Elo gain by core doubling - Komodo 14, Stockfish 11 - Contempt 0

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
fastgm
Posts: 575
Joined: Mon Aug 19, 2013 4:57 pm
Contact:

Elo gain by core doubling - Komodo 14, Stockfish 11 - Contempt 0

Post by fastgm » Sun Jun 28, 2020 5:52 am

AMD Ryzen Threadripper 3990X, 64 cores, 128 threads
Komodo 14 POPCNT vs Komodo 14 POPCNT, default, C0, 128 MB Hash, TC = 10 + 0.1 sec, 3000 games
Stockfish 11 POPCNT vs Stockfish 11 POPCNT, default, C0, 128 MB Hash, TC = 10 + 0.1 sec, 3000 games

Komodo 14

Code: Select all

Threads    2 vs 1          4 vs 2          8 vs 4          16 vs 8        32 vs 16        64 vs 32       128 vs 64
Elo          88              68              61               43             29              28              -1
Draw %      66.2            71.4            74.3             77.7           81.7            84.1            85.3
W-D-L   879-1986-135    718-2142-140    645-2229-126    520-2331-149    401-2451-148    360-2522-118    216-2559-225
Image

Stockfish 11

Code: Select all

Threads    2 vs 1          4 vs 2         8 vs 4           16 vs 8         32 vs 16       64 vs 32       128 vs 64
Elo          63              61             44                38              24             26               5
Draw %      70.8            72.7           77.5              80.1            84.6           86.1            88.2
W-D-L   709-2124-167    669-2180-151    525-2326-149    464-2403-133    336-2537-127    321-2582-97     199-2646-155
Image

tpm = time per move, d = depth, nps = nodes per second

Code: Select all

Komodo 14 T1      (tpm=218.0 d=19.07 nps=  2.184.031)
Komodo 14 T2      (tpm=214.8 d=20.52 nps=  4.356.131) - (tpm=219.6 d=20.38 nps=  4.371.401)
Komodo 14 T4      (tpm=217.7 d=21.72 nps=  8.598.857) - (tpm=221.8 d=21.58 nps=  8.427.235)
Komodo 14 T8      (tpm=222.9 d=23.01 nps= 17.083.765) - (tpm=220.0 d=22.89 nps= 16.319.575)
Komodo 14 T16     (tpm=224.5 d=24.06 nps= 30.843.341) - (tpm=225.8 d=24.44 nps= 33.266.657)
Komodo 14 T32     (tpm=224.1 d=24.77 nps= 57.449.961) - (tpm=222.8 d=25.60 nps= 73.557.642)
Komodo 14 T64     (tpm=214.5 d=26.26 nps=145.082.937) - (tpm=216.2 d=26.17 nps=145.036.085)
Komodo 14 T128    (tpm=200.2 d=25.29 nps=215.777.613)

Code: Select all

Stockfish 11 T1   (tpm=215.7 d=22.64 nps=  1.965.980)
Stockfish 11 T2   (tpm=212.1 d=25.49 nps=  3.851.322) - (tpm=211.4 d=24.46 nps=  4.008.706)
Stockfish 11 T4   (tpm=208.1 d=26.95 nps=  7.746.021) - (tpm=210.0 d=26.11 nps=  8.278.643)
Stockfish 11 T8   (tpm=207.2 d=28.15 nps= 15.410.676) - (tpm=207.4 d=27.96 nps= 18.544.535)
Stockfish 11 T16  (tpm=206.4 d=29.14 nps= 35.251.188) - (tpm=205.3 d=29.64 nps= 32.642.384)
Stockfish 11 T32  (tpm=205.1 d=30.18 nps= 59.852.477) - (tpm=203.6 d=30.64 nps= 77.849.924)
Stockfish 11 T64  (tpm=202.0 d=32.04 nps=147.277.339) - (tpm=202.3 d=31.59 nps=142.762.608)
Stockfish 11 T128 (tpm=201.9 d=31.20 nps=213.444.223)

Werewolf
Posts: 1281
Joined: Thu Sep 18, 2008 8:24 pm

Re: Elo gain by core doubling - Komodo 14, Stockfish 11 - Contempt 0

Post by Werewolf » Sun Jun 28, 2020 7:54 am

Surely the result of 128 v 64 will be impacted by moving from physical cores to logical cores..?

I think Dual EPYCs could still show a decent gain at 128 threads

corres
Posts: 3118
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Elo gain by core doubling - Komodo 14, Stockfish 11 - Contempt 0

Post by corres » Sun Jun 28, 2020 9:03 am

Werewolf wrote:
Sun Jun 28, 2020 7:54 am
Surely the result of 128 v 64 will be impacted by moving from physical cores to logical cores..?
I think Dual EPYCs could still show a decent gain at 128 threads
In general the the number of cores are only physical cores or (physical + logical) cores?

fastgm
Posts: 575
Joined: Mon Aug 19, 2013 4:57 pm
Contact:

Re: Elo gain by core doubling - Komodo 14, Stockfish 11 - Contempt 0

Post by fastgm » Sun Jun 28, 2020 2:12 pm

The AMD Ryzen Threadripper 3990X has 64 cores and 128 (hypertreading) threads.

lkaufman
Posts: 4297
Joined: Sun Jan 10, 2010 5:15 am
Location: Maryland USA
Contact:

Re: Elo gain by core doubling - Komodo 14, Stockfish 11 - Contempt 0

Post by lkaufman » Sun Jun 28, 2020 4:54 pm

So with Contempt off for both engines, Komodo gained more than Stockfish for every doubling of real cores. It seems that using Contempt significantly inflates the advantages of more cores, due to reducing draws.
Komodo rules!

corres
Posts: 3118
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Elo gain by core doubling - Komodo 14, Stockfish 11 - Contempt 0

Post by corres » Wed Jul 01, 2020 8:32 am

fastgm wrote:
Sun Jun 28, 2020 2:12 pm
The AMD Ryzen Threadripper 3990X has 64 cores and 128 (hypertreading) threads.
I do not ask you for the parameters of 3990x but I ask you, you used SMT or not used SMT at every point of your measurement.
If you would not use SMT at Threads = 2 the Elo difference would be higher between 1 Thread and 2 Threads and the full curve wuld be smaller about ~10% and maybe the curve of Elo would be more monotonous.

fastgm
Posts: 575
Joined: Mon Aug 19, 2013 4:57 pm
Contact:

Re: Elo gain by core doubling - Komodo 14, Stockfish 11 - Contempt 0

Post by fastgm » Wed Jul 01, 2020 9:41 am

The subject speaks for itself, core doubling.
Up to 64 = physical cores, 128 = hyperthreading threads

2 vs 1, 4 vs 2 ... 64 vs 32 physical cores and the last test
128 hyperthreading threads vs 64 physical cores

Post Reply