Leela Chess Zero 42478 vs Komodo 13.02 MCTS - experimental

[mwyoung]

Leela Chess Zero 42478 vs Komodo 13.02 MCTS - experimental

Komodo 13 MCTS shows significant speed gains when using SMT and 32 cores on my system configuration.

If this setting is somehow hurting K13. Komodo's setting instructions do advise against using this many logical cores.

komodo-13.02-64bit.jpg

But the potential speed gains of 28% to 35% need to be tested in real games.

Hardware 2950x, RTX 2080 ti
Komodo 32 Thread, 32 Gb HT
Leela Chess Zero 2 threads, 9900000 NN CS.
6 man Tablebase
12 move opening book
TC= 1h/40 +30m/20

Live Stream: https://www.youtube.com/watch?v=px82fNsWy8Y

[mjlef]

Leela Chess Zero 42478 vs Komodo 13.02 MCTS - experimental

Komodo 13 MCTS shows significant speed gains when using SMT and 32 cores on my system configuration.

If this setting is somehow hurting K13. Komodo's setting instructions do advise against using this many logical cores.
komodo-13.02-64bit.jpg
But the potential speed gains of 28% to 35% need to be tested in real games.

Hardware 2950x, RTX 2080 ti
Komodo 32 Thread, 32 Gb HT
Leela Chess Zero 2 threads, 9900000 NN CS.
6 man Tablebase
12 move opening book
TC= 1h/40 +30m/20

Live Stream: https://www.youtube.com/watch?v=px82fNsWy8Y

I would love to hear which tests better. Note for the 16 thread run you should have HT off, although Komodo's default thread binding, binds each program thread to every other system thread if it detects multiple thread per core (basically that HT is on). Once it runs out of threads (such as you setting it for 32 threads) it just lets the OS stick them wherever it wants. I see a few other programs doing the same thing as a way to ensure reasonable binding.

[mwyoung]

Leela Chess Zero 42478 vs Komodo 13.02 MCTS - experimental

Komodo 13 MCTS shows significant speed gains when using SMT and 32 cores on my system configuration.

If this setting is somehow hurting K13. Komodo's setting instructions do advise against using this many logical cores.
komodo-13.02-64bit.jpg
But the potential speed gains of 28% to 35% need to be tested in real games.

Hardware 2950x, RTX 2080 ti
Komodo 32 Thread, 32 Gb HT
Leela Chess Zero 2 threads, 9900000 NN CS.
6 man Tablebase
12 move opening book
TC= 1h/40 +30m/20

Live Stream: https://www.youtube.com/watch?v=px82fNsWy8Y

I would love to hear which tests better. Note for the 16 thread run you should have HT off, although Komodo's default thread binding, binds each program thread to every other system thread if it detects multiple thread per core (basically that HT is on). Once it runs out of threads (such as you setting it for 32 threads) it just lets the OS stick them wherever it wants. I see a few other programs doing the same thing as a way to ensure reasonable binding.

The 16 thread run was with SMT (ht) off.