Search found 75 matches

by ankan
Thu Feb 27, 2020 2:48 am
Forum: Computer Chess Club: General Topics
Topic: Better settings, best net, best backend on RTX GPU to LTC
Replies: 11
Views: 1407

Re: Better settings, best net, best backend on RTX GPU to LTC

Laskos wrote:
Wed Feb 26, 2020 8:24 pm
Yes, I also tried to increase the batch size to 512, but DX backend seems buggy in this respect. I haven't tried very large nncaches.
Max supported batch size for dx backend is 256. However you should be able to set any nncache size.
by ankan
Thu Feb 27, 2020 2:10 am
Forum: Computer Chess Club: General Topics
Topic: Better settings, best net, best backend on RTX GPU to LTC
Replies: 11
Views: 1407

Re: Better settings, best net, best backend on RTX GPU to LTC

Great info. I look forward to trying the 1.9 CPuct. I do have a completely different experience on the DX12 backend however. I run an ASUS Strix 2080TI OC (watercooled). And the DX12 backend is nearly 25% slower. It also goes nuts if I set a batch size of 512 like I run under cuda. It likewise goes...
by ankan
Thu Feb 13, 2020 3:37 pm
Forum: Computer Chess Club: General Topics
Topic: Lc0 v0.24 dev DX backend for AMD Radeon GPU
Replies: 11
Views: 1607

Re: Lc0 v0.24 dev DX backend for AMD Radeon GPU

Thanks Laskos for testing. It's good to know that the speed increase translates to improvement in playing strength. The dx backend (that defaults to fp16 precision) uses a different algorithm for convolution (winograd) that scales better with bigger networks compared to what cudnn-fp16 uses (implic...
by ankan
Thu Feb 13, 2020 11:48 am
Forum: Computer Chess Club: General Topics
Topic: Lc0 v0.24 dev DX backend for AMD Radeon GPU
Replies: 11
Views: 1607

Re: Lc0 v0.24 dev DX backend for AMD Radeon GPU

Thanks Laskos for testing. It's good to know that the speed increase translates to improvement in playing strength. The dx backend (that defaults to fp16 precision) uses a different algorithm for convolution (winograd) that scales better with bigger networks compared to what cudnn-fp16 uses (implici...
by ankan
Wed Jan 01, 2020 12:52 pm
Forum: Computer Chess Club: General Topics
Topic: External vs internal gpu for LC0
Replies: 5
Views: 798

Re: External vs internal gpu for LC0

I was thinking of getting a 2 in 1 laptop with an external gpu (2080ti) connected to thunderbolt or a desktop with the 2080ti. Would there be a significant difference in performance for Lela? I was thinking if I got the 2 in 1 then I wouldn’t need a desktop. lc0 isn't much bottlenecked by PCIe band...
by ankan
Thu Jun 20, 2019 3:19 pm
Forum: Computer Chess Club: Tournaments and Matches
Topic: Leela Chess Zero 42565 vs Stockfish 140619
Replies: 75
Views: 9484

Re: Leela Chess Zero 42565 vs Stockfish 140619

Pondering makes sense when playing against Leela on a many-core machine, as Leela wouldn't use many CPU threads while thinking, and all other threads would then go to waste. Likewise the GPUs would be idle during Stockfish' turn when Leela is not pondering. If Leela need two unshared cores, you can...
by ankan
Fri Mar 01, 2019 5:13 pm
Forum: Computer Chess Club: General Topics
Topic: Lc0 settings for Leela on GPU- Nvidia RTX 2080
Replies: 11
Views: 3354

Re: Lc0 settings for Leela on GPU- Nvidia RTX 2080

I agree that 2 threads is optimal for single GPU, and you must have --backend=cudnn-fp16 (to get ~2.5X speedup over the default cudnn fp32 backend). Changing some of the settings for more NPS can weaken the search. E.g, I wouldn't suggest increasing minibatch-size more than the default (256). Increa...
by ankan
Fri Mar 01, 2019 4:37 pm
Forum: Computer Chess Club: General Topics
Topic: AMD RYZEN 2950 Speed issue
Replies: 31
Views: 5951

Re: AMD RYZEN 2950 Speed issue

I read all the replies and did not see the biggest issue of the 2950x addressed. On the 2950x 16 cores are directly connected to the memory bus. The other 16 cores are connected to the memory bus through the first 16 cores. This creates a memory bottleneck for the 2nd 16 cores. Chess programs are b...
by ankan
Tue Feb 12, 2019 5:03 pm
Forum: Computer Chess Club: General Topics
Topic: Scorpio+Lc0-32742 vs Stockfish 9
Replies: 13
Views: 3358

Re: Scorpio+Lc0-32742 vs Stockfish 9

./lc0 -t 2 --backend=cudnn --verbose-move-stats --weights=../../../lcnets/ID-11248.gz | _ | | |_ |_ |_| v0.19.1 built Feb 11 2019 go movetime 100000 Loading weights file from: ../../../lcnets/ID-11248.gz Creating backend [cudnn]... For fair comparison, you should use cudnn-fp16 backend for lc0. Jus...
by ankan
Sat Jan 26, 2019 11:42 am
Forum: Computer Chess Club: General Topics
Topic: Big boost in strength for LC0 by simply changing VGA card
Replies: 19
Views: 4288

Re: Big boost in strength for LC0 by simply changing VGA card

Card I am using is now an RTX 2080Ti which reading on the forums seems to indicate would do very well. The previous card was a Quadro M4000 with around 1700 Cuda cores while the RTX 2080Ti has about 3.5x that. I would expect a bigger jump in NPS going from Quadro M4000 to the RTX 2080Ti (something ...