and Intel's Xeon Phi and Data
Center GPU lines of machine learning and GPGPU cards.

Mild Shock schrieb:

Hi,

It seems I am having problems pacing with
all the new fancy toys. Wasn't able to really
benchmark my NPU from a Desktop AI machine,

picked the wrong driver. Need to try again.
What worked was benchmarking Mobile AI machines.
I just grabbed Geekbench AI and some devices:

USA Fab, M4:

    sANN    hANN    qANN
iPad CPU    4848    7947    6353
iPad GPU    9752    11383    10051
iPad NPU    4873    36544    *51634*

China Fab, Snapdragon:

    sANN    hANN    qANN
Redmi CPU    1044    950    1723
Redmi GPU    480    905    737
Redmi NNAPI    205    205    469
Redmi QNN    226    226    *10221*

Speed-Up via NPU is factor 10x. See the column
qANN which means quantizised artificial neural
networks, when NPU or QNN is picked.

The mobile AI NPUs are optimized using
mimimal amounts of energy, and minimal amounts
of space squeezing (distilling) everything

into INT8 and INT4.

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Vertex AI Training is more expensive:

"Vertex AI provides a managed training
service that enables you to operationalize
large scale model training. You can use
Vertex AI to run training applications
based on any machine learning (ML) framework
on Google Cloud infrastructure. Vertex AI
also has integrated support that simplifies
the preparation process for model training
and serving for the PyTorch, TensorFlow,
scikit-learn, and XGBoost frameworks. https://cloud.google.com/products/calculator

For 10 Jobs à 10 hours, it wants 1000 USD
from me. The accelerator is TPU V3. Nevertheless
might be the cheaper option to use the RTX™ 6000

for 720 hours. ChatGPT gives me this calculation:

Scenario I: Single TPU, RTX 6000 wins
4.43×10^19 FLOPs versus ~2.36×10^20 FLOPs

Scenario II: 8 chips TPU, RTX 6000 looses
~3.54×10^20 FLOPs versus ~2.36×10^20 FLOPs

Bye

Mild Shock schrieb:

Hi,

Give Julio Di Egidio the bloody money. He is
craving for 300 USD so that he can buy the
ISO Prolog core standard. Just imagine he would want

to build a MiniMind. Just lets put some more
prespective on the current costs:

This open-source project aims to train a super-small
language model MiniMind with only 3 RMB cost and
2 hours, starting completely from scratch. The
MiniMind series is extremely lightweight, with the
smallest version being 1/7000 the size of GPT-3,
making it possible to train quickly on even the
most ordinary personal GPUs. https://github.com/jingyaogong/minimind/blob/master/README_en.md

ChatGPT tells me that most of the numbers
are correct when you rent a GPU by the hour.
But what about a 100% ownership of a GPU for

a year. I find this might cost 12'000 USD.
One has to separate platforms for execution from
those platforms for training:

GEX44: for AI inference
Nvidia RTX™ 4000, 184 EUR / month

GEX130: for AI training
NVIDIA RTX™ 6000, 813 EUR / month https://www.hetzner.com/dedicated-rootserver/matrix-gpu/

Bye

Mild Shock schrieb:

Hi,

Thinks are definitively accelerating. I really would
like to use an AI that knows about all the News of today.
This bloody cut date is so annoying.

Further indicative that AI is accelerating:

In August 2025, Sam Altman dropped a bombshell:

*months, not years: Rushing GPT-6*
In August 2025, Sam Altman dropped a bombshell:
GPT-6 is already in development and coming sooner
than you think. Not in two years, but
potentially in months.
https://www.youtube.com/watch?v=44mJb5sKji0

Karpathy, coined vibe coding, released in October 2025:

*nanochat: The best ChatGPT that $100 can buy*
This repo is a full-stack implementation of an
LLM like ChatGPT in a single, clean, minimal,
hackable, dependency-lite codebase. nanochat is
designed to run on a single 8XH100 node via
scripts like speedrun.sh, that run the
entire pipeline start to end.
https://github.com/karpathy/nanochat

Bye

Mild Shock schrieb:

Hi,

But not only Mobie AI and Desktop AI are making
a broader imprint now. We might also experience
Workstation AI, with a 3'000.- USD price tag:

You Can't Buy This... Yet! The NVIDIA GB10 from Dell
The New Superchip that Terrifies the Cloud!
https://www.youtube.com/watch?v=x1qViw4xyVo

So whats going on? I was asking Phind, which is
driven by a 70B model tailored towards developers:

Q: Is there an AI inflection point right now ,
    with NPUs in mobile, desktop and workstation

A: Evidence of the Inflection Point

- Mobile Leadership
   NPUs originated in smartphones
   Now becoming ubiquitous across all device types
   Enabling sophisticated AI features at consumer price points

- Desktop Revolution
   Major manufacturers implementing NPUs across product lines
   Apple's Neural Engine integrated into M-series chips
   Qualcomm, Intel, and AMD incorporating AI accelerators

- Workstation Transformation
   Professional-grade NPUs in mobile workstations
   Demonstrated superior performance for AI-specific tasks
   Enabling local processing of previously cloud-dependent workloads

https://www.phind.com/search/cmgs1s6jv00023h67g5z2aaa0

Bye

Mild Shock schrieb:

Hi,

It seems I am having problems pacing with
all the new fancy toys. Wasn't able to really
benchmark my NPU from a Desktop AI machine,

picked the wrong driver. Need to try again.
What worked was benchmarking Mobile AI machines.
I just grabbed Geekbench AI and some devices:

USA Fab, M4:

     sANN    hANN    qANN
iPad CPU    4848    7947    6353
iPad GPU    9752    11383    10051
iPad NPU    4873    36544    *51634*

China Fab, Snapdragon:

     sANN    hANN    qANN
Redmi CPU    1044    950    1723
Redmi GPU    480    905    737
Redmi NNAPI    205    205    469
Redmi QNN    226    226    *10221*

Speed-Up via NPU is factor 10x. See the column
qANN which means quantizised artificial neural
networks, when NPU or QNN is picked.

The mobile AI NPUs are optimized using
mimimal amounts of energy, and minimal amounts
of space squeezing (distilling) everything

into INT8 and INT4.

Bye

Mild Shock schrieb:

Hi,

The change from 378 ms to 286 ms is around 25-30%
is insane. But I did both tests on a novel AI CPU.
To be precise on a AMD Ryzen AI 7 350.

But somehow I picked up rumors that AI CPUs now
might do Neural Network Branch Prediction. The
idea seems to exist in hardware at least since (2012):

Machine learning and artificial intelligence are
the current hype (again). In their new Ryzen
processors, AMD advertises the Neural Net
Prediction. It turns out this is was already
used in their older (2012) Piledriver architecture
used for example in the AMD A10-4600M. It is also
present in recent Samsung processors such as the
one powering the Galaxy S7. What is it really?
https://chasethedevil.github.io/post/the_neural_network_in_your_cpu/ >>>>>
It can be done with Convoluted Neural Networks (CNN):

BranchNet: A Convolutional Neural Network to
Predict Hard-To-Predict Branches
To this end, Tarsa et al. proposed using convolutional
neural networks (CNNs) that are trained at
compiletime to accurately predict branches that
TAGE cannot. Given enough profiling coverage, CNNs
learn input-independent branch correlations.
https://microarch.org/micro53/papers/738300a118.pdf

Interstingly the above shows cases a PGO based
Machine Learning for Branch Predictors. No clue
how they construct the CPU, that they can feed

it with offline constructed neural neutworks for
their own execution. Maybe an optimizer uses it?
But I guess a more modern  solutions would not only

use CNN, but also an Attention Mechanism.

Bye

Mild Shock schrieb:

Hi,

I spent some time thinking about my primes.pl
test. And came to the conclusion that it
mainly tests the Prolog ALU. Things like

integer successor or integer modulo. Then
I found that Java has Math.floorMod() which
I wasn't using yet. And peng results are better:

/* Dogelog Player 2.1.2 for Java, today */
?- time(test).
% Zeit 286 ms, GC 1 ms, Lips 26302430, Uhr 15.10.2025 02:31
true.

Maybe the Java backend picks a CPU instruction
for Math.floorMod() instead of executing the
longer code sequence that is needed to correct

rem/2 into mod/2. Who knows. I also reorganized
the code a little bit, and eliminated an extra
method call in all arithmetic functions, by

inlining the arithmetic function body in the
evaluable predicate definition code. Comparison
to old measurements and some measurements of

other Prolog systems:

/* Dogelog Player 2.1.2 for Java, weeks ago */
?- time(test).
% Zeit 378 ms, GC 1 ms, Lips 19900780, Uhr 28.08.2025 17:44
true.

/* SWI-Prolog 9.0.4 */
?- time(test).
% 7,506,639 inferences, 0.363 CPU in 0.362 seconds
(100% CPU, 20693560 Lips)
true.

/* Scryer Prolog 0.9.4-639 */
?- time(test).
% CPU time: 0.365s, 7_517_613 inferences
true.

/* Trealla Prolog 2.82.23-3 */
?- time(test).
% Time elapsed 0.868s, 11263917 Inferences, 12.983 MLips
true.

Bye

P.S.: The code uses the hated mathematical mod/2,
and not the cheaper rem/2 that CPUs usually have:

test :-
    len(L, 1000),
    primes(L, _).

primes([], 1).
primes([J|L], J) :-
    primes(L, I),
    K is I+1,
    search(L, K, J).

search(L, I, J) :-
    mem(X, L),
    I mod X =:= 0, !,
    K is I+1,
    search(L, K, J).
search(_, I, I).

mem(X, [X|_]).
mem(X, [_|Y]) :-
    mem(X, Y).

len([], 0) :- !.
len([_|L], N) :-
    N > 0,
    M is N-1,
    len(L, M).

Mild Shock schrieb:

Hi,

WebPL is already outdated I guess. It doesn't
show the versions of the other Prolog systems
it is using. While I had these results for

the primes example in the WebPL playground:

/* Trealla Prolog WASM */
(23568.9ms)

When I run the example here:

https://php.energy/trealla.html

I get better results:

/* trealla-js 0.27.1 */

?- time(test).
% Time elapsed 9.907s, 11263917 Inferences, 1.137 MLips

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Candidate Recommendation Draft - 30 September 2025
https://www.w3.org/TR/webnn

WebNN samples by Ningxin Hu, Intel, Shanghai https://github.com/webmachinelearning/webnn-samples

Bye

Mild Shock schrieb:

Hi,

It seems I am having problems pacing with
all the new fancy toys. Wasn't able to really
benchmark my NPU from a Desktop AI machine,

picked the wrong driver. Need to try again.
What worked was benchmarking Mobile AI machines.
I just grabbed Geekbench AI and some devices:

USA Fab, M4:

    sANN    hANN    qANN
iPad CPU    4848    7947    6353
iPad GPU    9752    11383    10051
iPad NPU    4873    36544    *51634*

China Fab, Snapdragon:

    sANN    hANN    qANN
Redmi CPU    1044    950    1723
Redmi GPU    480    905    737
Redmi NNAPI    205    205    469
Redmi QNN    226    226    *10221*

Speed-Up via NPU is factor 10x. See the column
qANN which means quantizised artificial neural
networks, when NPU or QNN is picked.

The mobile AI NPUs are optimized using
mimimal amounts of energy, and minimal amounts
of space squeezing (distilling) everything

into INT8 and INT4.

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Boris the Loris and Julio Di Egidio the Nazi Retard,
are going for an afterwork beer. They are still
highly confused by Fuzzy Testing:

Star Trek - The 70's Disco Generation https://www.youtube.com/watch?v=505zvAvnreg

The favorite hangout is Spock's Logic Dancefloor,
which is known for its sharp unfuzzy wit. They
have a chat with Data about Disco Math,

the only Math which has no Fuzzy Logic in it.

Bye

Mild Shock schrieb:

Hi,

Candidate Recommendation Draft - 30 September 2025 https://www.w3.org/TR/webnn

WebNN samples by Ningxin Hu, Intel, Shanghai https://github.com/webmachinelearning/webnn-samples

Bye

Mild Shock schrieb:

Hi,

It seems I am having problems pacing with
all the new fancy toys. Wasn't able to really
benchmark my NPU from a Desktop AI machine,

picked the wrong driver. Need to try again.
What worked was benchmarking Mobile AI machines.
I just grabbed Geekbench AI and some devices:

USA Fab, M4:

     sANN    hANN    qANN
iPad CPU    4848    7947    6353
iPad GPU    9752    11383    10051
iPad NPU    4873    36544    *51634*

China Fab, Snapdragon:

     sANN    hANN    qANN
Redmi CPU    1044    950    1723
Redmi GPU    480    905    737
Redmi NNAPI    205    205    469
Redmi QNN    226    226    *10221*

Speed-Up via NPU is factor 10x. See the column
qANN which means quantizised artificial neural
networks, when NPU or QNN is picked.

The mobile AI NPUs are optimized using
mimimal amounts of energy, and minimal amounts
of space squeezing (distilling) everything

into INT8 and INT4.

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2