16 Core (E5-2687W) Matches and Tournaments

[ouachita]

Also Bobby, I wonder if your Fritz bench can use all of your cores, some of the Fritz bench tests were limited in core usage.thanks
kgburcham

It runs a bit better on 15.

[ouachita]

I just ran the Autotune for Houdini 4 and my best min split depth is 12 same as with Houdini 3.

H4 SD default is 10; mine runs about 8% faster at 14. SF doesn't have "autotune", but does it have something similar?

[ouachita]

If you are using the default, then I believe you're using Min Split Depth=7 (which might be sub-optimal for 16 threads).

In my tournaments I leave it to zero, so I think the engine will choose SD7 as default value. Clemens

Where did you find SD 7 as the default?

[zullil]

If you are using the default, then I believe you're using Min Split Depth=7 (which might be sub-optimal for 16 threads).

In my tournaments I leave it to zero, so I think the engine will choose SD7 as default value. Clemens

Where did you find SD 7 as the default?

thread.cpp --- lines 207-226:

Code: Select all

// read_uci_options() updates internal threads parameters from the corresponding
// UCI options and creates/destroys threads to match the requested number. Thread
// objects are dynamically allocated to avoid creating all possible threads
// in advance (which include pawns and material tables), even if only a few
// are to be used.

void ThreadPool::read_uci_options() {

  maxThreadsPerSplitPoint = Options["Max Threads per Split Point"];
  minimumSplitDepth       = Options["Min Split Depth"] * ONE_PLY;
  size_t requested        = Options["Threads"];

  assert(requested > 0);

  // Value 0 has a special meaning: We determine the optimal minimum split depth
  // automatically. Anyhow the minimumSplitDepth should never be under 4 plies.
  if (!minimumSplitDepth)
      minimumSplitDepth = &#40;requested < 8 ? 4 &#58; 7&#41; * ONE_PLY;
  else
      minimumSplitDepth = std&#58;&#58;max&#40;4 * ONE_PLY, minimumSplitDepth&#41;;

[ouachita]

thread.cpp --- lines 207-226:

It varies by engine, but if we're just looking for speed or kN/s, SD 10-12 is best on my machine.

[zullil]

thread.cpp --- lines 207-226:

It varies by engine, but if we're just looking for speed or kN/s, SD 10-12 is best on my machine.

I assumed you were asking about Stockfish, so I posted the relevant source code for Stockfish. If you leave the setting for Min Split Depth as 0, SF will set it to 7 if the setting for Threads is 8 or more.

Based on your experimenting, it looks like that's suboptimal for 16 threads.

[ouachita]

I assumed you were asking about Stockfish, . . . it looks like that's suboptimal for 16 threads.

I meant to say, it varies by SF version, and, yes, the fastest SD for mine is 10-12, depending. Thanks for pointing out that default 7.

[ouachita]

SF010214-16 v. H4B-16, Blitz 1m+1s 0

Code: Select all

                                      
1   Houdini 4 Pro x64B           +28  +29/=50/-21 54.00%   54.0/100
2   Stockfish 020114 64 SSE4.2   -28  +21/=50/-29 46.00%   46.0/100

SF010214-32T v. H4B-16C, 1+1  0
                                      
1   Houdini 4 Pro x64B            +7  +21/=60/-19 51.00%   51.0/100
2   Stockfish 020114 64 SSE4.2    -7  +19/=60/-21 49.00%   49.0/100

The single difference between these two test bases, aside from kPa and RH, was changing SF to 32 threads.

It is what it is.

[ouachita]

Code: Select all

Physical cores v. Physical Cores

SF010214-16 v. H4B-1, Blitz 1m+1s  0
                                      
1   Stockfish 020114 64 SSE4.2  +174  +41/=35/-4 73.13%   58.5/80
2   Houdini 4 Pro x64B          -174  +4/=35/-41 26.88%   21.5/80

SF010214-1 v. H4B-16, Blitz 1m+1s  0

                                      
1   Houdini 4 Pro x64B          +205  +47/=30/-4 76.54%   62.0/81
2   Stockfish 020114 64 SSE4.2  -205  +4/=30/-47 23.46%   19.0/81

Run XXXX games with these bases, the answer should be basically the same. I do not have the data to plot the relationship between decreasing core deltas (1-16=15, 2-16=14, etc.) and decreasing ELO deltas, but the direction of the trend lines are obvious.

8C v. 16C is running now.

[Vinvin]

Run XXXX games with these bases, the answer should be basically the same. I do not have the data to plot the relationship between decreasing core deltas (1-16=15, 2-16=14, etc.) and decreasing ELO deltas, but the direction of the trend lines are obvious.

Some old rules :
1) Multiply speed by 2 : +60-70 elo.
2) 1 core to 2 cores (that's not the same as multiply speed by 2 because the search have to be split somewhere) : +50-60 elo
3) 2 cores to 4 cores : a bit less than previous -> +45-50?
4) 4 cores to 8 cores : a bit less than previous ... ?
5) 8 cores to 16 cores : a bit less than previous ... ?

8C v. 16C is running now.

That will be interesting to know !
Thanks !