Stockfish with 16 threads --- big news?

[JJJ]

Just Kidding. How could you think I was serious ^^

[Laskos]

Looks like a patch submitted by Joona has generated a 50 point ELO gain for Stockfish running 16 threads. So perhaps a big step forward in Stockfish's multi-threaded searching. Seems very promising.

http://tests.stockfishchess.org/tests/v ... 3e9fac133b

I can get 16 Threads if I enable Hyper-Threading on my 8-Core i7 5960 X.
I use Windows 8.1.
Will my CPU also benefit ?

I have now a Haswell (4 physical cores).

1/ If not using HT threads, disable them from the BIOS or use "start /affinity 55 chessprogram.exe" (in my case, for 8 threads reduced to 4, it's 55), it might make up to 10% difference.
2/ From 8 to 16 threads (your case) it's very unlikely you will get something from HT, as Louis had only a 1.45 NPS speed-up on 16 core machine, while from 4 to 8 threads it's about 1.8 or so on 8 physical cores.
3/ Time-to-depth seems a good measure to compare the performance on many cores, as SF doesn't widen from 4 to 8 or from 8 to 16 threads. On my machine, on 150 opening-middlegame positions to depth 29 I got shorter time-to-depth with HT on using all 8 threads. You can do the same TTD test on your machine.

Engine: Stockfish 200215 64 BMI2 (2048 MB)
by Tord Romstad, Marco Costalba and Joona

4 threads:
TTD: 139:09 minutes

8 threads:
TTD: 116:14 minutes

Impressive is the EBF:
Last 7 plies:

Code: Select all

  Ply:22   Positions:150   Avg Nodes: 24794888   Branching = 1.55
  Ply:23   Positions:150   Avg Nodes: 37607067   Branching = 1.52
  Ply:24   Positions:150   Avg Nodes: 53597615   Branching = 1.43
  Ply:25   Positions:150   Avg Nodes: 83414641   Branching = 1.56
  Ply:26   Positions:150   Avg Nodes:127148481   Branching = 1.52
  Ply:27   Positions:150   Avg Nodes:181211891   Branching = 1.43
  Ply:28   Positions:150   Avg Nodes:258165515   Branching = 1.42

The average EBF for relevant plies: 1.49

[Jouni]

I was wrong 32 cpu improvement is only 222 ELO :

Code: Select all

    Program                       Elo    +   -   Games   Score   Av.Op.  Draws
 -------------------------------------------------------------------------------
  1 Stockfish 15021621 SMP T32  : 3111   35  34   250    78.2 %   2889   38.8 %
  2 Stockfish 15021602 T32      : 2889   34  35   250    21.8 %   3111   38.8 %


Result     : 195.5/250 (+147,=97,-6)
Perf.      : 78.2 %
Elo        : 3222

But note only for Intel Xeon E5-2666v3 @ 2.9 GHz with AMD Opteron +-0.

[Laskos]

I was wrong 32 cpu improvement is only 222 ELO :

Code: Select all

    Program                       Elo    +   -   Games   Score   Av.Op.  Draws
 -------------------------------------------------------------------------------
  1 Stockfish 15021621 SMP T32  : 3111   35  34   250    78.2 %   2889   38.8 %
  2 Stockfish 15021602 T32      : 2889   34  35   250    21.8 %   3111   38.8 %


Result     : 195.5/250 (+147,=97,-6)
Perf.      : 78.2 %
Elo        : 3222

But note only for Intel Xeon E5-2666v3 @ 2.9 GHz with AMD Opteron +-0.

Are you serious? What time control that was? You mean on Xeon it's 222 Elo points, on AMD 32 cores it's 0?

[Jouni]

I only copied Andreas Strangmüller's stunning results from CSS forum - he should have posted also here. Time control was 5" + 0.05". Longer time gave still "some" gain

Code: Select all

TC = 60" + 0.05"

    Program                       Elo    +   -   Games   Score   Av.Op.  Draws
 --------------------------------------------------------------------------------
  1 Stockfish 15021621 SMP T32  : 3063   43  38   100    67.5 %   2937   61.0 %
  2 Stockfish 15021602 T32      : 2937   38  43   100    32.5 %   3063   61.0 %

Result     : 67.5/100 (+37,=61,-2)
Perf.      : 67.5 %
Elo        : 3127