Lc0 on RTX 2080 Super

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

User avatar
Nordlandia
Posts: 2821
Joined: Fri Sep 25, 2015 9:38 pm
Location: Sortland, Norway

Lc0 on RTX 2080 Super

Post by Nordlandia »

Received 2080 Super yesterday replacing my 1070 Ti that have been in use for about 2 years.

Speed for Lc0: 16-35-55 Kilo nodes per second.

Game observation: look at that miserable bishop at h1 later in the game. Komodo was lucky this time.

[pgn][Event "DESKTOP-N1IDKRV, Rapid 30m+30s"] [Site "Expert-Desktop"] [Date "2019.08.16"] [Round "1"] [White "Komodo 13.02 64-bit"] [Black "Lc0 v0.22.0"] [Result "1/2-1/2"] [Annotator "1.44;0.91"] [SetUp "1"] [FEN "N1bk1b1r/p2pq2p/1pnn2p1/3Qpp2/8/1B6/PPPP1PPP/R1B1K1NR w KQ - 0 1"] [PlyCount "87"] [TimeControl "1800+30"] 1. Ne2 {1.44/32 131 Both last book move} Bb7 {0.91/15 29} 2. Qd3 {1.08/30 100} Bxa8 {0.92/32 0} 3. O-O {1.33/29 47 (c3)} h5 {0.89/19 67 (f4)} 4. a3 {1.19/32 154 (c3)} h4 {0.86/17 94} 5. Bd5 {1.10/31 2} h3 {0.89/17 53 (Bg7)} 6. g3 { 1.07/32 76} Bg7 {0.88/19 30} 7. Qb3 {1.00/33 85} Ne8 {0.90/13 0} 8. Bh1 { 1.06/33 243} Nc7 {0.94/21 74} 9. d3 {1.22/31 32} g5 {0.95/20 32 (Rf8)} 10. Bd2 {1.21/31 78 (a4)} Kc8 {0.90/16 210 (f4)} 11. a4 {1.12/31 155} f4 {0.92/26 0} 12. Rae1 {1.20/32 179 (a5)} Rf8 {0.61/18 81 (Qf8)} 13. Nc3 {1.25/31 62} f3 { 0.56/33 1 (Kb8)} 14. a5 {1.13/33 158} Nd4 {0.48/26 3} 15. Qa4 {1.30/32 52} Qf6 {0.45/33 0 (Qc5)} 16. axb6 {1.48/32 53} axb6 {0.46/38 0} 17. Qa7 {1.10/33 77} Qg6 {0.41/32 0 (Bb7)} 18. Be3 {1.02/31 57} g4 {0.36/31 30} 19. Rb1 {0.90/34 72 (Bxd4)} Rf6 {0.17/27 222} 20. Bxd4 {0.84/41 0} exd4 {0.06/27 91} 21. Nb5 { 0.54/37 0 (Na4)} Bb7 {0.28/24 138} 22. Nxc7 {0.29/37 0 (Nxd4)} Kxc7 {0.01/20 113} 23. Rbc1 {0.21/38 0 (Rfe1)} Rf5 {-0.73/15 119 (Rc6)} 24. Qa2 {0.00/39 48 (Qa4)} Ra5 {-0.75/16 67} 25. Qc4+ {0.00/38 0} Rc5 {-0.75/16 34} 26. Qb4 { 0.00/39 14 (Qb3)} Qf5 {-0.71/17 64} 27. Qa4 {0.00/38 33} Rb5 {-0.72/26 53 (Qg6) } 28. b4 {0.00/36 72 (Rb1)} Re5 {-0.85/18 88 (Bh6)} 29. c3 {0.00/39 58 (Rce1)} Re2 {-1.04/22 71 (dxc3)} 30. cxd4+ {0.00/36 26} Kb8 {-1.08/26 0} 31. Qb3 { 0.00/41 62 (b5)} Bxd4 {-0.26/28 218} 32. Qg8+ {0.00/44 0} Ka7 {-0.17/30 11} 33. Ra1+ {0.00/43 42} Bxa1 {-0.10/33 0} 34. Rxa1+ {0.00/43 56} Ba6 {-0.09/32 19} 35. b5 {0.00/45 53} Qe5 {-0.10/27 13 (Qxb5)} 36. Rxa6+ {0.00/39 48} Kb7 { -0.07/25 0} 37. Qa8+ {0.00/41 58} Kc7 {-0.07/21 0} 38. Ra7+ {0.00/42 79} Kd6 { -0.08/23 0} 39. Qf8+ {0.00/42 62} Ke6 {-0.08/22 58} 40. Qh6+ {0.00/43 12} Kd5 { -0.07/20 78 (Ke7)} 41. Bxf3+ {0.00/43 52 (Rxd7+)} gxf3 {-0.04/19 199} 42. Rxd7+ {0.00/49 0} Kc5 {-0.04/18 49} 43. Qc6+ {0.00/47 162 (Qc1+)} Kb4 {-0.03/15 0} 44. Qc4+ {0.00/50 41 adjud. Draw accepted} 1/2-1/2 [/pgn]

Brief lichess analysis - https://lichess.org/NGEUTwmi
jp
Posts: 1470
Joined: Mon Apr 23, 2018 7:54 am

Re: Lc0 on RTX 2080 Super

Post by jp »

As with your other thread, I don't think you can say that Komodo was "lucky" without seriously analysing the position first. By move 36, that "miserable" bishop is a full extra bishop. I'd be happy to play with many bad pieces if they're all extra pieces. Even before move 36, Leela had two bishops for Komodo's rook plus "miserable" bishop.
corres
Posts: 3657
Joined: Wed Nov 18, 2015 11:41 am
Location: hungary

Re: Lc0 on RTX 2080 Super

Post by corres »

Nordlandia wrote: Fri Aug 16, 2019 8:44 am Received 2080 Super yesterday replacing my 1070 Ti that have been in use for about 2 years.
Speed for Lc0: 16-35-55 Kilo nodes per second.
...
What is 15-35-55 Kns/sec?
What were the parameters? (net files, Lc0 parameters, etc)
User avatar
Nordlandia
Posts: 2821
Joined: Fri Sep 25, 2015 9:38 pm
Location: Sortland, Norway

Re: Lc0 on RTX 2080 Super

Post by Nordlandia »

Speed ranging from 17 to middle of 30s into 50 kilo nodes per second.

Network is 128x10-2019_0812_1326_54_239 - is it good and is there any better at this point?
User avatar
Laskos
Posts: 10948
Joined: Wed Jul 26, 2006 10:21 pm
Full name: Kai Laskos

Re: Lc0 on RTX 2080 Super

Post by Laskos »

Nordlandia wrote: Fri Aug 16, 2019 5:33 pm Speed ranging from 17 to middle of 30s into 50 kilo nodes per second.

Network is 128x10-2019_0812_1326_54_239 - is it good and is there any better at this point?
Wrong speeds for that smallnet and 2080 Super. On your GPU, JHorthos 256x20b latest nets are the strongest to rapid TC, and his latest bignet 320x24b is the strongest to LTC. That 128x10b net might be the strongest in blitz on 1070.
Daniel Shawul
Posts: 4185
Joined: Tue Mar 14, 2006 11:34 am
Location: Ethiopia

Re: Lc0 on RTX 2080 Super

Post by Daniel Shawul »

Nordlandia wrote: Fri Aug 16, 2019 5:33 pm Speed ranging from 17 to middle of 30s into 50 kilo nodes per second.

Network is 128x10-2019_0812_1326_54_239 - is it good and is there any better at this point?
On my rtx 2070-super lc0 gets

Code: Select all

FP16 28k nps
FP32 10k nps
using a 20x256 net, so it is a bigger net than yours.
corres
Posts: 3657
Joined: Wed Nov 18, 2015 11:41 am
Location: hungary

Re: Lc0 on RTX 2080 Super

Post by corres »

Daniel Shawul wrote: Sat Aug 17, 2019 3:00 am On my rtx 2070-super lc0 gets

Code: Select all

FP16 28k nps
FP32 10k nps
using a 20x256 net, so it is a bigger net than yours.
It is interesting to see there is only 2.8 times difference between speed of FP16 and of FP32.
For older GPUs the difference was about four times.
Which net you used in concrete and what was the parameters of Leela?
User avatar
Nordlandia
Posts: 2821
Joined: Fri Sep 25, 2015 9:38 pm
Location: Sortland, Norway

Re: Lc0 on RTX 2080 Super

Post by Nordlandia »

What is the best network for RTX 2080 Super on i7-5960X 4.3GHz with 128 GB RAM and partial 7-piece syzygy.
mwyoung
Posts: 2727
Joined: Wed May 12, 2010 10:00 pm

Re: Lc0 on RTX 2080 Super

Post by mwyoung »

Nordlandia wrote: Sun Aug 18, 2019 6:06 am What is the best network for RTX 2080 Super on i7-5960X 4.3GHz with 128 GB RAM and partial 7-piece syzygy.
TB don't matter. Your CPU is fine and should not matter. The best NN I have tested are 42841, and 42850. The end of the line for the T40000. I have tested T6000, but it is still too weak. Needs more training. Right now SF seems to be King of the hill once again with the latest patch. Maybe T6000 can change this, but it needs more time...
"The worst thing that can happen to a forum is a running wild attacking moderator(HGM) who is not corrected by the community." - Ed Schröder
But my words like silent raindrops fell. And echoed in the wells of silence.
User avatar
Nordlandia
Posts: 2821
Joined: Fri Sep 25, 2015 9:38 pm
Location: Sortland, Norway

Re: Lc0 on RTX 2080 Super

Post by Nordlandia »

I tried Lc0 in selfplay in this borderline endgame (by borderline i mean considerable advantage for white but not certain that it's winning edge).

i7-5960X 4.4GHz
Net used: T40e-swa-3000
TC:30m+30s
EGTB: Partial 7-piece syzygy
GUI: Fritz GUI

[pgn][Event "DESKTOP-N1IDKRV, Rapid 30m+30s"] [Site "Expert-Desktop"] [Date "2019.08.19"] [Round "1"] [White "Lc0 v0.22.0"] [Black "Lc0 v0.22.0"] [Result "1-0"] [SetUp "1"] [FEN "2r1bk2/pp4pp/1nn5/8/4P3/2N2P2/PP1R2PP/3R2K1 w - - 0 1"] [PlyCount "117"] [TimeControl "1800+30"] {Intel(R) Core(TM) i7-5960X CPU @ 3.00GHz 3000 MHz W=18.1 plies; 35kN/s; 684 732 TBAs B=18.1 plies; 35kN/s; 603 832 TBAs} 1. Rc2 {1.89/12 58 Both last book move} Bf7 {1.85/13 37} 2. b3 {1.90/15 22} Nb4 {1.87/14 32} 3. Rcc1 { 1.90/14 29} Ke7 {1.87/14 31} 4. Rd2 {1.90/14 35} a5 {1.88/14 47} 5. h4 { 1.88/15 90} Rc6 {1.85/17 139} 6. Ne2 {1.88/16 63} g6 {1.85/17 11} 7. Kh2 { 1.91/17 38} Nd7 {1.89/17 100} 8. Rxc6 {1.92/18 55} bxc6 {1.90/18 1} 9. g4 { 1.93/17 36} c5 {1.91/16 87} 10. Kg3 {1.92/17 205} Nb6 {1.88/17 58} 11. Nc3 { 1.93/20 0} c4 {1.89/18 19} 12. bxc4 {1.93/17 4} Nxc4 {1.90/18 24} 13. Rd1 { 1.95/14 38} Be6 {1.89/15 82} 14. g5 {1.90/16 103} Bf7 {1.85/14 77} 15. f4 { 1.85/15 127} Nd6 {1.83/21 63} 16. a3 {1.86/23 0} Nc2 {1.86/22 29} 17. a4 { 1.86/19 18} Bb3 {1.87/22 46} 18. h5 {1.83/17 63} gxh5 {1.82/18 52} 19. Kh4 { 1.87/18 193 (f5)} Ne3 {1.85/18 175} 20. Rd2 {1.88/20 14} Bf7 {1.89/20 26} 21. f5 {1.91/20 34} Be8 {1.99/20 169} 22. Rd3 {1.97/21 47} Nec4 {1.99/24 1} 23. Rd5 {2.03/20 43} Bf7 {2.02/23 26} 24. Rd4 {2.08/21 46} Ne3 {2.14/21 144} 25. Rd3 { 2.15/21 54} Ng4 {2.13/19 0} 26. Kg3 {2.19/21 45} Ne5 {2.14/22 17} 27. Rd4 { 2.20/24 36} Nc6 {2.17/23 72} 28. Rd1 {2.24/24 61} Bb3 {2.20/23 36} 29. Rh1 { 2.27/24 46} Bf7 {2.23/24 33} 30. Kf4 {2.29/22 35} Nb4 {2.26/22 33} 31. Rh3 { 2.32/20 37} Nb7 {2.29/20 59} 32. e5 {2.34/21 45} Nc5 {2.32/20 36} 33. Kg3 { 2.37/22 51} Nbd3 {2.37/20 168} 34. e6 {2.40/22 77} Nxe6 {2.36/22 1} 35. fxe6 { 2.41/21 36} Kxe6 {2.38/21 23} 36. Rh1 {2.43/20 44} Bg6 {2.42/19 88} 37. Rb1 { 2.46/20 68} Nb4 {2.43/19 21} 38. Re1+ {2.48/19 47} Kd6 {2.45/18 45} 39. Ne4+ { 2.51/18 32} Kc6 {2.50/18 156} 40. Nf6 {2.56/19 110} Bc2 {2.54/19 76} 41. Rc1 { 2.60/19 87} Kd6 {2.55/18 13} 42. Kh4 {2.63/18 54} Bxa4 {2.58/17 80} 43. Ra1 { 2.66/18 74} Bc2 {2.60/17 33} 44. Rxa5 {2.66/16 27} Bg6 {2.65/15 141} 45. Ra7 { 2.70/15 81} Nc6 {2.64/14 0} 46. Ra6 {2.72/14 37} Kc5 {2.69/13 126} 47. Nxh5 { 2.76/12 91} Bd3 {2.73/12 149} 48. Ra3 {2.84/12 136} Bg6 {2.81/12 145} 49. Nf6 { 2.95/12 107} Kd6 {2.88/12 42} 50. Ra6 {3.08/12 82} Kc5 {3.03/11 55} 51. Kg3 { 3.22/11 104} Kd6 {3.09/10 24} 52. Kf4 {3.28/11 34} Bc2 {3.19/10 64} 53. Rb6 { 3.92/12 95 (Ng4)} Bg6 {3.44/11 30} 54. Rb7 {4.07/13 13} Ne7 {3.71/12 26} 55. Ne4+ {4.32/13 30} Ke6 {4.05/11 21} 56. Rb6+ {4.63/11 23} Kd7 {4.40/11 92} 57. Nc5+ {5.16/10 62} Ke8 {4.85/10 46} 58. Ke5 {5.93/9 62} Bh5 {5.05/8 43} 59. Ne6 {6.68/9 101 (Rb7)} 1-0 [/pgn]

Brief lichess analysis - https://lichess.org/KoaEaf6K#0