Lc0 in GTX video card

[Jouni]

My impression after some weeks testing. Interesting alternative to CPU engines. Can find some moves missed by SF. But I seldom start it anymore. Why? Too weak loses badly to SF12 and 4 cores! Weak in test suites. Don't show current move in analysis. Speed with NVIDIA drivers is about 3 knps/big net. Obviously Lc0 is usable only with RTX cards .

[Pi4Chess]

I agree. Leela needs so much computing power to achieve what SF 12 can do with normal hardware setup.

My joke is that "Leela is for the rich and Stockfish for the poor"

[MMarco]

My impression after some weeks testing. Interesting alternative to CPU engines. Can find some moves missed by SF. But I seldom start it anymore. Why? Too weak loses badly to SF12 and 4 cores! Weak in test suites. Don't show current move in analysis. Speed with NVIDIA drivers is about 3 knps/big net. Obviously Lc0 is usable only with RTX cards .

What card at what time control did you used?

I've a laptop with a GTX 1660Ti + Ryzen-7 3750H. Here are results at 2m+2s (with the Ryzen @ 3.3 Ghz):

Code: Select all

   # PLAYER                  :  RATING  ERROR  PLAYED    (%)   CFS    W    D    L   D(%)
   1 lc0-263-j94-80          :     9.1   22.9     300  51.17    78   62  183   55  61.00
   2 stockfish-12-4cpu       :     0.0   ----    1200  56.96    99  347  673  180  56.08
   3 lc0-263-ls-15           :   -35.1   27.7     300  45.50    99   58  157   85  52.33
   4 lc0-PhoenixStein-1.1    :   -85.5   28.1     300  39.17    87   36  163  101  54.33
   5 lc0-263-703810          :  -108.8   29.0     300  36.33   ---   24  170  106  56.67

White advantage = 114.47 +/- 6.16
Draw rate (equal opponents) = 67.30 % +/- 1.69

Yes...I get around 3knps for the big nets thanks to improvements in drivers and Lc0 backend. At faster time control, SFNNUE will kill it though.

As for Lc0 not showing analysis, it always does in console mode. I guess your GUI cuts the output at low depth, so that is why your missing the analysis lines. Is that a Chessbase GUI? I think there might be a workaround, but I'm not sure how to.

Anyway, like you said a RTX card is much better. Today there are laptops at 800-900$ usd with a RTX 2060, but the processors get stronger too! I have another one with a RTX 2060 and a 8-cores Ryzen-9 4900H. With SFNNUE on 8-cores I need a time control of 5m+3s to have Lc0 about on par. Note that a mobile Ryzen-9 is just as strong as some desktop CPU like a 3700x or a 9900k (without over overclocking).

Otherwise, I run micro-bullets on the GTX (2s + 0.2s) with 10x128 networks. Those are about on par with SFNNUE on one Ryzen-7 core in these conditions.

Code: Select all

   # PLAYER                 :  RATING  ERROR  PLAYED    (%)   CFS     W     D     L   D(%)
   1 lc0-263-tinker-6430    :    13.4    9.6    1952  51.87    95   495  1035   422  53.02
   2 lc0-263-703810         :     1.8   10.5    1994  50.25    63   469  1066   459  53.46
   3 stockfish-12-1cpu      :     0.0   ----   11881  50.74    88  2831  6395  2655  53.83
   4 lc0-263-730730         :    -6.1   10.4    1994  49.15    70   435  1090   469  54.66
   5 lc0-263-702820         :   -10.3   10.7    1994  48.57    59   441  1055   498  52.91
   6 lc0-263-jio-3450       :   -11.9   10.2    1953  48.34    83   397  1094   462  56.02
   7 lc0-263-730315         :   -18.6    9.0    1994  47.42   ---   418  1055   521  52.91

White advantage = 53.81 +/- 2.25
Draw rate (equal opponents) = 55.32 % +/- 0.43

[Nay Lin Tun]

I have 1060 GTX 3GB, on 4 cores PC, Lc0 and Stockfish are very close strength ( may be within 50 elo in blitz , although I havent tested enough games to get accurate elo).

If you have only built in intel GPU or > 5 years old GPU, Lc0 is not suitable for you, otherwise it is on par with Stockfish.

[phhnguyen]

My joke is that "Leela is for the rich and Stockfish for the poor"

Are you sure? You may buy 1K cheap computer + 1K graphics card to play Lc0. With 2K only it’s not easy to buy a multi-cores computer which is good enough for SF to vs Lc0 in above computer

[Nay Lin Tun]

My joke is that "Leela is for the rich and Stockfish for the poor"

Are you sure? You may buy 1K cheap computer + 1K graphics card to play Lc0. With 2K only it’s not easy to buy a multi-cores computer which is good enough for SF to vs Lc0 in above computer

Exactly, when you spend extra $500 or $1000 in upgrading PC ( a GPU upgrade or 4 cores to 32 cores cpu upgrade), a $ 500 or $1000 worth of GPU upgrade should be on par with similar upgrade in spending cpu, provided that you know which GPU is good for Lc0 and how to install Lc0 properly.

[jshriver]

Agree, my laptop has a 1660 in it and lc0 and Fat Fritz run horribly. So much I should do a match against tscp

[MMarco]

Agree, my laptop has a 1660 in it and lc0 and Fat Fritz run horribly. So much I should do a match against tscp

Did you use the recent cuda backend? That one works very bad with my 1660 Ti, but cudnn works fine.

[jshriver]

Agree, my laptop has a 1660 in it and lc0 and Fat Fritz run horribly. So much I should do a match against tscp

Did you use the recent cuda backend? That one works very bad with my 1660 Ti, but cudnn works fine.

I'm using the latest nvidia drivers and lc0 0.26.3 cuda that comes with Fritz 17.

Just looked and while i have the latest drivers, there is a newer cuda package. Installing now. Tried to run the version of lc0 from the official website but gave me a Cuda error: the provided ptx was compiled with an unsupported toolchain.

[Ferdy]

As for Lc0 not showing analysis, it always does in console mode. I guess your GUI cuts the output at low depth, so that is why your missing the analysis lines. Is that a Chessbase GUI? I think there might be a workaround, but I'm not sure how to.

Probably he meant this.