TalkChess.com

"There is no single "best net" for Leela but there a few worth recommending for various purposes. The most important consideration in choosing a net is picking the right size for your hardware and time controls (TC). In general, if you have a weak GPU (or no GPU at all) and you want to only spend milliseconds per move, then you want a smaller net that evaluates positions more quickly, i.e. higher NPS (nodes per second). On the other hand, if you have an RTX card (or several) and you want to run an analysis on a single move for hours at a time, then the quality of the evaluation is more important than the speed and a larger (but slower) net is probably going to work best."

https://github.com/LeelaChessZero/lc0/w ... ts-for-Lc0

Anything better than 128x10.T17.6-swa-20000 for CPU Leela?

Any way to use GPU Leela in a GPU that doesn't support CUDA or OpenCL?

Hi

I am searching a weightfile for a promising challenger against Stockfish.
Since 18 days, I started my tests again, 300 games, 5m+3s , ponder on.
4 matches already done, and I hope I do not loose my motivation and a lot of tests will follow.
https://docs.google.com/spreadsheets/d/ ... =609748570
I decided to use contempt=0 for Stockfish, for Leela its much harder to beat SF with contempt 0.

No challenger found yet ... merry Christmas to all
C.K.

Hugo wrote: ↑Tue Dec 24, 2019 11:08 am Hi

I am searching a weightfile for a promising challenger against Stockfish.
Since 18 days, I started my tests again, 300 games, 5m+3s , ponder on.
4 matches already done, and I hope I do not loose my motivation and a lot of tests will follow.
https://docs.google.com/spreadsheets/d/ ... =609748570
I decided to use contempt=0 for Stockfish, for Leela its much harder to beat SF with contempt 0.

No challenger found yet ...

Code: Select all

   # PLAYER                          :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W     D    L  D(%)
   1 Lc0 v0.22.0-leele11.1           :       0     14   150.0     300  50.0      50   16   268   16  89.3
   2 Stockfish 031219 64 BMI2-T13    :       0   ----   609.0    1200  50.8      62   91  1036   73  86.3
   3 Lc0 v0.22.0-sergio1207          :      -2     15   149.0     300  49.7      62   15   268   17  89.3
   4 Lc0 v0.22.0-T40.B4-160          :      -6     15   147.5     300  49.2      74   24   247   29  82.3
   5 Fat Fritz                       :     -13     15   144.5     300  48.2     ---   18   253   29  84.3

merry Christmas to all
C.K.

I have a little issue. Being Ponder ON of an 8 core CPU with hyperthreading, the 13 threads of SF will slow down the 2 threads of Lc0. Windows distributes 13 threads on 16 hyperthreads that they all slow down individually compared to the 8 threads used on 8 cores. Lc0 with its 2 threads won't slow down SF significantly in Ponder ON mode. But LcO needs 2 full blown cores, not 2 crippled hyperthreads. My opinion is that you abandon Ponder ON.

Hi

when the latest test is done, I will make a comparison with several constellations. For example let SF analyse endless with 13 T and compare Lc0 speed with 2 threads. same with Lc0 standalone.
Right now I could not see any difference in Lc0 with two cpu or two threads. But I did not look at it when other threads are bussy.

C.K.

Laskos wrote: ↑Tue Dec 24, 2019 11:49 am I have a little issue. Being Ponder ON of an 8 core CPU with hyperthreading, the 13 threads of SF will slow down the 2 threads of Lc0. Windows distributes 13 threads on 16 hyperthreads that they all slow down individually compared to the 8 threads used on 8 cores. Lc0 with its 2 threads won't slow down SF significantly in Ponder ON mode. But LcO needs 2 full blown cores, not 2 crippled hyperthreads. My opinion is that you abandon Ponder ON.

Hugo wrote: ↑Tue Dec 24, 2019 3:11 pm Hi

when the latest test is done, I will make a comparison with several constellations. For example let SF analyse endless with 13 T and compare Lc0 speed with 2 threads. same with Lc0 standalone.
Right now I could not see any difference in Lc0 with two cpu or two threads. But I did not look at it when other threads are bussy.

C.K.

Laskos wrote: ↑Tue Dec 24, 2019 11:49 am I have a little issue. Being Ponder ON of an 8 core CPU with hyperthreading, the 13 threads of SF will slow down the 2 threads of Lc0. Windows distributes 13 threads on 16 hyperthreads that they all slow down individually compared to the 8 threads used on 8 cores. Lc0 with its 2 threads won't slow down SF significantly in Ponder ON mode. But LcO needs 2 full blown cores, not 2 crippled hyperthreads. My opinion is that you abandon Ponder ON.

Yes, the difference, if any, should be very small. I just prefer to keep 2 full cores for Lc0, but it's maybe I am being too pedantic.

Ovyron wrote: ↑Tue Dec 24, 2019 6:56 am Anything better than 128x10.T17.6-swa-20000 for CPU Leela?

Any way to use GPU Leela in a GPU that doesn't support CUDA or OpenCL?

Early GPU-programs were purely written in "shader" form. A Pixel shader transformed data in a 2d-grid into another 2d-grid. Its probably possible to write a CNN interpreter in that form, but it'd be way more difficult than just using CUDA or OpenCL.

There is a sergio1572 and a LS up to 12.2 (It is recommended to use Leelenstein cuda/opencl binary and tournament search params instead of default, like used at ccc or Navs stream). You might also consider testing terminator 2-136 24block. Fat Fritz and regular T40 networks are definitely not the strongest now.

Hugo wrote: ↑Tue Dec 24, 2019 11:08 am I am searching a weightfile for a promising challenger against Stockfish.

Hugo wrote: ↑Tue Dec 24, 2019 11:08 am Hi

I am searching a weightfile for a promising challenger against Stockfish.
Since 18 days, I started my tests again, 300 games, 5m+3s , ponder on.
4 matches already done, and I hope I do not loose my motivation and a lot of tests will follow.
https://docs.google.com/spreadsheets/d/ ... =609748570
I decided to use contempt=0 for Stockfish, for Leela its much harder to beat SF with contempt 0.

No challenger found yet ...

Code: Select all

   # PLAYER                          :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W     D    L  D(%)
   1 Lc0 v0.22.0-leele11.1           :       0     14   150.0     300  50.0      50   16   268   16  89.3
   2 Stockfish 031219 64 BMI2-T13    :       0   ----   609.0    1200  50.8      62   91  1036   73  86.3
   3 Lc0 v0.22.0-sergio1207          :      -2     15   149.0     300  49.7      62   15   268   17  89.3
   4 Lc0 v0.22.0-T40.B4-160          :      -6     15   147.5     300  49.2      74   24   247   29  82.3
   5 Fat Fritz                       :     -13     15   144.5     300  48.2     ---   18   253   29  84.3

merry Christmas to all
C.K.

The latest Lc0 61xxxx nets are now matching or beating the latest SF on my hardware. Even at 5s per move.

Laskos wrote: ↑Tue Dec 24, 2019 11:49 am

Hugo wrote: ↑Tue Dec 24, 2019 11:08 am Hi

I am searching a weightfile for a promising challenger against Stockfish.
Since 18 days, I started my tests again, 300 games, 5m+3s , ponder on.
4 matches already done, and I hope I do not loose my motivation and a lot of tests will follow.
https://docs.google.com/spreadsheets/d/ ... =609748570
I decided to use contempt=0 for Stockfish, for Leela its much harder to beat SF with contempt 0.

No challenger found yet ...

Code: Select all

   # PLAYER                          :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W     D    L  D(%)
   1 Lc0 v0.22.0-leele11.1           :       0     14   150.0     300  50.0      50   16   268   16  89.3
   2 Stockfish 031219 64 BMI2-T13    :       0   ----   609.0    1200  50.8      62   91  1036   73  86.3
   3 Lc0 v0.22.0-sergio1207          :      -2     15   149.0     300  49.7      62   15   268   17  89.3
   4 Lc0 v0.22.0-T40.B4-160          :      -6     15   147.5     300  49.2      74   24   247   29  82.3
   5 Fat Fritz                       :     -13     15   144.5     300  48.2     ---   18   253   29  84.3

merry Christmas to all
C.K.

I have a little issue. Being Ponder ON of an 8 core CPU with hyperthreading, the 13 threads of SF will slow down the 2 threads of Lc0. Windows distributes 13 threads on 16 hyperthreads that they all slow down individually compared to the 8 threads used on 8 cores. Lc0 with its 2 threads won't slow down SF significantly in Ponder ON mode. But LcO needs 2 full blown cores, not 2 crippled hyperthreads. My opinion is that you abandon Ponder ON.

Ponder on works with Lc0 just fine. For a AMD the best solution to not impact Lc0 is to give SF all 16 logical cores but run SF in a lower priority. And give Lc0 what it wants normally. The slow down for Lc0 will be zero, and SF will only be impacted by the exact percentage of utilization when Lc0 is run on the system by itself. In my case Lc0 uses about 4% of the CPUs. And the exact slowdown to SF with Ponder on playing Lc0 is 4%. This is meaningless with a 12% overclock to all cores. I have not tested this with a Intel CPU, but give it a try. Let the system do the core splitting.