Using eGPU for Chess  

Some more background about chess engines:

When I started programming chess engines in the year 1996, human were still stronger. The ELO rating of the best human is 2700 (as it is today), the best chess engines were 2600, OliThink is about 2400. A ELO difference of 100 points is already a sign for a much stronger player. A typical club player (like myself) has ELO 1700, which still is much better than a player, who just knows the rules with 800.

Meanwhile things have changed, Stockfish and Leelo Chess Zero do have ELO ratings about 3600. Stockfish does have a lot of chess knowledge from the programming, but Leelo Chess Zero does not. It just has the power of eGPUs 😉

Hello dear community, I have some information:
First of all the Lc0 benchmarks for different OpenCL GPUs:

Intel UHD 630: 325
Radeon Pro 560X: 1301
Radeon RX 570: 2346
GTX 1080 Ti: 4453

The interesting part is, when you combine two or more CPUs, there is the interesting fact, that not every combination yields better result, especially when more than two GPUs are being used. The best result, so far is:

GTX 1080 Ti + Radeon RX 570 (OpenCL): 6908
Benchmark final time 10.0896s calculating 6908.48 nodes per second.

This looks promising but actually it isn't. Because I was successful in compiling Leelo Chess Zero with CUDA Support for MacOSX and the result is amazing:

./lc0 benchmark --backend=cudnn
GTX 1080 Ti (CUDA): 40938
Benchmark final time 10.0133s calculating 40937.8 nodes per second.

It means that CUDA is about 9 times (900%) more effective in calculation than OpenCL. It also means, that you lose a lot of calculation power once you upgrade you MacOSX to Mojave or Catalina, because they do no support CUDA.

PS: Any Combination with other OpenCL Gpus lowers the score:

./lc0 benchmark --backend=multiplexing --backend-opts='(backend=cudnn),(gpu=2,backend=opencl)'
Benchmark final time 10.1303s calculating 31887.1 nodes per second.

@itsage: Is Chess an interesting application for the eGPU community?

Here is a match between LC0 (0.42.1) and Stockfish 11. It's quite typical, LC0 dominating midgame but doing strange things in the endgame. This happens when you have hardly any chess knowledge. The match was win since about move 35, but still neede 160 more moves :/

1. e4 {+0.30/6} c5 {+0.11/26 9} 2. Nf3 {+0.29/10 2.9} d6 {+0.06/25 3} 3. d4
{+0.30/9 2.8} cxd4 {+0.10/26 0.1} 4. Nxd4 {+0.33/9 2.1} Nf6 {+0.11/23 4} 5.
Nc3 {+0.30/9 1.0} a6 {+0.17/20 0.3} 6. f3 {+0.27/8 9} e5 {+0.53/21 1.7} 7.
Nb3 {+0.25/10 0.6} Be7 {+0.44/23 4} 8. Be3 {+0.27/10 2.5} Be6
{+0.44/23 1.3} 9. Qd2 {+0.28/11 1.8} Nbd7 {+0.65/23 4} 10. g4 {+0.33/9 5}
O-O {+0.14/26 9} 11. g5 {+0.42/10 4} Nh5 {+0.00/22 1.7} 12. O-O-O
{+0.44/9 1.0} Nb6 {+0.00/27 5} 13. Rg1 {+0.46/8 10} Rc8 {+0.00/30 1.4} 14.
Kb1 {+0.46/8 1.8} g6 {+0.00/27 0.9} 15. h4 {+0.51/8 3} Qc7 {+0.00/28 13}
16. Qf2 {+0.35/10 4} Nc4 {+0.00/27 0.3} 17. Bxc4 {+0.34/11 0.7} Bxc4
{-0.15/30 3} 18. Bb6 {+0.51/11 1.8} Qb8 {+0.08/25 0.3} 19. Qd2 {+0.46/8 7}
Be6 {+0.00/29 0.1} 20. Nd5 {+0.45/9 2.9} Bxd5 {+0.00/30 1.6} 21. Qxd5
{+0.45/10 1.3} Rc6 {+0.00/29 2.1} 22. Be3 {+0.43/9 1.2} Qc8 {+0.00/24 1.6}
23. Na5 {+0.46/14 5} Rxc2 {+0.27/25 1.7} 24. Qxb7 {+0.48/14 1.2} Rh2
{+0.00/27 4} 25. Qxc8 {+0.49/10 5} Rxc8 {+0.13/26 1.8} 26. Rc1 {+0.44/13 3}
Rxc1+ {+0.00/23 1.6} 27. Rxc1 {+0.50/15 1.3} Rxh4 {+0.00/26 0.5} 28. Rc8+
{+0.51/14 4} Kg7 {+0.00/30 0.1} 29. Nc6 {+0.41/12 2.5} Rh1+ {+0.00/32 0.1}
30. Kc2 {+0.14/11 3} Rh2+ {+0.00/32 0.1} 31. Kb1 {+0.20/14 7} Bf8
{+0.00/35 0.1} 32. a4 {+0.60/10 4} Nf4 {-1.16/29 17} 33. b4 {+0.98/11 1.6}
Re2 {-0.82/28 5} 34. Bb6 {+1.73/9 4} Ne6 {-2.54/30 5} 35. a5 {+2.05/10 2.9}
Re1+ {-3.00/26 1.6} 36. Kc2 {+2.04/10 6} Re2+ {-1.79/27 1.7} 37. Kc3
{+1.90/10 2.6} Ra2 {-1.86/29 0.1} 38. Kb3 {+1.82/12 2.8} Ra1 {-1.99/29 0.8}
39. Kb2 {+1.80/12 2.4} Ra4 {-2.01/30 0.8} 40. Be3 {+2.04/10 6} f6
{-1.91/31 0.2} 41. Ra8 {+2.32/8 2.4} fxg5 {-1.87/29 0.8} 42. Rxa6
{+2.11/10 1.9} h5 {-1.92/27 1.4} 43. Rb6 {+2.12/10 5} h4 {-2.24/25 4} 44.
a6 {+2.22/9 0.4} Kh6 {-3.86/27 12} 45. Bg1 {+1.92/8 4} h3 {-3.32/26 2.8}
46. Bh2 {+2.35/9 1.4} g4 {-4.80/27 2.7} 47. fxg4 {+2.59/8 1.8} Kg5
{-4.56/25 1.1} 48. a7 {+2.75/8 2.6} Kxg4 {-5.49/26 4} 49. Rb7 {+2.66/8 3}
Rxa7 {-5.49/25 1.0} 50. Rxa7 {+2.57/10 1.1} Kf3 {-5.30/23 0.2} 51. Rf7+
{+2.63/8 5} Kg2 {-2.67/23 0.7} 52. Rf6 {+2.69/9 0.5} Nf4 {-2.66/26 0.8} 53.
b5 {+2.31/10 2.1} Kxh2 {-2.70/24 1.1} 54. Rxf8 {+2.33/12 0.9} Kg3
{-2.82/27 2.0} 55. b6 {+2.57/8 2.0} h2 {-3.17/25 0.1} 56. Rh8 {+2.58/9 0.9}
Nh5 {-3.50/29 6} 57. b7 {+2.71/10 0.5} h1=Q {-3.17/26 0.6} 58. b8=Q
{+2.66/9 0.4} Qd1 {-3.06/22 0.4} 59. Qb3+ {+3.59/7 1.5} Qxb3+
{-3.97/27 2.3} 60. Kxb3 {+2.86/8 0.3} Kf3 {-4.30/29 6} 61. Rg8
{+3.40/8 0.7} Nf4 {-3.16/22 0.7} 62. Ne7 {+3.81/7 0.7} Kxe4 {-3.59/22 0.7}
63. Nxg6 {+5.04/6 0.7} d5 {-2.99/22 1.2} 64. Nxf4 {+13.97/5 1.0} exf4
{-1000.35/30 2.2} 65. Kc2 {+16.51/6 0.7} Ke3 {-1000.24/39 0.9} 66. Kd1
{+20.03/5 0.3} Kd3 {-1000.25/36 0.6} 67. Rf8 Ke3 {-1000.26/38 0.3} 68. Ke1
{+66.41/2 0.3} d4 {-1000.24/40 0.5} 69. Rg8 {+44.43/4 0.2} d3
{-1000.23/39 0.5} 70. Re8+ {+19.96/4 0.2} Kd4 {-1000.21/42 0.2} 71. Kd2
{+34.95/3 0.1} f3 {-1000.20/43 0.3} 72. Rf8 {+59.54/2 0.2} Ke4
{-1000.19/41 0.2} 73. Rf7 {+71.96/4 0.1} f2 {-1000.18/40 0.3} 74. Rxf2 Kd4
{-1000.17/41 0.5} 75. Rf4+ {+102.32/4 0.1} Ke5 {-1000.15/42 0.7} 76. Rh4
{+104.86/3 0.1} Kf5 {-1000.15/41 0.4} 77. Kxd3 Ke5 {-1000.14/39 0.2} 78.
Rc4 Ke6 {-1000.13/39 0.4} 79. Ke2 Kd5 {-1000.15/40 0.4} 80. Ra4 Ke6
{-1000.14/39 0.3} 81. Rb4 Kd7 {-1000.14/38 0.3} 82. Re4 {+84.20/2 0.2} Kd6
{-1000.14/39 0.4} 83. Rd4+ Ke6 {-1000.13/39 0.3} 84. Kf1 {+65.32/2 0.2} Ke5
{-1000.16/36 0.3} 85. Rd3 Kf4 {-1000.15/38 0.2} 86. Rd5 Ke4
{-1000.15/38 0.4} 87. Rd8 Ke3 {-1000.15/39 0.3} 88. Rh8 Ke4
{-1000.15/41 0.7} 89. Kg2 Kd4 {-1000.14/39 0.3} 90. Rh1 Kc5
{-1000.15/38 0.3} 91. Rb1 Kc4 {-1000.15/37 0.3} 92. Rc1+ {+30.98/1 0.2} Kd3
{-1000.15/39 0.3} 93. Kh2 Kd4 {-1000.16/34 0.3} 94. Rb1 Kd5
{-1000.16/36 0.4} 95. Rf1 Ke5 {-1000.16/39 0.5} 96. Rf3 Ke4
{-1000.17/42 1.6} 97. Rf2 {+20.74/4 0.1} Kd5 {-1000.16/40 0.7} 98. Kh3 Ke4
{-1000.16/38 0.3} 99. Rb2 Kd4 {-1000.15/37 0.4} 100. Ra2 Kc5
{-1000.15/41 1.3} 101. Rf2 {+14.61/6 17} Kd4 {-1000.15/52 4} 102. Rf8
{+14.12/6 15} Ke5 {-1000.14/56 4} 103. Ra8 {+14.78/6 13} Kd6
{-1000.14/55 4} 104. Ra4 {+13.67/6 12} Kd5 {-1000.15/52 4} 105. Rh4
{+12.89/6 11} Kd6 {-1000.15/52 4} 106. Kg3 {+13.05/6 10} Kd5
{-1000.14/53 4} 107. Kg4 {+12.57/5 9} Ke5 {-1000.14/53 2.6} 108. Rh7
{+12.56/5 8} Ke4 {-1000.14/58 3} 109. Rg7 {+13.79/5 7} Kd5 {-1000.13/64 3}
110. Kf5 {+13.77/5 6} Kc4 {-1000.12/65 0.1} 111. Ke4 {+14.22/5 6} Kc5
{-1000.11/62 3} 112. Rg3 {+14.23/5 5} Kc6 {-1000.12/52 4} 113. Ke5
{+14.21/5 5} Kc5 {-1000.11/63 3} 114. Rc3+ {+13.11/5 4} Kb4
{-1000.10/70 0.2} 115. Kd4 {+16.98/5 4} Kb5 {-1000.09/70 0.1} 116. Kd5
{+15.85/5 4} Kb6 {-1000.08/66 0.1} 117. Rc5 {+16.49/5 3} Kb7
{-1000.07/74 0.1} 118. Kd6 {+18.79/5 3} Kb6 {-1000.06/86 0.1} 119. Rf5
{+25.37/5 2.7} Kb7 {-1000.05/154 2.9} 120. Rb5+ {+128.00/6 2.2} Ka6
{-1000.04/245 0.2} 121. Kc5 {+128.00/5 2.1} Ka7 {-53.93/1 0.1} 122. Kc6
{+128.00/5 2.1} Ka6 {-1000.03/245 0.1} 123. Rd5 {+128.00/3 1.9} Ka7
{-1000.03/1 0.1} 124. Kc7 {+128.00/3 1.7} Ka6 {-1000.02/245 0.1} 125. Rg5
{+128.00/2 1.6} Ka7 {-1000.01/1 0.1} 126. Rg6 {+128.00/2 1.4} Ka8
{-1000.01/1 0.1} 127. Ra6# {+128.00/2 0.6}
{Xboard adjudication: Checkmate} 1-0

@oliverb, Thank you for sharing your chess engine, OliThink and how effective external graphics can be with the right application. I wish I had learnt to play chess to understand this better.

@oliverb, Do professional chess players read those numbers and understand the play in real time?

Posted by: @jangoloti_1

@oliverb, Do professional chess players read those numbers and understand the play in real time?

 

Yes, they do. I am a non-professional mediocre chess player (about low club level) and I understand the numbers. Of course it's easier and quicker to use a graphical interface like xboard/winboard where you can replay the match. This is the position after the 34. Nd6 when it was very clear, that black was in serious trouble.

The most interesting part is, that black (Stockfish 11) is actually the best chess program/player ever, but self learning Leela Chess Zero combined with eGPU power looks even stronger.

It is a scandal that the feud between Apple and nVidia causes that CUDA on MacOSX is dead after High Sierra.
They spit on science and progress.