lc0-win-20180512-cuda90-cudnn712-00

[shrapnel]

No, it is completely unnecessary. You can ignore that warning too. It relates only to CUDA items for VS, and nothing to do with Leela. There is now a 9.2 version, so by all means install it and use the latest LC0 built for it.

lc0-win-20180519-cuda92-cudnn714-00.7z

Thank You, that was very helpful.
Which CUDA Installer should I use, the exe (network) or exe (local) ?

[Albert Silver]

No, it is completely unnecessary. You can ignore that warning too. It relates only to CUDA items for VS, and nothing to do with Leela. There is now a 9.2 version, so by all means install it and use the latest LC0 built for it.

lc0-win-20180519-cuda92-cudnn714-00.7z

Thank You, that was very helpful.
Which CUDA Installer should I use, the exe (network) or exe (local) ?

I used local. It is a big install, with some unnecessary items such as the Visual Studio extras. Don't worry if it complains it can't find it. Once all is installed, including cuDNN, you need to locate three DLLs in all, and copy them in the LCZero folder:

cublas64_92.dll
cudart64_92.dll
cudnn64_7.dll

After that just enjoy LC0 in https://crem.xyz/lc0/ Be sure to configure its UCI options, including the path of the weights file. For example, mine shows: C:\Users\Albert\Chess\LCZero\weights.txt.

Feel free to use the optimized settings described in a previous post in this thread.

[shrapnel]

I used local. It is a big install, with some unnecessary items such as the Visual Studio extras. Don't worry if it complains it can't find it. Once all is installed, including cuDNN, you need to locate three DLLs in all, and copy them in the LCZero folder:

cublas64_92.dll
cudart64_92.dll
cudnn64_7.dll

After that just enjoy LC0 in https://crem.xyz/lc0/ Be sure to configure its UCI options, including the path of the weights file. For example, mine shows: C:\Users\Albert\Chess\LCZero\weights.txt.

Feel free to use the optimized settings described in a previous post in this thread.

Thanks again.
Unfortunately, CUDNN shows only versions for Windows 7 and Windows 10.
I'm using Windows 8.1.
Which should I try first ?

[Laskos]

Wow, thanks for the tips, I am new to my settings with GPU. Indeed, with these settings, with ID237 and with May 19 LC0 CUDA, against Zurichess Neuchatel, a modern AB engine, LC0 performs at 3120 Elo level in CCRL 40/4' conditions (still not very many games), a rating I have never seen even remotely with any LC0 (master or CUDA) in these TC conditions. The same GTX 1060 6GB card as yours and 2 i7 threads. Thanks also for the revised WAC and the result with PUCT values on it. You really seem to hit a sweet point, as tactics is the most important cause of its misses. Time and again, a tactical blunder gives a half a point or a full point away.

I actually have interesting news for settings based on a discussion in Discord, the official LC0 channel. Someone ran CLOP on it to finetune all the settings to find optimal results. He came up with slowmover (the time management setting) best at 2.2-2.3, cPUCT at about 2.8, and FPU Reduction at -0.08 (yes, negative value). I have not tested this myself, but am sharing:

I misread and slowmover should be about 2.75

I am getting some pretty incredible results. With ID237 and these settings, CUDA version beats convincingly in 10 rapid games at 10'+ 10'' TC Houdini 1.5a (3170 CCRL 40/4' Elo):

Code: Select all

Score of LC0_GPU_CUDA vs Houdini 1.5a: 4 - 1 - 5 [0.650]
Elo difference: 107.54 +/- 166.19

10 of 10 games finished.

Achieved performance is about 3280 CCRL 40/4' Elo points. Here is the first game, which was a strangulation win by LCO over Houdini 1.5a. Quite a seen.

[Event "My Tournament"]
[Site "?"]
[Date "2018.05.20"]
[Round "1"]
[White "LC0_GPU_CUDA"]
[Black "Houdini 1.5a"]
[Result "1-0"]
[FEN "rnbqkbnr/1pp2ppp/p3p3/3p4/3PP3/8/PPPN1PPP/R1BQKBNR w KQkq - 0 1"]
[PlyCount "147"]
[SetUp "1"]
[TimeControl "600+10"]

1. Bd3 {+0.12/2 21s} c5 {-0.18/19 29s} 2. dxc5 {+0.16/2 14s} Bxc5 {-0.15/19 18s}
3. Ngf3 {+0.18/2 8.4s} Nf6 {-0.08/20 31s} 4. O-O {+0.22/2 11s}
O-O {-0.09/20 30s} 5. a3 {+0.22/2 32s} a5 {-0.09/19 78s} 6. b3 {+0.21/2 28s}
Nc6 {-0.03/18 23s} 7. Bb2 {+0.17/2 15s} Qe7 {-0.10/19 24s} 8. Qe2 {+0.44/2 24s}
Rd8 {-0.07/18 19s} 9. e5 {+0.39/2 18s} Nd7 {-0.08/19 21s} 10. c3 {+0.40/2 23s}
Bb6 {-0.03/18 31s} 11. Bc2 {+0.41/2 12s} Bc7 {0.00/19 63s} 12. c4 {+0.45/2 12s}
b6 {+0.03/18 32s} 13. cxd5 {+0.43/2 22s} exd5 {0.00/17 8.9s}
14. Rfe1 {+0.44/2 6.7s} Bb7 {+0.01/18 18s} 15. Rad1 {+0.53/2 15s}
Nc5 {-0.01/18 15s} 16. Nf1 {+0.54/2 13s} Ba6 {-0.05/18 14s}
17. Qe3 {+1.02/2 15s} Rac8 {-0.29/17 36s} 18. Ng3 {+1.26/2 16s}
Ne6 {-0.38/18 14s} 19. Bb1 {+1.42/2 10.0s} Qc5 {-0.41/18 14s}
20. Qd2 {+1.40/2 8.5s} Ne7 {-0.54/17 23s} 21. Nd4 {+1.85/2 16s}
Ng6 {-0.86/19 64s} 22. Nh5 {+2.12/2 12s} Nxd4 {-1.14/17 38s}
23. Bxd4 {+2.06/2 6.0s} Qe7 {-1.25/18 41s} 24. Bf5 {+2.02/2 7.2s}
a4 {-1.30/18 35s} 25. bxa4 {+2.65/2 16s} Ra8 {-1.38/17 32s} 26. e6 {+2.40/2 13s}
f6 {-1.81/19 32s} 27. Qc2 {+2.53/2 7.4s} Rdc8 {-1.91/18 12s}
28. Qb1 {+3.26/2 14s} Nf8 {-2.30/18 24s} 29. Qb2 {+3.31/2 8.5s}
Qd6 {-2.35/18 20s} 30. g3 {+3.98/2 14s} Bd8 {-1.99/16 0.052s}
31. Bc3 {+4.10/2 10s} Be7 {-2.73/17 9.1s} 32. Bb4 {+4.29/2 9.5s}
Qd8 {-3.02/18 17s} 33. Nf4 {+4.65/2 12s} Bb7 {-4.54/17 18s}
34. Be4 {+4.90/2 11s} Qe8 {-5.21/16 15s} 35. Bxd5 {+5.02/2 12s}
Rc7 {-5.46/17 11s} 36. Qd4 {+5.50/2 12s} Qc8 {-5.71/17 13s}
37. Bxb7 {+6.69/2 14s} Rxb7 {-4.93/18 8.6s} 38. Bxe7 {+7.03/2 8.5s}
Rxe7 {-4.79/17 0.003s} 39. Nd5 {+7.21/2 5.0s} Rxe6 {-5.33/19 16s}
40. Ne7+ {+7.28/2 4.1s} Rxe7 {-5.33/18 0.001s} 41. Rxe7 {+7.26/2 2.5s}
Qc5 {-5.18/20 9.2s} 42. Qxc5 {+7.68/2 15s} bxc5 {-4.97/16 4.3s}
43. Rb1 {+8.17/2 11s} c4 {-4.96/18 12s} 44. Rbb7 {+9.45/2 15s}
Rxa4 {-5.46/19 18s} 45. Ra7 {+10.77/2 21s} Rxa7 {-4.36/17 4.3s}
46. Rxa7 {+10.79/2 7.8s} Ne6 {-4.06/19 10s} 47. Ra4 {+11.20/2 9.9s}
Kf7 {-4.44/20 17s} 48. Rxc4 {+11.90/2 14s} Ke7 {-4.55/20 11s}
49. a4 {+11.86/2 13s} Kd6 {-4.71/19 15s} 50. a5 {+11.86/2 11s}
Nc7 {-4.92/19 11s} 51. h4 {+12.33/2 27s} f5 {-4.89/18 13s}
52. Kg2 {+13.45/2 20s} g6 {-5.14/18 8.5s} 53. Rxc7 {+17.94/2 15s}
Kxc7 {-5.57/17 0.077s} 54. Kf3 {+18.45/2 9.8s} Kd6 {-11.67/26 12s}
55. Kf4 {+19.72/3 14s} h6 {-12.88/26 15s} 56. h5 {+20.91/3 14s}
gxh5 {-16.48/25 13s} 57. Kxf5 {+21.01/2 9.0s} Kc6 {-14.65/23 0s}
58. Kg6 {+21.47/3 22s} Kd5 {-14.87/18 4.1s} 59. f4 {+25.58/2 23s}
Kc6 {-19.15/20 17s} 60. f5 {+25.80/2 19s} Kb5 {-M26/25 11s}
61. f6 {+25.06/2 18s} h4 {-M24/27 12s} 62. gxh4 {+26.80/2 12s} Ka6 {-M20/29 14s}
63. f7 {+29.38/5 17s} Kxa5 {-M18/28 0.087s} 64. f8=Q {+30.21/3 18s}
Kb5 {-M16/26 0.004s} 65. Qxh6 {+38.01/3 15s} Kc4 {-M16/26 2.0s}
66. Kf5 {+42.44/4 18s} Kd3 {-M14/24 0.39s} 67. Ke5 {+46.84/3 21s}
Kc4 {-M12/22 0.15s} 68. Ke4 {+48.19/4 22s} Kc5 {-M12/22 0.11s}
69. Qa6 {+53.79/2 22s} Kb4 {-M10/1 0s} 70. Kd4 {+61.35/3 13s} Kb3 {-12.55/1 0s}
71. Kd3 {+76.44/3 17s} Kb2 {-M6/16 0s} 72. Qa4 {+112.92/3 13s} Kb1 {-M4/14 0s}
73. Kc3 {+127.83/3 9.7s} Kc1 {-M2/1 0s} 74. Qa1# {+128.00/2 9.8s, White mates}
1-0

With these settings, there are much fewer tactical blunders, and the eval graph of LC0 looks much smoother.
I tried ID317 against Zurichess Neuchatel in some 40 games at 1'+ 1'', it achieved a 3200 Elo performance, even higher than ID237. It also beat in these conditions Komodo 12 MCTS on 2 threads (3050 CCRL), achieving again a performance of 3200 Elo points.

Next I will play 10 rapid games at 10'+ 10'' against a year and a half old top engine Komodo 10.2 (3369 CCRL 40/4' Elo points). LC0 will most probably lose heavily the match, but maybe I will see one win over a top recent engine Komodo 10.2.

With CUDA 9.2 engine (new CUDA and cuDNN installed) the results seem bad, it seems the engine is buggy or my setup is wrong. So, for now I am using CUDA 9.0 and its latest engine with NN319. I am risking a bit in the match against Komodo 10.2, as NN237 might still be the best, but we will see, 10 games will take some 7 hours in these conditions.

EDIT: indeed, it seems NN319 blunders tactically a bit more often than NN237, in an almost won positions against Komodo 10.2, it blundered first the win, and several moves later blundered again, losing. Maybe Komodo 10.2 is tactically too strong. The almost won position was achieved slowly and positionally by LC0, which was pretty amazing to see against such a good positional engine like Komodo.

[shrapnel]

Interesting results, Mr Laskos.
Which version of Windows are you using ?

[Laskos]

Interesting results, Mr Laskos.
Which version of Windows are you using ?

Win8.1.
I am using Win7 cuDNN, as I am not sure Win10 one will work here.

[pohl4711]

No, it is completely unnecessary. You can ignore that warning too. It relates only to CUDA items for VS, and nothing to do with Leela. There is now a 9.2 version, so by all means install it and use the latest LC0 built for it.

lc0-win-20180519-cuda92-cudnn714-00.7z

Thank You, that was very helpful.
Which CUDA Installer should I use, the exe (network) or exe (local) ?

I used local. It is a big install, with some unnecessary items such as the Visual Studio extras. Don't worry if it complains it can't find it. Once all is installed, including cuDNN, you need to locate three DLLs in all, and copy them in the LCZero folder:

cublas64_92.dll
cudart64_92.dll
cudnn64_7.dll

After that just enjoy LC0 in https://crem.xyz/lc0/ Be sure to configure its UCI options, including the path of the weights file. For example, mine shows: C:\Users\Albert\Chess\LCZero\weights.txt.

Feel free to use the optimized settings described in a previous post in this thread.

There is no need to install CUDA. Just download the installer file (.exe) and unpack it (with 7z-packer for example). Then use the search-funcion of windows-explorer to locate the dll (cublas64_92.dll, cudart64_92.dll) in the huge folder, which you got, and copy them in your Leela Cuda folder. Then you can delete the installer and the unpacked folders and files.

[shrapnel]

There is no need to install CUDA. Just download the installer file (.exe) and unpack it (with 7z-packer for example). Then use the search-funcion of windows-explorer to locate the dll (cublas64_92.dll, cudart64_92.dll) in the huge folder, which you got, and copy them in your Leela Cuda folder. Then you can delete the installer and the unpacked folders and files.

Interesting shortcut.
So you mean to say that no need to install CUDA, but CUDNN has to be installed ?
And what do you mean by "Leela CUDA" Folder ? I have LCZero Folder.

[shrapnel]

There is no need to install CUDA. Just download the installer file (.exe) and unpack it (with 7z-packer for example). Then use the search-funcion of windows-explorer to locate the dll (cublas64_92.dll, cudart64_92.dll) in the huge folder, which you got, and copy them in your Leela Cuda folder. Then you can delete the installer and the unpacked folders and files.

Interesting shortcut.
So you mean to say that no need to install CUDA, but CUDNN has to be installed ?
And what do you mean by "Leela CUDA" Folder ? I have LCZero Folder.
Also all the binaries at https://crem.xyz/lc0/ contain executable files named lc0-win.exe.
If I unzip them to the LCZero Folder, Windows asks if each new lco-win.exe should over-write the other ??
Confusing.

[pohl4711]

There is no need to install CUDA. Just download the installer file (.exe) and unpack it (with 7z-packer for example). Then use the search-funcion of windows-explorer to locate the dll (cublas64_92.dll, cudart64_92.dll) in the huge folder, which you got, and copy them in your Leela Cuda folder. Then you can delete the installer and the unpacked folders and files.

Interesting shortcut.
So you mean to say that no need to install CUDA, but CUDNN has to be installed ?
And what do you mean by "Leela CUDA" Folder ? I have LCZero Folder.

CUDNN is a download of a zip-file, only. Just unzip and get the file cudnn64_7.dll out of it.
And it doesnt matter, how you name the folder, in which your LCZero is. I name it Leela Cuda. You can name it "cold_beer", if you like.
Just put the 3 dll-files in that folder, where your LCZero exe-file is, which you downloaded here https://crem.xyz/lc0/
Then load the network-file. Unpack it. Name the unpacked file "weights.txt" and put in the same folder.
With these 5 files, it works.

No installation of anything is required. Works with Cuda 9.0. I did not try Cuda 9.2 until now.