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

Groups > sci.logic > #342280

Benchmark results DirectML versus QNN [Challenge for Geekbench AI] (Re: AnythingLLM QNN/ONNX: Massiv Computations versus John Sowa)

Cross-posted to 3 groups.

Show all headers | View raw

Hi,

This super cute Snapdragon X box, has massive
benchmark score for quantisized neural networks (qANN):

     sCPU    mCPU    GPU    sANN    hANN    qANN
AcerSwift    2835    13393    25395    6744    10167    5175
YogaUltra    2785    9844    30545    7270    13936    4830
ThinkCentre    2145    9754    13782    1414    16456    40721

AcerSwift    OpenCL    DirectML
YogaUltra    OpenCL    DirectML
ThinkCentre    Vulkan    QNN

But maybe the low qANN numbers are a problem of
Geekbench AI, and how it uses DirectML, that it cannot
yet address teh full potential of the NPUs on the

other two Local AI machines. But impressively the
QNN API and the ONNX format, goes very smooth on
the ThinkCentre. For machine translation via quantisized

neural networks (qANN). I see that the ThinkCentre
is 10 times faster than the other two machines. But I
guess with a suitable version of Geekbench AI,

the gap between to the other two machines will close.
They are just too new, so that Geekbench AI is
lacking behind.

Bye

Mild Shock schrieb:
> Hi,
> 
> Ha Ha, remember this post on SWI-Prolog
> discourse, the primary source for morons such
> as Boris the Loris and Nazi Retard Julio:
> 
> "The idea that LLM-based methods can become
> more intelligent by using massive amounts
> of computation is false. They can generate
> more kinds of BS, but at an enormous cost in
> hardware and in the electricity to run that
> massive hardware. But without methods of
> evaluation, the probability that random mixtures
> of data are true or useful or worth the cost
> of generating them becomes less and less likely."
> - John Sowa, 10 Jul 2024
> https://swi-prolog.discourse.group/t/prolog-and-llms-genai/8699
> 
> Guess what my new ThinkCentre, that just arrived
> via Lenovo, China, with a Snapdragon X, for around
> 700.- USD could easily run locally some inferencing.
> 
> I was using AnythingLLM, it has little idioctic
> electron user interface, but can support
> Snapdragon X NPU and models, via QNN/ONNX:
> 
> The all-in-one AI application
> https://anythingllm.com/
> 
> Tested a LLama Model, a little bit chatty to
> be honest, and a Phi Silica model, not yet that
> good in coding. Where did the massive computation
> 
> come from? From the SOC and the unified memory
> of the Snapdragon. I had 32 GB, and 16 GB was
> shared with the NPU. So you don't need to
> 
> buy an Aura Yoga laptop, which has separate
> NVIDIA Graphics card, with only 8 GB. This
> graphic card will be useless, many interesting
> 
> models are above 8 GB. And yes the massive
> computation obviously leads to more intelligence.
> The later is a riddle for every Prologer, how
> 
> could more LIPS (logical inference per second)
> lead to more intelligence?
> 
> Bye
> 
> Mild Shock schrieb:
>> Hi,
>>
>> How it started:
>>
>> https://conceptbase.sourceforge.net/
>>
>> How its going:
>>
>> https://www.ibm.com/products/datastax
>>
>> The problem with claims such as " Formal languages,
>> such as KAOS, are based on predicate logic and
>> capture additional details about an application
>> in a precise manner. They also provide a foundation
>> for reasoning with information models." is that
>> every thing in the quoted sentence is wrong.
>>
>> Real AI systems scale by approximation,
>> vectorization, distributed representations,
>> and partial knowledge — not by globally
>> consistent logical models. No classical requirements
>> language or ontology captures the informal
>> cognitive machinery that makes
>> intelligence flexible. Intelligence needs the
>> whole messy cognitive spectrum.
>>
>> Somehow DataStax looks like n8n married AI embedding.
>> I hope Amazon, Meta, Google, etc.. get the message.
>> I don't worry about Microsoft, they might come with
>>
>> something from their Encarta corner and Copilot+ is
>> more Local AI. After all we need things like Wikidata
>> in a Robot and not in a Data Center.
>>
>> LoL
>>
>> Bye
> 

Back to sci.logic | Previous | Next — Previous in thread | Find similar | Unroll thread

Thread

Rene Descartes "Discours de la méthode" has fizzled out Mild Shock <janburse@fastmail.fm> - 2025-11-14 11:27 +0100
  Philosophical Twist due to negligence (Re: Rene Descartes "Discours de la méthode" has fizzled out) Mild Shock <janburse@fastmail.fm> - 2025-11-14 11:44 +0100
    Re: Philosophical Twist due to negligence (Re: Rene Descartes "Discours de la méthode" has fizzled out) Ross Finlayson <ross.a.finlayson@gmail.com> - 2025-11-14 11:10 -0800
      NY is the next Detroit in 2035 (Was: Philosophical Twist due to negligence) Mild Shock <janburse@fastmail.fm> - 2025-11-14 23:12 +0100
        Its a little sad story with NY (Was: NY is the next Detroit in 2035) Mild Shock <janburse@fastmail.fm> - 2025-11-14 23:25 +0100
          Re: Its a little sad story with NY (Was: NY is the next Detroit in 2035) Mild Shock <janburse@fastmail.fm> - 2025-11-14 23:45 +0100
    How to not be Artificial Intelligent [Boris the Loris deeply shocked] (Re: Philosophical Twist due to negligence) Mild Shock <janburse@fastmail.fm> - 2025-11-16 11:24 +0100
      Abstraction refinement (CEGAR) etc.. [Community Blind Spot] (Was: How to not be Artificial Intelligent) Mild Shock <janburse@fastmail.fm> - 2025-11-16 12:05 +0100
        The illusion of set theories [Computational Logic Primate] (Was: Abstraction refinement (CEGAR) etc.. [Community Blind Spot]) Mild Shock <janburse@fastmail.fm> - 2025-11-16 13:08 +0100
          Not all logicians are primarily interested in "computation" (Re: The illusion of set theories [Computational Logic Primate]) Mild Shock <janburse@fastmail.fm> - 2025-11-16 16:32 +0100
  AnythingLLM QNN/ONNX: Massiv Computations versus John Sowa (Re: Rene Descartes "Discours de la méthode" has fizzled out) Mild Shock <janburse@fastmail.fm> - 2025-11-25 20:07 +0100
    Benchmark results DirectML versus QNN [Challenge for Geekbench AI] (Re: AnythingLLM QNN/ONNX: Massiv Computations versus John Sowa) Mild Shock <janburse@fastmail.fm> - 2025-11-25 20:17 +0100

csiph-web