I ran the same experiment, here the results.petero2 wrote:It seems 1 is not a big problem for YBWC either. I ran some tests with stockfish 6 at time control 1s+0.08s/move, 4 cores vs 1 core:Joerg Oster wrote:At least lazy smp wouldn't suffer from 2 limitations as the current implementation.mar wrote:I'm not sure it would scale better, but certainly it would be interesting to compare.
1. takes some time to fully kick in because of the min split depth parameter
2. not enough workload especially for threads >= 8 due to heavy pruning and reductionsCode: Select all
Rank Name Elo + - games score oppo. draws 1 sf6_4c 70 7 7 2282 73% -70 46% 2 sf6 -70 7 7 2282 27% 70 46%
(I know, only 500 games each, nevertheless it gives an impression.)
2 Threads against 1, tc=1+0.08
Code: Select all
Rank Name Elo + - games score oppo. draws
1 SF6-2T 49 13 13 500 66% -49 44%
2 SF6-1T -49 13 13 500 34% 49 44%
Code: Select all
Rank Name Elo + - games score oppo. draws
1 SF6-4T 76 13 13 500 73% -76 38%
2 SF6-1T -76 13 13 500 27% 76 38%
Code: Select all
Rank Name Elo + - games score oppo. draws
1 SF6-8T 71 14 13 500 72% -71 39%
2 SF6-1T -71 13 14 500 28% 71 39%
No gain from 4 to 8 Threads. I know this is to be expected, but also a bit disappointing.
My idea to improve on this is to allow 2 threads to kick in earlier, and then continuously adding more threads with increasing depth. I don't know if this will work, but I think it is worth a try.