TalkChess.com

Posted: **Sun Mar 31, 2024 9:35 pm**

When testing different Berserk versions I was surprised how much avx2 can give:

Score of berserk-12.1-avx2-pext vs berserk-12.1-ssse3: 105 - 69 - 226 [0.545]
...      berserk-12.1-avx2-pext playing White: 105 - 0 - 95  [0.762] 200
...      berserk-12.1-avx2-pext playing Black: 0 - 69 - 131  [0.328] 200
...      White vs Black: 174 - 0 - 226  [0.718] 400
Elo difference: 31.4 +/- 22.4, LOS: 99.7 %, DrawRatio: 56.5 %
400 of 400 games finished.

Nps speed up is 27%. I was expecting minor gain at this level only.

Posted: **Sun Mar 31, 2024 11:47 pm**

Jouni wrote: ↑Sun Mar 31, 2024 9:35 pm When testing different Berserk versions I was surprised how much avx2 can give:
Code: Select all
Score of berserk-12.1-avx2-pext vs berserk-12.1-ssse3: 105 - 69 - 226 [0.545]
...      berserk-12.1-avx2-pext playing White: 105 - 0 - 95  [0.762] 200
...      berserk-12.1-avx2-pext playing Black: 0 - 69 - 131  [0.328] 200
...      White vs Black: 174 - 0 - 226  [0.718] 400
Elo difference: 31.4 +/- 22.4, LOS: 99.7 %, DrawRatio: 56.5 %
400 of 400 games finished.
Nps speed up is 27%. I was expecting minor gain at this level only.

What is the time control?

I guess you used UHO book(with unbiased book it is certainly smaller gain).

Posted: **Mon Apr 01, 2024 2:08 pm**

Yes UHO book and 20+0,2 level. Similar test with SF16.1. Bmi compile is 25% faster than sse41 compile. So Berserk no exception.

Code: Select all

Score of stockfish16.1 vs stockfish-sse41: 116 - 83 - 201 [0.541]
...      stockfish16.1 playing White: 113 - 2 - 85  [0.777] 200
...      stockfish16.1 playing Black: 3 - 81 - 116  [0.305] 200
...      White vs Black: 194 - 5 - 201  [0.736] 400
Elo difference: 28.7 +/- 24.0, LOS: 99.0 %, DrawRatio: 50.2 %
400 of 400 games finished.

TalkChess.com

Surprise: avx2 compile gives +31 Elo to Berserk

Surprise: avx2 compile gives +31 Elo to Berserk

Re: Surprise: avx2 compile gives +31 Elo to Berserk

Re: Surprise: avx2 compile gives +31 Elo to Berserk