[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: For Mac OS X users [Was Re: [Bug-gnubg] gnubg on mac os x]

From: Joern Thyssen
Subject: Re: For Mac OS X users [Was Re: [Bug-gnubg] gnubg on mac os x]
Date: Wed, 9 Apr 2003 14:02:07 +0000
User-agent: Mutt/1.4i

On Wed, Apr 09, 2003 at 03:32:45PM +0200, Olivier Baur wrote
> Le mercredi, 9 avr 2003, à 10:09 Europe/Paris, Øystein O Johansen a 
> écrit :
> >Fantastic!
> >How is the speed? How many static evaluations can you get each sec?
> >(Analyse->Evaluation Speed->Calibrate)
> >I get about 17000 on my 2 GHz Intel XEON, and about 23000 on my 2.53 
> >GHz P4
> >
> >-Øystein
> On my PowerBook G4 768 MHz (no L3 cache :-)), I get about 9000 static 
> evals/sec. Quite poor compared to the numbers you give. I'm waiting for 
> someone to test it on a desktop G4 :-)

I get 11k evals/sec on a 1GHz laptop. 

I'm not very familiar with the G4 architecture, so I don't know what a
768MHz roughly corresponds to in Intel MHz.

Jim once did some profiling. I can't remember how much time is spend in
NeuralNetEvaluate versus CalculateInputs?!

> But I know I can get an increase to about 11000-12000 evals/sec by 
> "tweaking" some code (actually rewriting the eval algorithm to 
> superscalar with the G4's Velocity Engine) -- this was a quick try, not 
> released for now. My G4 code perf analyser also tells me there are many 
> things to correct about float usage in the eval, sigmoid and 
> calculateinputs functions -- things that dramatically matter on G4 
> performance.

I would very much like to see these changes; they may improve
performance for i386 as well.  If it's possible to rewrite the neural
net evaluation to calls to matrix or vector operations, then we can use
heavily optimised libraries that do exist for many platforms.


reply via email to

[Prev in Thread] Current Thread [Next in Thread]