The effective speedup from 1 to 8 cpus for SF and Komodo

[Adam Hair]

I have completed my measurement of the effective speedup from 1 to 8 cores for Stockfish 230315 and Komodo 9, where effective speedup = time odds for 1 core version to be equal in strength to 8 cores version.

(My computer - dual L5420 Xeon processors @ 2.5 GHz)

SF 230315 8 cores used a tc of 60"+0.05", 1 core used 240" + 0.2" = 4 * 8 core tc

Code: Select all

   # PLAYER              : RATING  ERROR   POINTS  PLAYED    (%)
   1 SF 230315 1core     :    0.0    3.6   1000.0    2000   50.0%
   2 SF 230315 8cores    :    0.0    3.6   1000.0    2000   50.0%

White advantage = 45.11 +/- 3.77
Draw rate (equal opponents) = 80.00 % +/- 0.97

Komodo 9 8 cores used a tc of 84" + 0.07". At this tc on my computer, Komodo 9 8 cores is roughly equal in strength to SF 230315 8 cores @ 60" + 0.05":

Code: Select all

   # PLAYER              : RATING  ERROR   POINTS  PLAYED    (%)
   1 Komodo 9 8cores     :    1.7    9.5    257.5     510   50.5%
   2 SF 230315 8cores    :   -1.7    9.5    252.5     510   49.5%

White advantage = 44.92 +/- 9.15
Draw rate (equal opponents) = 63.15 % +/- 2.11

The results for Komodo 9 8 cores @ 84" + 0.7" versus 1 core @ 378" + 0.315" (4.5 * 8 core tc):

Code: Select all

   # PLAYER             : RATING  ERROR   POINTS  PLAYED    (%)
   1 Komodo 9 8cores    :    4.6    3.5   1026.0    2000   51.3%
   2 Komodo 9 1core     :   -4.6    3.5    974.0    2000   48.7%

White advantage = 38.03 +/- 3.60
Draw rate (equal opponents) = 81.92 % +/- 0.93

Under these conditions, the effective speedup from 1 to 8 cores for SF 230315 is approximately 4, while for K9 it is probably greater than 4.5.

I have an incomplete test for SF 6 that seems to indicate its effective speedup is approximately 3.6.

Link to the pgns: http://www.mediafire.com/download/3irtt ... speedup.7z

[mjlef]

Adam,

Great stuff. I would love to see how this continues with more cores. It would take a lot of computer time.

Mark

[Terry McCracken]

Good work, thanks Adam!

[Laskos]

Thanks, very time and CPU consuming test.

The central value for Komodo 9 effective speedup is a bit larger than 4.5, having 9 Elo points difference. The doubling at 40/4' is worth about 70 Elo points, so in your case about 70-80 Elo points, and the effective speedup must be multiplied by 2^(9/70 to 80) for a total of ~4.9 for Komodo 9 on 8 cores. Therefore,
4.0 Stockfish 230315
4.9 Komodo 9
going from 1 to 8 cores, at blitz.
Error margins are harder to compute.

These are difficult to misinterpret, important numbers. As Stockfish shows hardly any widening, maybe in its case time to depth is the easier way to compute the effective speedup. Komodo does widen, so playing games is the only way.

[Werewolf]

8^0.76 = 4.86

very close to theoretical in Komodo's case.

[lkaufman]

8^0.76 = 4.86

very close to theoretical in Komodo's case.

What is the significance of the number 0.76 above? Where does it come from?

[Werewolf]

8^0.76 = 4.86

very close to theoretical in Komodo's case.

What is the significance of the number 0.76 above? Where does it come from?

It's a number I've always used handed down by the Rybka team, LC in particular.

You didn't come across this when you were working with them?

[bob]

8^0.76 = 4.86

very close to theoretical in Komodo's case.

What is the significance of the number 0.76 above? Where does it come from?

It's a number I've always used handed down by the Rybka team, LC in particular.

You didn't come across this when you were working with them?

I would not call it "theoretical" however. That implies some sort of global truth. That is purely a number that worked for Rybka.

[lkaufman]

8^0.76 = 4.86

very close to theoretical in Komodo's case.

What is the significance of the number 0.76 above? Where does it come from?

It's a number I've always used handed down by the Rybka team, LC in particular.

You didn't come across this when you were working with them?

If so I don't recall it now. It seems too simplistic, since each doubling is worth less than the one before.

[Sean Evans]

8^0.76 = 4.86

very close to theoretical in Komodo's case.

What is the significance of the number 0.76 above? Where does it come from?

It's a number I've always used handed down by the Rybka team, LC in particular.

You didn't come across this when you were working with them?

If so I don't recall it now. It seems too simplistic, since each doubling is worth less than the one before.

It represents a point of diminishing returns.