Replies: 2 comments 1 reply
-
That's clever. Should be strictly better indeed, with a cost of about 500kB of RAM per thread in the current network architecture. |
Beta Was this translation helpful? Give feedback.
1 reply
-
complete with 49ef4c9 |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
There is a possibility that we can speedup refreshes by using something called a "Finny table", introduced by Luecx (one of the authors of Koivisto). A relevant commit: AndyGrant/Ethereal@c5ba24d (note that this is in Ethereal instead)
The way this works is that we have an array of computed accumulators + bitboards indexed by bucket index. This way, when we require a refresh, we can simply update the difference from the computed accumulator rather than from scratch.
In addition, Andrew has been speaking quite a bit on potential NNUE speedups that Torch has but Stockfish doesn't, and there is a high chance that this is one of them.
So far, I have checked and it seems Stockfish doesn't have this idea. Maybe we can explore this? I am not sure how easy this is to implement.
cc @mstembera I think you might be interested :)
Beta Was this translation helpful? Give feedback.
All reactions