Stockfish NNUE speed is improved a lot

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Dann Corbit, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
Jouni
Posts: 2387
Joined: Wed Mar 08, 2006 7:15 pm

Stockfish NNUE speed is improved a lot

Post by Jouni » Mon Jan 11, 2021 3:24 pm

GitHub says: On CPUs supporting modern vector instructions (avx2 and similar), the NNUE evaluation results in stronger playing strength, even if the nodes per second computed by the engine is somewhat lower (roughly 60% of nps is typical). I also remember first NNUE versions speed was about 65-70% of classic. I tested 8.1.2021 version (bmi) with bench. Now NNUE runs with about 88% of classic speed :!: . Nice speed-up and optimization. Soon 100%?
Jouni

Raphexon
Posts: 368
Joined: Sun Mar 17, 2019 11:00 am
Full name: Henk Drost

Re: Stockfish NNUE speed is improved a lot

Post by Raphexon » Mon Jan 11, 2021 10:03 pm

Because in the past NNUE was pure and now it's hybrid.

Most of the speedup comes from classical eval being used on some nodes.

Jouni
Posts: 2387
Joined: Wed Mar 08, 2006 7:15 pm

Re: Stockfish NNUE speed is improved a lot

Post by Jouni » Mon Jan 11, 2021 10:08 pm

If I use evalType == "NNUE" for bench still classical eval?
Jouni

RogerC
Posts: 35
Joined: Tue Oct 29, 2019 7:33 pm
Location: French Polynesia
Full name: Roger C.

Re: Stockfish NNUE speed is improved a lot

Post by RogerC » Tue Jan 12, 2021 12:11 am

Jouni wrote:
Mon Jan 11, 2021 3:24 pm
GitHub says: On CPUs supporting modern vector instructions (avx2 and similar), the NNUE evaluation results in stronger playing strength, even if the nodes per second computed by the engine is somewhat lower (roughly 60% of nps is typical). I also remember first NNUE versions speed was about 65-70% of classic. I tested 8.1.2021 version (bmi) with bench. Now NNUE runs with about 88% of classic speed :!: . Nice speed-up and optimization. Soon 100%?
Hi, i have a 102% speed between pure NNUE vs hybrid on SD865+ processor on last Cfish and 8 cores.
Optimizations pushed the pure eval to be faster on heavy vector processors + big cache levels. Should be the same on Intel i9.

Post Reply