lc0-win-20180512-cuda90-cudnn712-00

[Albert Silver]

Just got GTX 1060 6GB GPU, performed some tests. NPS from starting position to depth 25-26 on ID305 with v0.10 GPU master is about 820. Latest CUDA LC0 from here https://crem.xyz/lc0/ is about 4000. People were saying that this build is buggy, and is barely performing above master non-CUDA v0.10 version. I found a huge advantage of CUDA version, almost to the level of its NPS advantage.

So are you saying it has been fixed? Is there a document explaining how to change over (I also have a 1060)

It seems no more buggy than the master itself v0.10 is . So, it seems its NPS are almost fair to compare to master NPS.
You have to decompress the weights file and put in the same folder with CUDA lc0 (renaiming it weights.txt). Then you have to install CUDA 9.0, after that cuDNN 7.1 for CUDA 9.0. After that, take the installed dll libraries and put them into lc0 folder. Should work.

I cannot speak for others, but the last version of May 9 (I did not try later ones), is a beast: much faster and much stronger than the normal version.

[Laskos]

So are you saying it has been fixed? Is there a document explaining how to change over (I also have a 1060)

It seems no more buggy than the master itself v0.10 is . So, it seems its NPS are almost fair to compare to master NPS.
You have to decompress the weights file and put in the same folder with CUDA lc0 (renaiming it weights.txt). Then you have to install CUDA 9.0, after that cuDNN 7.1 for CUDA 9.0. After that, take the installed dll libraries and put them into lc0 folder. Should work.

I cannot speak for others, but the last version of May 9 (I did not try later ones), is a beast: much faster and much stronger than the normal version.

Hmmm...
In direct play with ID312, CUDA version beats Master v0.10 by 50-100 Elo points in CCRL 40/4' conditions. But against Zurichess Neuchatel (2894 CCRL 40/4' Elo points), a modern AB engine, LC0 Master v0.10 in CCRl 40/4' conditions seems to be 2970, while CUDA version seems weaker, about 2900. Only 100 games each match, but still quite relevant. It seems that the direct match Master-CUDA is not highletening the tactics so much, but against an AB engine like Zurichess, CUDA version blunders tactically more often (just my impression). What CPUCT are you using for CUDA version? I used the default 1.20.

[Werewolf]

It seems no more buggy than the master itself v0.10 is . So, it seems its NPS are almost fair to compare to master NPS.
You have to decompress the weights file and put in the same folder with CUDA lc0 (renaiming it weights.txt). Then you have to install CUDA 9.0, after that cuDNN 7.1 for CUDA 9.0. After that, take the installed dll libraries and put them into lc0 folder. Should work.

I cannot speak for others, but the last version of May 9 (I did not try later ones), is a beast: much faster and much stronger than the normal version.

Hmmm...
In direct play with ID312, CUDA version beats Master v0.10 by 50-100 Elo points in CCRL 40/4' conditions. But against Zurichess Neuchatel (2894 CCRL 40/4' Elo points), a modern AB engine, LC0 Master v0.10 in CCRl 40/4' conditions seems to be 2970, while CUDA version seems weaker, about 2900. Only 100 games each match, but still quite relevant. It seems that the direct match Master-CUDA is not highletening the tactics so much, but against an AB engine like Zurichess, CUDA version blunders tactically more often (just my impression). What CPUCT are you using for CUDA version? I used the default 1.20.

How does switching to the Nvidia dlls (CUDA) do anything except affect speed? Seems strange to hear it is tactically weaker etc.

[Laskos]

I cannot speak for others, but the last version of May 9 (I did not try later ones), is a beast: much faster and much stronger than the normal version.

Hmmm...
In direct play with ID312, CUDA version beats Master v0.10 by 50-100 Elo points in CCRL 40/4' conditions. But against Zurichess Neuchatel (2894 CCRL 40/4' Elo points), a modern AB engine, LC0 Master v0.10 in CCRl 40/4' conditions seems to be 2970, while CUDA version seems weaker, about 2900. Only 100 games each match, but still quite relevant. It seems that the direct match Master-CUDA is not highletening the tactics so much, but against an AB engine like Zurichess, CUDA version blunders tactically more often (just my impression). What CPUCT are you using for CUDA version? I used the default 1.20.

How does switching to the Nvidia dlls (CUDA) do anything except affect speed? Seems strange to hear it is tactically weaker etc.

No, it is effectively a different engine. And as I understood, it was pretty buggy and improvisional. Seems better now, but still something is not right.

[Albert Silver]

Hmmm...
In direct play with ID312, CUDA version beats Master v0.10 by 50-100 Elo points in CCRL 40/4' conditions. But against Zurichess Neuchatel (2894 CCRL 40/4' Elo points), a modern AB engine, LC0 Master v0.10 in CCRl 40/4' conditions seems to be 2970, while CUDA version seems weaker, about 2900. Only 100 games each match, but still quite relevant. It seems that the direct match Master-CUDA is not highletening the tactics so much, but against an AB engine like Zurichess, CUDA version blunders tactically more often (just my impression). What CPUCT are you using for CUDA version? I used the default 1.20.

How does switching to the Nvidia dlls (CUDA) do anything except affect speed? Seems strange to hear it is tactically weaker etc.

No, it is effectively a different engine. And as I understood, it was pretty buggy and improvisional. Seems better now, but still something is not right.

I had trouble with the May 12 version, just advising. As to PUCT values, I set FPU Reduction to 0.0, and cPUCT to 3.0. Also, the strongest NN, if you wish, is NN237, by a very large margin.

[Laskos]

How does switching to the Nvidia dlls (CUDA) do anything except affect speed? Seems strange to hear it is tactically weaker etc.

No, it is effectively a different engine. And as I understood, it was pretty buggy and improvisional. Seems better now, but still something is not right.

I had trouble with the May 12 version, just advising. As to PUCT values, I set FPU Reduction to 0.0, and cPUCT to 3.0. Also, the strongest NN, if you wish, is NN237, by a very large margin.

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.

[Albert Silver]

No, it is effectively a different engine. And as I understood, it was pretty buggy and improvisional. Seems better now, but still something is not right.

I had trouble with the May 12 version, just advising. As to PUCT values, I set FPU Reduction to 0.0, and cPUCT to 3.0. Also, the strongest NN, if you wish, is NN237, by a very large margin.

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:

[Albert Silver]

I had trouble with the May 12 version, just advising. As to PUCT values, I set FPU Reduction to 0.0, and cPUCT to 3.0. Also, the strongest NN, if you wish, is NN237, by a very large margin.

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

[shrapnel]

I installed CUDA 9.0 and cuDNN 7.12, with the libraries in CUDA lc0 folder.

Is it necessary to have Visual Studio installed before installing CUDA 9.0 ?
I installed some free version of Visual Studio but I'm getting the message "nVidia Installer failed" while trying to install CUDA 9.0.

[Albert Silver]

I installed CUDA 9.0 and cuDNN 7.12, with the libraries in CUDA lc0 folder.

Is it necessary to have Visual Studio installed before installing CUDA 9.0 ?
I installed some free version of Visual Studio but I'm getting the message "nVidia Installer failed" while trying to install CUDA 9.0.

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