TalkChess.com

Don't you already have Zeta running most of the engine in the gpu?

But, It has been estimated that AGZ would take about 1,700 years of "typical" cpu/gpu time. Google burned about $25M.

There is a distributed effort for Go making quite good progress
https://github.com/gcp/leela-zero

GCP's prior work on "plain" Leela both in terms of playing and portability is most impressive.

Don't you already have Zeta running most of the engine in the gpu?

Yes, but in this design it was not intended to query a NN during search....
maybe i have to ponder on this a bit.

But, It has been estimated that AGZ would take about 1,700 years of "typical" cpu/gpu time. Google burned about $25M.

Ups

I just had the numbers of Giraffe in my mind, 3 days to one week for training on 12 core resp. 4 core machine.

There is a distributed effort for Go making quite good progress

Hmm,
maybe the chess community needs an commercial sponsor for the compute cycles for such an NN project.

--
Srdja

Quote:
Don't you already have Zeta running most of the engine in the gpu?

Yes, but in this design it was not intended to query a NN during search....
maybe i have to ponder on this a bit.

So far i have implemented parallel search and a dummy NN in Zeta,
but it turns out that plain GPU assist approach like LC0,
seems the way to go.

My parallel search scales bad above 32 workers,
so it would be more efficient to do the search on CPU,
and use the GPU for neural network evaluation only,
that would utilize both of them better.

--
Srdja

Rémi Coulom wrote:I took a longer look at what you did. Max pooling is used for image recognition, but makes little sense for chess. You'll need several layers of convolutions before reducing the resolution. I would simply stack a dozen 3x3 convolutions, and then directly go to a couple scalar layers, without any max pooling.

How do you arrive at this conclusion? Is there any rule of thumb how to design the NN architecture? What would you do in 8x8 american checkers?