TalkChess.com

Modern Times wrote: ↑Sat Sep 07, 2019 12:17 am

lkaufman wrote: ↑Sat Sep 07, 2019 12:02 am
OK, I'll rephrase the question. What is the best network and settings for any NN on 7 threads of an 8 core, very fast i7 machine? LD2 is beating 11248 on that config by 345 elo in my current test, so the question is whether some other network is even better, and whether default settings are appropriate?

So how strong is LD2 on that set-up Larry ?

I don't have any other test results yet, but 11248 cpu only is several hundred elo below the best cpu engines so I suppose even LD2 isn't one of the top CPU engines, but pretty good.

lkaufman wrote: ↑Sat Sep 07, 2019 12:02 am

dkappe wrote: ↑Fri Sep 06, 2019 11:24 pm

pohl4711 wrote: ↑Fri Sep 06, 2019 9:05 pm

lkaufman wrote: ↑Fri Sep 06, 2019 8:14 pm Which NN network, Lc0 or other, is currently believed to be the strongest one for use only on a CPU, no GPU? I've seen posts on this before, but this is always changing. I understand it should be a small network.

No doubt, that LD2 is the strongest for CPU. 10x128 built out of T40 training games.

https://lc0.org/ld2

There is significant doubt. There's a sweet spot for CPU which depends on nps. Generally, the slower the CPU, the better 11258-112x9-se performs vs nets like LD2. Indeed, on slow Android devices, 11258-48x5-se outperforms LD2 comfortably. The best thing is to grab a few nets and test them out on your particular setup

OK, I'll rephrase the question. What is the best network and settings for any NN on 7 threads of an 8 core, very fast i7 machine? LD2 is beating 11248 on that config by 345 elo in my current test, so the question is whether some other network is even better, and whether default settings are appropriate?

After running over 10,000 test games with Lc0. Here is what I know. The smaller nets perform best on CPU, the biggest nets run best on RTX GPU's (320x24b). And by best I mean better Elo gains with everything else being equal.

And for the love of god don't core cut your processor by using 7 threads, and a 8 core machine running Lc0. That is a waste of resources. You are wasting 1 core of your system to run windows, and it does not need that much run time. There is a better way. One easy way if using the Fritz GUI is to configure Lc0 to run at below normal priority. Then run Lc0 or a A/B engine with all 8 cores for best performance. You can do this by checking the below normal box when setting up Lc0. This is the best setup for running Lc0 by itself, not 2 engine testing....

lkaufman wrote: ↑Sat Sep 07, 2019 12:02 am OK, I'll rephrase the question. What is the best network and settings for any NN on 7 threads of an 8 core, very fast i7 machine? LD2 is beating 11248 on that config by 345 elo in my current test, so the question is whether some other network is even better, and whether default settings are appropriate?

I'm surprised that 11248 does even that well. Part of my work distilling 11258 to smaller sizes was finding the best combination of NPS and smarts. For a reasonable CPU, LD2 may be among the best. For a less powerful CPU 11258-112x9-se (112 filters by 9 blocks) may do better than LD2.

TC is also a factor.

What about other settings (smart prunning, max prefetch, cpuct etc) Ferdy? Are you use default ?

mesilikas wrote: ↑Sun Sep 08, 2019 12:25 am What about other settings (smart prunning, max prefetch, cpuct etc) Ferdy? Are you use default ?

Other options are on their default values.

Update, added more games in tc 5m+3s and add new tc 10m+3s.

Added 256x20 at tc 10m+3s. It only scored around 17% after 36 games.

Ferdy wrote: ↑Sat Sep 07, 2019 11:58 pm TC is also a factor.

NN-cache is a factor too.
But only if it gets filled up.

Ferdy wrote: ↑Sun Sep 08, 2019 6:40 pm Added 256x20 at tc 10m+3s. It only scored around 17% after 36 games.

Very good Ferdy. A proof that longer TC favor bigger, slower nets.

Laskos wrote: ↑Sun Sep 08, 2019 8:49 pm

Ferdy wrote: ↑Sun Sep 08, 2019 6:40 pm Added 256x20 at tc 10m+3s. It only scored around 17% after 36 games.

Very good Ferdy. A proof that longer TC favor bigger, slower nets.erers er

Or bigger hardware. That is why ccrl test and others can not gadge Lc0 correctly. Faster Hardware equals longer time controls.