16 vs 32 threads - Stockfish 180104

[fastgm]

Conditions:
Dual AMD Opteron 6376, Windows 10 Professional 64-Bit
Cutechess-Cli, Stockfish 180104, tc = 10 min + 6 sec, 4096 MB Hash, opening book: Noomen_2-move_Testsuite.pgn, changing colors, adjudication rules, no tablebases

First result after 80 games:

Code: Select all

1   Stockfish 180104 T32   +22  +5/=75/-0 53.13%   42.5/80
2   Stockfish 180104 T16   -22  +0/=75/-5 46.88%   37.5/80

 Engine                 Depth       Time   Games     Moves  Average
--------------------------------------------------------------------
 Stockfish 180104 T32   38.46   20:21:11      80      5259    13.93
 Stockfish 180104 T16   37.82   20:26:05      80      5266    13.97

Depth     : Average search depth
Time      : Total time engine used
Moves     : Total moves engine played
Average   : Average time per move in centi-seconds

.

[Laskos]

Conditions:
Dual AMD Opteron 6376, Windows 10 Professional 64-Bit
Cutechess-Cli, Stockfish 180104, tc = 10 min + 6 sec, 4096 MB Hash, opening book: Noomen_2-move_Testsuite.pgn, changing colors, adjudication rules, no tablebases

First result after 80 games:

Code: Select all

1   Stockfish 180104 T32   +22  +5/=75/-0 53.13%   42.5/80
2   Stockfish 180104 T16   -22  +0/=75/-5 46.88%   37.5/80

 Engine                 Depth       Time   Games     Moves  Average
--------------------------------------------------------------------
 Stockfish 180104 T32   38.46   20:21:11      80      5259    13.93
 Stockfish 180104 T16   37.82   20:26:05      80      5266    13.97

Depth     : Average search depth
Time      : Total time engine used
Moves     : Total moves engine played
Average   : Average time per move in centi-seconds

.

Thanks Andreas, interesting as always. How many NUMA nodes this 32 core thing has? The main result for me here is 0.64 plies depth difference, it is probably a more reliable and easier to quantify variable here than Elo. If EBF of SF for Lazy SMP on so many cores is about 1.5-1.6, then effective speedup would be in 1.3-1.4 range from 16 -> 32 cores on probably double NUMA nodes.

[fastgm]

Thanks Andreas, interesting as always. How many NUMA nodes this 32 core thing has? The main result for me here is 0.64 plies depth difference, it is probably a more reliable and easier to quantify variable here than Elo. If EBF of SF for Lazy SMP on so many cores is about 1.5-1.6, then effective speedup would be in 1.3-1.4 range from 16 -> 32 cores on probably double NUMA nodes.

The workstation has 4 NUMA nodes.
I would like to play at least 500 games. Then the result is statistically more meaningful.

[cdani]

Thanks Andreas, interesting as always. How many NUMA nodes this 32 core thing has? The main result for me here is 0.64 plies depth difference, it is probably a more reliable and easier to quantify variable here than Elo. If EBF of SF for Lazy SMP on so many cores is about 1.5-1.6, then effective speedup would be in 1.3-1.4 range from 16 -> 32 cores on probably double NUMA nodes.

Will be nice to know the number of nodes visited also. I know is not something that is stored normally, was just a way of telling that other than the 0.64 extra plies, the widening also accounts for a big % of the strength increase.

[Laskos]

Thanks Andreas, interesting as always. How many NUMA nodes this 32 core thing has? The main result for me here is 0.64 plies depth difference, it is probably a more reliable and easier to quantify variable here than Elo. If EBF of SF for Lazy SMP on so many cores is about 1.5-1.6, then effective speedup would be in 1.3-1.4 range from 16 -> 32 cores on probably double NUMA nodes.

Will be nice to know the number of nodes visited also. I know is not something that is stored normally, was just a way of telling that other than the 0.64 extra plies, the widening also accounts for a big % of the strength increase.

Right, I tried some sort of convoluted and confusing thinking while writing the post. I wrote EBF 1.5-1.6 going from 16 to 32 cores, which EBF includes widening. In fact EBF of SF is about 1.4-1.5 on the same number of cores. And widening may account for 30% of strength improvement, deepening 70%. Or something like that. So, 0.64 plies is maybe about 1.3 effective speed-up time to strength from deepening. And that means 16->32 cores gives about 1.5 total effective speed-up time to strength from deepening + widening. That is all very speculative. Could be even 1.7.

[fastgm]

Will be nice to know the number of nodes visited also.

I can determine the average node numbers by playing a few additional games with the LittleBlitzer GUI.

[fauzi]

Can you please share the games after the test is finished?

[fastgm]

After 150 games:

Code: Select all

1   Stockfish 180104 T32   +21  +10/=139/-1 53.00%   79.5/150
2   Stockfish 180104 T16   -21  +1/=139/-10 47.00%   70.5/150

.

[fastgm]

Can you please share the games after the test is finished?

Yes, I'll do that.

[cdani]

Will be nice to know the number of nodes visited also.

I can determine the average node numbers by playing a few additional games with the LittleBlitzer GUI.

As you wish. Will be nice Thanks