WIP, Eta - GPGPU ANN based engine, RFC

[smatovic]

I don't know, maybe you have to touch the source code, maybe SF also does Root-Pruning which should also be disabled for such an run.

--
Srdja

This was done long ago in Aloril's old sheet. There was a hack by Tilps, I think, to have Leela pick moves based on value head assessment of all moves at root, effectively a d1 search using only value head.

Aloril tested this against an SF depth 1 search, and as I recall the best 64x6 net was a bit weaker than SF9 d1 and the best 128x10 was a bit better, so that size was where the crossover was.

The best 256x20 nets at the time were something like 500-600 points stronger with the d1 value head hack vs SF9 d1. The d1 value head hack with the strongest 256x20 was also much stronger than SF9 at 1k nodes per move, and dropped to something like -150 rating performance against SF9 at 10k nodes.

Of course, this was a very long time ago, early 2018, I think, so much has changed (and my memory's probably not 100% accurate).

Cheers!

Interesting. Thanks for sharing. Now I am really keen to see what I am able to tinker with this idea.

--
Srdja

[smatovic]

Just for the files, some basic benchmarks of the Zeta NNUE AB framework showed 100 Knps to max 200 Knps throughput per worker for current GPUs, definitely to slow for an Eta v0501 hybrid approach considering modern CPUs with ~2 Mnps per core.

--
Srdja