TalkChess.com

[Srdja Matovic]

hi,

last time i asked for an possible Board Presentation and Move Generation suited for GPUs,
these two topics are solved by use of Quad-Bitboards and a Magic Bitboard Move Generator (thanks to Gerd Isenberg and the Stockfish Team).

Now its up on the Search Algorithm which is required to feed thousands of threads.

Monte Carlo Tree Search would be easy to implement, but my simple implementation on a cpu shows random game play without any statistical significance. UCT may be a way to improve the results, but i don't think that this is an easy part to port on the GPU.

Parallel AlphaBeta Search like YBWC or DTS need, from GPU-view, sophisticated ways for communication and i am not sure how they will perform with thousands of threads.

Plain MiniMax could be another option to get chess running on a gpu, but i am not sure how to distribute work across threads. A Elimination Backoff Stack could be a solution...

Looking for some ideas or inspiration,
Srdja