Leela Chess Zero 42518 vs Stockfish 100619

[mwyoung]

Leela Chess Zero 42518 vs Stockfish 100619

Hardware 2950x, RTX 2080 ti
Stockfish 32 Threads, 32 Gb HT
Leela Chess Zero 2 threads, 9900000 NN CS.
6 man Tablebase
12 move opening book
TC= 30m + 30s

Live Stream: https://www.youtube.com/watch?v=M-u8l7F-m9s

[Nordlandia]

Considering SF play on CPU and Lz0 on GPU. It probably make sense to play with permanent brain turned on. I tried on my system couple of days ago. As long as one core is left for the OS and 1-2 core for Lz0. Have you tried this?

[jorose]

Considering SF play on CPU and Lz0 on GPU. It probably make sense to play with permanent brain turned on. I tried on my system couple of days ago. As long as one core is left for the OS and 1-2 core for Lz0. Have you tried this?

He only has 16 physical cores so how many threads would you give Leela? Regardless of how you do this, one engine will be interfering with the other in an unpredictable way. Im also not too sure about the effects on CPU cache or memory, but I imagine the it could lead to some very serious memory thrashing.

While he isn't running this for scientific reasons, so he could do that if he wanted to, you are still handicapping one of the engines (I would guess SF, but could be wrong) so it would detract quite a bit from the weight of the result.

[mwyoung]

Considering SF play on CPU and Lz0 on GPU. It probably make sense to play with permanent brain turned on. I tried on my system couple of days ago. As long as one core is left for the OS and 1-2 core for Lz0. Have you tried this?

I designed my system to test this way. And it works very well. Lc0 has such a low overhead on my system. That this way of testing is playing each engine at its best.

I just don't share this information with this site. As it gets everyone's panties tied into a bunch when I posted those matches.

But the results are the same either way of testing.

[Nordlandia]

Earlier today i watched your stream and the following position was reached.

[d]r7/5kPK/8/7P/8/2B5/8/8 w - - 0 0

From this moment the game continues for 30-45 minutes. It's very tiresome and annoying to keep playing inside EGTB position when both engines know it's draw. Fritz GUI need to update few of the adjudication settings. Otherwise several hours if not a days will be wasted in a long testing event.

6-piece Syzygy adjudication is a time saver BIG Time. Chess.com (CCC) use it and TCEC use it.

CuteChess GUI adjudication settings is alot better and saved alot of time as result. Have you tried it?

[jorose]

I don't think you will actually be able to judge the effect it is having without a reasonably large sample size. Even if Leela has a reasonably low overhead in terms of CPU usage the effect on cache is harder to measure. It's also unclear what you would compare the performance to, perhaps one engine is getting handicapped, but since it scales better this effect is masked?

That being said, if you would do single machine ponder on matches, I think you chose very good hardware for that.

[mwyoung]

I don't think you will actually be able to judge the effect it is having without a reasonably large sample size. Even if Leela has a reasonably low overhead in terms of CPU usage the effect on cache is harder to measure. It's also unclear what you would compare the performance to, perhaps one engine is getting handicapped, but since it scales better this effect is masked?

That being said, if you would do single machine ponder on matches, I think you chose very good hardware for that.

You are correct. I would not be able to judge without serious testing......