[Bug-gnubg] Re: Bug-gnubg Digest, Vol 48, Issue 6

[Frank Berger]

> Message: 10
> Date: Mon, 06 Nov 2006 14:27:50 +0100
> From: ? ystein Johansen<address@hidden>
> Subject: Re: [Bug-gnubg] 50% decrease in evaluation speed

> There was also a suggestion of moving the whole neural net
> evaluation over to the GPU to let the strong processor in
> a modern computer do the hard work. The neural net evaluation
> is basically just a matrix multiplication and GPUs are just
> made for that. No work has been done thought.
There is a master thesis(?) realted to neural nets on GPU's. didnt  
was that promising. The setup is so expensive, that it has only a  
small gain with very large nets. O.k. This article is 1-2 years old  
so the circumstances may have changed.



> So, should we maybe make another benchmark test?
> A test that tests all the neural nets, all the databases, and
> all input calculations and so on. I guess we'll have to make
> set of positions for such a benchmark, and these positions
> should then not be selected at random, as todays speed
> evaluation.

Good idea. Forward the positions to me and I will come up with the  
same test.

I suggest a small set of positions:
3 contact
1 running because running is roughly 25% of the positions

or

6 contact
1 running
1 Bear off db... where with the later it depends a lot on the size of  
the DB
so the first makes more appropriate to me.

IMHo it makes sense to differentiate with the ply 0/1/2 (or 1/2/3- 
rest of the world counting)

Caching should be turned of IMHO

ciao
Frank