Page 3 of 4
Re: Congrats the Komodo Team
Posted: Fri Aug 16, 2019 6:24 am
by mjlef
zullil wrote: ↑Thu Aug 15, 2019 6:06 pm
mehmet karaman wrote: ↑Thu Aug 15, 2019 5:58 pm
Komodo won the WCCC.
After 9 rounds Komodo has 7 points at 9 game. Komodo is 1.5 points ahead of Chiron. Komodo guaranteed to be champion with one game left.
We now that Stockfish and Stockfish derivatives like Cfish, asmFish are the strongest engines. Lc0 is as strong as Stockfish with very powerful GPU's.
At WCCC not only chess engine is important. Also hardware, opening book and tablebase are important. The operator is important too. (We saw this last WCCC)
No doubt that TCEC is the best chess engine tournament in the world. But this tournament isn' t same category with TCEC.
WCCC is a whole system (software+ hardware+ operator) tournament. This success isn' t only belongs to Larry Kaufman and Mark Lefler (and Don Dailey). Erdogan Gunes has a share in success too.
And maybe the biggest share of the success goes to
8 x Intel (R) Xeon(R) Platinum 8168 CPU @ 2.7 GHz, 128 cores used.
But now Komodo can be marketed as the "World Computer Chess Champion".
Having the big machine certainly helps. I think it helps more in longer time controls. For Speed Chess we might have done better if we used a faster per core 16 core machine. It was not NUMA, so avoids the whole slow memory access, and it was over clocked above 4 GHz. But the 128 core was used in Speed Chess and WCCC this time. I think I need to do some more programming to take advantage of an 8 NUMA node machine. All those cores trying to access memory requires more speical memory management, so we will work on it. Komodo was getting about 70-90 million nps for much of the positions, but faster in endgames. With some work it can probably do better, but each processor doubling gives less and less elo and at some point, more is less.
Re: Congrats the Komodo Team
Posted: Fri Aug 16, 2019 7:10 am
by Graham Banks
Congratulations to the Komodo team.
Re: Congrats the Komodo Team
Posted: Fri Aug 16, 2019 11:38 am
by phhnguyen
zullil wrote: ↑Thu Aug 15, 2019 6:06 pm
And maybe the biggest share of the success goes to
8 x Intel (R) Xeon(R) Platinum 8168 CPU @ 2.7 GHz, 128 cores used.
Can someone give me an estimate about the price of that computer? Thanks
Re: Congrats the Komodo Team
Posted: Fri Aug 16, 2019 12:09 pm
by zullil
mjlef wrote: ↑Fri Aug 16, 2019 6:24 am
I think I need to do some more programming to take advantage of an 8 NUMA node machine. All those cores trying to access memory requires more speical memory management, so we will work on it. Komodo was getting about 70-90 million nps for much of the positions, but faster in endgames. With some work it can probably do better, but each processor doubling gives less and less elo and at some point, more is less.
Out of curiosity, what are the "node distances" on that machine? Under Linux, that table is provided by running
Re: Congrats the Komodo Team
Posted: Fri Aug 16, 2019 12:13 pm
by zullil
phhnguyen wrote: ↑Fri Aug 16, 2019 11:38 am
zullil wrote: ↑Thu Aug 15, 2019 6:06 pm
And maybe the biggest share of the success goes to
8 x Intel (R) Xeon(R) Platinum 8168 CPU @ 2.7 GHz, 128 cores used.
Can someone give me an estimate about the price of that computer? Thanks
Well, each CPU is about $6000 (6000 USD).
Re: Congrats the Komodo Team
Posted: Sat Aug 17, 2019 3:25 am
by mjlef
zullil wrote: ↑Fri Aug 16, 2019 12:09 pm
mjlef wrote: ↑Fri Aug 16, 2019 6:24 am
I think I need to do some more programming to take advantage of an 8 NUMA node machine. All those cores trying to access memory requires more speical memory management, so we will work on it. Komodo was getting about 70-90 million nps for much of the positions, but faster in endgames. With some work it can probably do better, but each processor doubling gives less and less elo and at some point, more is less.
Out of curiosity, what are the "node distances" on that machine? Under Linux, that table is provided by running
here they are:
node distances:
node 0 1 2 3 4 5 6 7
0: 10 20 20 20 20 20 20 20
1: 20 10 20 20 20 20 20 20
2: 20 20 10 20 20 20 20 20
3: 20 20 20 10 20 20 20 20
4: 20 20 20 20 10 20 20 20
5: 20 20 20 20 20 10 20 20
6: 20 20 20 20 20 20 10 20
7: 20 20 20 20 20 20 20 10
Re: Congrats the Komodo Team
Posted: Sat Aug 17, 2019 3:31 am
by bob
that is actually pretty amazing numbers...
Re: Congrats the Komodo Team
Posted: Sat Aug 17, 2019 10:13 am
by phhnguyen
zullil wrote: ↑Fri Aug 16, 2019 12:13 pm
phhnguyen wrote: ↑Fri Aug 16, 2019 11:38 am
zullil wrote: ↑Thu Aug 15, 2019 6:06 pm
And maybe the biggest share of the success goes to
8 x Intel (R) Xeon(R) Platinum 8168 CPU @ 2.7 GHz, 128 cores used.
Can someone give me an estimate about the price of that computer? Thanks
Well, each CPU is about $6000 (6000 USD).
Amazing expensiveness!!!
So the price of Komodo computer should be from $60K.
Another curious/stupid question: if we spend $60K to buy a computer/station with strong GPUs (not to hire special ones from Google/Amazon) for Leela0, can it beat Komodo with its $60K computer? (all counted recent status/versions)
Re: Congrats the Komodo Team
Posted: Sat Aug 17, 2019 12:00 pm
by mehmet karaman
Elo of Stockfish 10 is 47 points higher than Komodo 13.02
https://ssdf.bosjo.net/list.htm
At TCEC conditions elo of Stockfish 10 is 35 points higher than Komodo 13.02
Elo of Stockfish 18042019 (version played at TCEC 15 Super Final) is ~ 50 elo higher than Komodo 13.02
Elo of Lc0 0.21.2 T40.T8.610 (version played at TCEC 15 Super Final) is ~20 elo higher than Stockfish 18042019
Elo of last version Lc0 (Lc0 427xx) is ~10 elo higher than Lc0 0.21.2 T40.T8.610
Komodo XXX (WCCC 2019) is 5-10 elo higher 13.02
Elo of Lc0 427xx is ~ 70 elo higher than Komodo
Lc0 (RTX 2080 Tİ + RTX 2080) is 70 elo higher than Komodo XXX (44 CPU)
Lc0 (RTX 2080 Tİ + RTX 2080) is 10 elo higher than Komodo XXX (128 CPU)
My conclusion is strongest Lc0 version at RTX 2080 Tİ + RTX 2080 is a bit stronger than Komodo WCCC 2019 version at 128 CPU (with same opening book)
Re: Congrats the Komodo Team
Posted: Sat Aug 17, 2019 12:07 pm
by zullil
mjlef wrote: ↑Sat Aug 17, 2019 3:25 am
zullil wrote: ↑Fri Aug 16, 2019 12:09 pm
mjlef wrote: ↑Fri Aug 16, 2019 6:24 am
I think I need to do some more programming to take advantage of an 8 NUMA node machine. All those cores trying to access memory requires more speical memory management, so we will work on it. Komodo was getting about 70-90 million nps for much of the positions, but faster in endgames. With some work it can probably do better, but each processor doubling gives less and less elo and at some point, more is less.
Out of curiosity, what are the "node distances" on that machine? Under Linux, that table is provided by running
here they are:
node distances:
node 0 1 2 3 4 5 6 7
0: 10 20 20 20 20 20 20 20
1: 20 10 20 20 20 20 20 20
2: 20 20 10 20 20 20 20 20
3: 20 20 20 10 20 20 20 20
4: 20 20 20 20 10 20 20 20
5: 20 20 20 20 20 10 20 20
6: 20 20 20 20 20 20 10 20
7: 20 20 20 20 20 20 20 10
Thanks. So all other nodes are "equidistant" from any one fixed node. So even if a given search thread wanders off its preferred node, the penalty is never too severe.