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


Groups > sci.physics > #894449 > unrolled thread

NPUs (Neural Processing Units) are the new normal

Started byMild Shock <janburse@fastmail.fm>
First post2025-10-15 16:15 +0200
Last post2025-11-28 15:13 +0100
Articles 4 — 1 participant

Back to article view | Back to sci.physics


Contents

  NPUs (Neural Processing Units) are the new normal Mild Shock <janburse@fastmail.fm> - 2025-10-15 16:15 +0200
    AI most hated by formal verification (Re: Fuzzy Alert: Boris the Loris on the Dancefloor) Mild Shock <janburse@fastmail.fm> - 2025-11-26 17:03 +0100
      Googles TPU muscle in 2017 [Prolog Community is Sleepy Joe] (Was: Zeus: A Language for Expressing Algorithms in Hardware) Mild Shock <janburse@fastmail.fm> - 2025-11-27 15:22 +0100
      The tables have turned: GigaLIP on the Laptop? (Re: 100% serious Giga Logical Inferences per Second (GLIPS)) Mild Shock <janburse@fastmail.fm> - 2025-11-28 15:13 +0100

#894449 — NPUs (Neural Processing Units) are the new normal

FromMild Shock <janburse@fastmail.fm>
Date2025-10-15 16:15 +0200
SubjectNPUs (Neural Processing Units) are the new normal
Message-ID<10coa9b$12ucf$6@solani.org>
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

[toc] | [next] | [standalone]


#894568 — AI most hated by formal verification (Re: Fuzzy Alert: Boris the Loris on the Dancefloor)

FromMild Shock <janburse@fastmail.fm>
Date2025-11-26 17:03 +0100
SubjectAI most hated by formal verification (Re: Fuzzy Alert: Boris the Loris on the Dancefloor)
Message-ID<10g78cv$l82g$2@solani.org>
In reply to#894449
Hi,

So Boris the Loris and Nazi Retartd Julio are
not alone. There is now a mobilization of the
kind of rage against the machine,

fighting for methods without randomness. Its
almost like  Albert Einstein ascendet from his
grave and is now preaching,

"God does not play dice"

So how it started:

PIVOT was an interactive program verifier designed by
L. Peter Deutsch for his Ph.D. dissertation.
Posted here by permission of L. Peter Deutsch.
https://softwarepreservation.computerhistory.org/pivot/

How its going:

Formal Methods: Whence and Whither?
The text also highlights the evolving role of formal
methods amidst technological advancements, such as
AI, and explores educational and standardization issues
related to their adoption.
https://de.slideshare.net/slideshow/formal-methods-whence-and-whither-keynote/273708245

Can the Don Quijotes win, and fight the AI windmills?

LoL

Bye

Mild Shock schrieb:
> 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
>>
> 

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


#894570 — Googles TPU muscle in 2017 [Prolog Community is Sleepy Joe] (Was: Zeus: A Language for Expressing Algorithms in Hardware)

FromMild Shock <janburse@fastmail.fm>
Date2025-11-27 15:22 +0100
SubjectGoogles TPU muscle in 2017 [Prolog Community is Sleepy Joe] (Was: Zeus: A Language for Expressing Algorithms in Hardware)
Message-ID<10g9mr8$oiad$1@solani.org>
In reply to#894568
Hi,

Well I am currently looking in Local AI when I
consider NPUs, which have very small floats.
An example of bigger AI accelerators are TPUs,

which can deal with larger floats. Subsequently
then can also deal with a larger integer range.
I only read this anecdote yesterday:

"In December 2017, Stockfish 8 was used as a
benchmark to test Google division DeepMind's
AlphaZero, with Stockfish running on CPU and
AlphaZero running on Google's proprietary
Tensor Processing Units (TPUs).

AlphaZero was trained through self-play for
a total of nine hours, and reached Stockfish's
level after just four. AlphaZero also played
twelve 100-game matches against Stockfish starting
from twelve popular openings for a final score
of 290 wins, 886 draws and 24 losses, for a
point score of 733:467."
https://en.wikipedia.org/wiki/Stockfish_(chess)#Stockfish_8_versus_AlphaZero

And then:

"AlphaZero's victory over Stockfish sparked a
flurry of activity in the computer chess community,
leading to a new open-source engine aimed at
replicating AlphaZero, known as Leela Chess Zero.
The two engines remained close in strength for a
while, but Stockfish has pulled away since the
introduction of NNUE, winning every TCEC
season since Season 18."

Meanwhile the Prolog community: Sleepy Joe

LoL

Bye

Mild Shock schrieb:
> Hi,
> 
> What mindset is needed to program an NPU. Mostlikely
> a mindset based on fork/join parallelism is nonsense.
> What could be more fruitful is view the AI accellerator
> 
> as a blackbox that runs a neural network, whereby
> a neural network can be effectively viewed as a form
> of hardware, although unter the hood, it is open weights
> 
> and matrix operations. So the mindest needs:
> 
> Zeus: A Language for Expressing Algorithms in Hardware
> K. J. Lieberherr -  01 February 1985
> https://dl.acm.org/doi/10.1109/MC.1985.1662799
> 
> What changed back to then?
> 
> - 80's Field Programmable Gate Array (FPGA)
> 
> - 20's AI Boom: NPUs, Unified Memory and Routing Fabric
> 
> Bye
> 
> Mild Shock schrieb:
>> Hi,
>>
>> I already posted how to do SAT and Clark Completion
>> with ReLU. This was a post from 15.03.2025, 16:13,
>> see also below. But can we do CLP as well? Here
>>
>> is a take on the dif/2 constraint, or more precisely
>> a very primitive (#\=)/2 from CLP(FD), going towards
>> analogical computing. Might work for domains that
>>
>> fit into the quantization size of a NPU:
>>
>> 1) First note that we can model abs() via ReLU:
>>
>> abs(x) = ReLU(x) + ReLU(- x)
>>
>> 2) Then note that for integer values, we can model
>> chi(x>0), the characteristic function of the predicate x > 0:
>>
>> chi(x>0) = 1 - ReLU(1 - x).
>>
>> 3) Now chi(x=\=y) is simply:
>>
>> chi(x=\=y) = chi(abs(x - y) > 0)
>>
>> Now insert the formula for chi(x>0) based on ReLU
>> and the formula for abs() based on ReLU. Eh voila you
>> got an manually created neural network for the
>>
>> (#\=)/2 condition of CLP(FD), constraint logic
>> programming for finite domains.
>>
>> Have Fun!
>>
>> Bye
>>
>> Mild Shock schrieb:
>>  > A storm of symbolic differentiation libraries
>>  > was posted. But what can these Prolog code
>>  > fossils do?
>>  >
>>  > Does one of these libraries support Python symbolic
>>  > Pieceweise ? For example one can define rectified
>>  > linear unit (ReLU) with it:
>>  >
>>  >                  /   x      x  >= 0
>>  >      ReLU(x) := <
>>  >                  \   0      otherwise
>>  >
>>  > With the above one can already translate a
>>  > propositional logic program, that uses negation
>>  > as failure, into a neural network:
>>  >
>>  > NOT     \+ p             1 - x
>>  > AND     p1, ..., pn      ReLU(x1 + ... + xn - (n-1))
>>  > OR      p1; ...; pn      1 - ReLU(-x1 - .. - xn + 1)
>>  >
>>  > For clauses just use Clark Completion, it makes
>>  > the defined predicate a new neuron, dependent on
>>  > other predicate neurons,
>>  >
>>  > through a network of intermediate neurons. Because
>>  > of the constant shift in AND and OR, the neurons
>>  > will have a bias b.
>>  >
>>  > So rule based in zero order logic is a subset
>>  > of neural network.
>>  >
>>  > Python symbolic Pieceweise
>>  > 
>> https://how-to-data.org/how-to-write-a-piecewise-defined-function-in-python-using-sympy/ 
>>
>>  >
>>  >
>>  > rectified linear unit (ReLU)
>>  > https://en.wikipedia.org/wiki/Rectifier_(neural_networks)
>>  >
>>  > Clark Completion
>>  > https://www.cs.utexas.edu/~vl/teaching/lbai/completion.pdf
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>> I am spekulating an NPU could give 1000x more LIPS.
>>> For certain combinatorial search problems. It all
>>> boils down to implement this thingy:
>>>
>>> In June 2020, Stockfish introduced the efficiently
>>> updatable neural network (NNUE) approach, based
>>> on earlier work by computer shogi programmers
>>> https://en.wikipedia.org/wiki/Stockfish_%28chess%29
>>>
>>> There are varying degrees what gets updated of
>>> a neural network. But the specs of an NPU tell
>>> me very simply the following:
>>>
>>> - An NPU can make 40 TFLOPS, all my AI Laptops
>>>    from 2025 can do that right now. The brands
>>>    are Intel Ultra, AMD Ryzen and Snapdragon X,
>>>
>>>    but I guess there might be more brands around,
>>>    which can do that with a price tag less
>>>    than 1000.- USD.
>>>
>>> - SWI Prolog can make 30 MLIPS, Dogelog Player
>>>    runs similar, some Prolog systems are faster.
>>>
>>> Now thats is 10^12 versus 10^6. If some of the
>>> LIPS can be delegated to a NPU, and if we assume
>>> for example less locality or more primitive
>>>
>>> operations that require a layering. Would could assume
>>> that from the NPU 10^12 a factor of 1000 goes
>>> away. So we might still see 10'9 LIPS emerge.
>>>
>>> Now make the calculation:
>>>
>>> - Without NPU: MLIPS
>>> - With NPU: GLIPS
>>> - Ratio: 1000x times faster
>>>
>>> Have fun!
>>>
>>> Bye
>>>
>>> Mild Shock schrieb:
>>>> Hi,
>>>>
>>>> So Boris the Loris and Nazi Retartd Julio are
>>>> not alone. There is now a mobilization of the
>>>> kind of rage against the machine,
>>>>
>>>> fighting for methods without randomness. Its
>>>> almost like  Albert Einstein ascendet from his
>>>> grave and is now preaching,
>>>>
>>>> "God does not play dice"
>>>>
>>>> So how it started:
>>>>
>>>> PIVOT was an interactive program verifier designed by
>>>> L. Peter Deutsch for his Ph.D. dissertation.
>>>> Posted here by permission of L. Peter Deutsch.
>>>> https://softwarepreservation.computerhistory.org/pivot/
>>>>
>>>> How its going:
>>>>
>>>> Formal Methods: Whence and Whither?
>>>> The text also highlights the evolving role of formal
>>>> methods amidst technological advancements, such as
>>>> AI, and explores educational and standardization issues
>>>> related to their adoption.
>>>> https://de.slideshare.net/slideshow/formal-methods-whence-and-whither-keynote/273708245 
>>>>
>>>>
>>>> Can the Don Quijotes win, and fight the AI windmills?
>>>>
>>>> LoL
>>>>
>>>> Bye
>>>>
>>>> Mild Shock schrieb:
>>>>> 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
>>>>>>
>>>>>
>>>>
>>>
>>
> 

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


#894571 — The tables have turned: GigaLIP on the Laptop? (Re: 100% serious Giga Logical Inferences per Second (GLIPS))

FromMild Shock <janburse@fastmail.fm>
Date2025-11-28 15:13 +0100
SubjectThe tables have turned: GigaLIP on the Laptop? (Re: 100% serious Giga Logical Inferences per Second (GLIPS))
Message-ID<10gcam8$ofi3$3@solani.org>
In reply to#894568
Hi,

Notably we try to do something with AI laptops
that came out in 2025. Tapping into their NPU.
Back in the late 80's GigaLIP were rather

hypothetical, small machines could hardly
do KLIPs. Today small machines do easily MLIPs.
But hardly any popular Prolog systems already taps

into the AI Boom, they are all Sleepy Joes.
Not to mention populate by morons like Boris the
Loris and Nazi Retard Julio.

They simply cannot connect the dots.

Bye

P.S.: Nice travel to the past is this paper:

Is a GigaLIP Fast Enough?
TOM W. KELLER - December 1988
DOI: 10.1007/BF00436711

Mild Shock schrieb:
> 
> Hi,
> 
> I am 100% serious about Giga Logical Inferences
> per Second (GLIPS). Leaving behind the sequential
> constraint solving world:
> 
> The Complexity of Constraint Satisfaction Revisited
> https://www.cs.ubc.ca/~mack/Publications/AIP93.pdf
> 
> Only I have missed the deep learning bandwagon,
> never programmed with PyTorch or Keras. So even
> for the banal problem of coding some
> 
> ReLU networks and shipping them to a GPU or NPU,
> or a hybrid, I don't have much experience. So
> I am marveling at papers such as:
> 
> Learning Variable Ordering Heuristics
> for Solving Constraint Satisfaction Problems
> https://arxiv.org/abs/1912.10762
> 
> Given that the AI Boom started after 2019,
> the above paper is already old, and it has
> currious antique terminology like Multilayer
> 
> Perceptron, which is not so common anymore?
> It does also more than what I want to demonstrate,
> it does also do policy learning.
> 
> Bye
> 
> Mild Shock schrieb:
>> Hi,
>>
>> I am spekulating an NPU could give 1000x more LIPS.
>> For certain combinatorial search problems. It all
>> boils down to implement this thingy:
>>
>> In June 2020, Stockfish introduced the efficiently
>> updatable neural network (NNUE) approach, based
>> on earlier work by computer shogi programmers
>> https://en.wikipedia.org/wiki/Stockfish_%28chess%29
>>
>> There are varying degrees what gets updated of
>> a neural network. But the specs of an NPU tell
>> me very simply the following:
>>
>> - An NPU can make 40 TFLOPS, all my AI Laptops
>>    from 2025 can do that right now. The brands
>>    are Intel Ultra, AMD Ryzen and Snapdragon X,
>>
>>    but I guess there might be more brands around,
>>    which can do that with a price tag less
>>    than 1000.- USD.
>>
>> - SWI Prolog can make 30 MLIPS, Dogelog Player
>>    runs similar, some Prolog systems are faster.
>>
>> Now thats is 10^12 versus 10^6. If some of the
>> LIPS can be delegated to a NPU, and if we assume
>> for example less locality or more primitive
>>
>> operations that require a layering. Would could assume
>> that from the NPU 10^12 a factor of 1000 goes
>> away. So we might still see 10'9 LIPS emerge.
>>
>> Now make the calculation:
>>
>> - Without NPU: MLIPS
>> - With NPU: GLIPS
>> - Ratio: 1000x times faster
>>
>> Have fun!
>>
>> Bye
>>
>> Mild Shock schrieb:
>>> Hi,
>>>
>>> So Boris the Loris and Nazi Retartd Julio are
>>> not alone. There is now a mobilization of the
>>> kind of rage against the machine,
>>>
>>> fighting for methods without randomness. Its
>>> almost like  Albert Einstein ascendet from his
>>> grave and is now preaching,
>>>
>>> "God does not play dice"
>>>
>>> So how it started:
>>>
>>> PIVOT was an interactive program verifier designed by
>>> L. Peter Deutsch for his Ph.D. dissertation.
>>> Posted here by permission of L. Peter Deutsch.
>>> https://softwarepreservation.computerhistory.org/pivot/
>>>
>>> How its going:
>>>
>>> Formal Methods: Whence and Whither?
>>> The text also highlights the evolving role of formal
>>> methods amidst technological advancements, such as
>>> AI, and explores educational and standardization issues
>>> related to their adoption.
>>> https://de.slideshare.net/slideshow/formal-methods-whence-and-whither-keynote/273708245 
>>>
>>>
>>> Can the Don Quijotes win, and fight the AI windmills?
>>>
>>> LoL
>>>
>>> Bye
>>>
>>> Mild Shock schrieb:
>>>> 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
>>>>>
>>>>
>>>
>>
> 

[toc] | [prev] | [standalone]


Back to top | Article view | sci.physics


csiph-web