TalkChess.com

Lc0 parameters: NNCachesize=10000000 SmartPruningFactor=0.0 Backend=cudnn-fp16, for DUAL GPU=Multiplexing Other parameters are Default Maybe you forgot to mention threads since default of 2 will be non optimal. Have you tried threads=3 with round-robin yet? TCEC uses this for a reason because it's ...

if threadcount and NPS are the same NPS only hold true. If you are using the same number of threads or cores. ... This is why if you look at both screen shots. And look at the total nodes searched the reach depth 36 ply. The 32 thread Stockfish had to search more then 1.1 billion more nodes. To rea...

Testrun of Lc0 0.20.2 N:32930 finished. I was wondering if you used the defaults or the Superfinal settings that improved tactical performance? Cpuct=3.4 and cpuctbase=10000 (and enough nncache for longer thinks). (I've tested Leelafish some and it seems that it's actually worse with these tactical...

I think NPS and thread count are what matters, not whether SMT is on or off. A modern computer could have a similar threadcount and NPS to a 5 year old computer that used actual cores instead of SMT, and it shouldn't really matter if threadcount and NPS are the same. Just cheaper hardware to get tha...

Eduard wrote: ↑

Fri Feb 22, 2019 10:36 pm

I see that it is also possible to combine two neural networks. One as the main engine and one as the AuxEngine. Interesting.

I tried that and it seemed to work for me.

using 3 threads for SFdev as aux engine, 1024MB hash, everything else default. Insomuch as the defaults are balanced for 1 thread, using more threads will reach the same depth faster, and thus have empty queues too often. So with more threads it should be necessary either to increase the default de...

Uma has greater memory bandwidth (since it can also use memory channels that are further away) but higher memory latency. So for hash tables numa is better.

Those are the best available, but almost useless now since they were before we had the a0 puct scaling formula. So it didn't scale well at all without that. Also T30 and T40 seem to scale differently than the T10 series in shorter benchmark results. See http://talkchess.com/forum3/viewtopic.php?f=2&...

Ankan's benchmark, is a bit out of date, but still quite relevant I think. Latest lc0 should be a bit faster, and some defaults have changed. I'm excited to follow your progress. with cudnn 7.3 and 411.63 driver available at nvidia.com minibatch-size=512, network id: 11250, go nodes 1000000 v0.17, d...

Two 2080s is certainly much stronger than one 2080 ti. You can see 2 x 2070 sometimes on TCEC bonus and it is faster than a 2080 ti for Leela, (but probably not anything else as it doesn't even support NVLINK/SLI). I do recommend the 2950x and x399 mobo, but if you can wait until October the 30xx CP...