LCZero: Progress and Scaling. Relation to CCRL Elo

[Laskos]

Now, with this post of Carlos Canavessi http://www.talkchess.com/forum/viewtopi ... 9&start=44 and the release of CPU and GPU versions, I have something odd happening. Although my GPU is weak, probably some 10 times slower than the best GTX cards, I didn't expect GPU-engine to perform so badly compared to CPU-only engine. In other GPU intensive tasks, it helps greatly over my Haswell four core i7 CPU + integrated GPU. Is LCZero using efficiently GPU?

Old LCZero shows the following:

Code: Select all

Generated 1924 moves
Detecting residual layers...v1...64 channels...6 blocks.
Initializing OpenCL.
Detected 1 OpenCL platforms.
Platform version: OpenCL 1.2 CUDA 9.1.75
Platform profile: FULL_PROFILE
Platform name:    NVIDIA CUDA
Platform vendor:  NVIDIA Corporation
Device ID:     0
Device name:   GeForce GT 730
Device type:   GPU
Device vendor: NVIDIA Corporation
Device driver: 388.13
Device speed:  1400 MHz
Device cores:  2 CU
Device score:  1112
Selected platform: NVIDIA CUDA
Selected device: GeForce GT 730
with OpenCL 1.2 capability.
Loaded existing SGEMM tuning.
Wavefront/Warp size: 32
Max workgroup size: 1024
Max workgroup dimensions: 1024 1024 64
BLAS Core: Haswell

New GPU LCZero shows the following (identical):

Code: Select all

Generated 1924 moves
Detecting residual layers...v1...64 channels...6 blocks.
Initializing OpenCL.
Detected 1 OpenCL platforms.
Platform version: OpenCL 1.2 CUDA 9.1.75
Platform profile: FULL_PROFILE
Platform name:    NVIDIA CUDA
Platform vendor:  NVIDIA Corporation
Device ID:     0
Device name:   GeForce GT 730
Device type:   GPU
Device vendor: NVIDIA Corporation
Device driver: 388.13
Device speed:  1400 MHz
Device cores:  2 CU
Device score:  1112
Selected platform: NVIDIA CUDA
Selected device: GeForce GT 730
with OpenCL 1.2 capability.
Loaded existing SGEMM tuning.
Wavefront/Warp size: 32
Max workgroup size: 1024
Max workgroup dimensions: 1024 1024 64
BLAS Core: Haswell

New CPU version shows:

Code: Select all

Generated 1924 moves
Detecting residual layers...v1...64 channels...6 blocks.
BLAS Core: Haswell

Using the old LCZero (actually the one from 1-2 days ago), and new CPU and GPU Windows binaries, I got the following:

Code: Select all

Games Completed = 300 of 300 (Avg game length = 97.535 sec)
Settings = RR/64MB/1000ms per move/M 500cp for 3 moves, D 140 moves/EPD:C:\LittleBlitzer\2moves_v1.epd(32000)
Time = 8427 sec elapsed, 0 sec remaining
 1.  LCZero Old               	53.0/200	45-139-16  	(L: m=139 t=0 i=0 a=0)	(D: r=6 i=9 f=1 s=0 a=0)	(tpm=972.7 d=11.29 nps=138)
 2.  LCZero GPU               	54.0/200	45-137-18  	(L: m=137 t=0 i=0 a=0)	(D: r=9 i=8 f=1 s=0 a=0)	(tpm=972.9 d=11.29 nps=98)
 3.  LCZero CPU               	193.0/200	190-4-6  	(L: m=4 t=0 i=0 a=0)	(D: r=3 i=3 f=0 s=0 a=0)	(tpm=960.7 d=14.66 nps=498)





   # PLAYER        : RATING  ERROR    POINTS  PLAYED     (%)   CFS(next)
   1 LCZero CPU    :  387.5   81.1     193.0     200    96.5     100    
   2 LCZero GPU    : -192.1   51.5      54.0     200    27.0      54    
   3 LCZero Old    : -195.4   52.6      53.0     200    26.5     ---    

White advantage = 0.00 +/- 28.99
Draw rate (equal opponents) = 12.69 % +/- 2.87

CPU LCZero has some 5 times higher NPS than GPU LCZero, and is much stronger. The old and the new Windows GPU version seem to perform comparably. With the old LCZero, I got

LCZero Old versus Zurichess Appenzeller (Elo 1821 CCRL):
1s/move:
9.0 - 91.0



Now, with the CPU version, I get a much better result:

LCZero CPU (one core) versus Zurichess Appenzeller (Elo 1821 CCRL):
1s/move:
33.0 - 67.0

Which gives a rating of about 1700 CCRL Elo points for CPU version. I have to revise my plot for progress and extrapolation, using stronger on my PC CPU version. From 27th to 28th of March, LCZero advanced very little, maybe it is a sign of plateauing? If not, the extrapolation on one core (for CPU version on my Haswell) would look as following:

[Werewolf]

When you analyse in Arena, how many rollouts / sec are you getting on with the CPU and GPU versions, please?

I'm getting 2000 - 5000 on GPU

[Laskos]

When you analyse in Arena, how many rollouts / sec are you getting on with the CPU and GPU versions, please?

I'm getting 2000 - 5000 on GPU

Interesting. I am getting about 150 on GPU and 600 on CPU.

[Werewolf]

When you analyse in Arena, how many rollouts / sec are you getting on with the CPU and GPU versions, please?

I'm getting 2000 - 5000 on GPU

Interesting. I am getting about 150 on GPU and 600 on CPU.

That does seem low for the GPU, though it might just be your card. The Titan V is at least 3x faster than my card which is eye opening.

By the way, do you know why there is such big regression on this site from time to time?

http://lczero.org/networks

[Jhoravi]

By the way, do you know why there is such big regression on this site from time to time?

http://lczero.org/networks

It was caused by a bug discussed here.
https://groups.google.com/forum/#!topic ... vf9p3biLRk
Now it's fixed.

[Werewolf]

Amazing. I'll wait for the training before doing tests on the new one.

[petero2]

do you know the elo of the strongest ai chess software today ?

http://www.computerchess.org.uk/ccrl/40 ... t_all.html

Look at 1 core performances, as LCZero is on one core too. Stockfish 9 on one core is about 3500 CCRL Elo.

I meant software which uses neural networks.is it giraffe ?

I think Texel Gi is the strongest that can run without a GPU/TPU. In my tests it is around 2820 on the CCRL 40/4 rating scale, so about 400 elo above Giraffe.

[Milos]

Now, with this post of Carlos Canavessi http://www.talkchess.com/forum/viewtopi ... 9&start=44 and the release of CPU and GPU versions, I have something odd happening. Although my GPU is weak, probably some 10 times slower than the best GTX cards, I didn't expect GPU-engine to perform so badly compared to CPU-only engine. In other GPU intensive tasks, it helps greatly over my Haswell four core i7 CPU + integrated GPU. Is LCZero using efficiently GPU?

Old LCZero shows the following:

Code: Select all

Generated 1924 moves
Detecting residual layers...v1...64 channels...6 blocks.
Initializing OpenCL.
Detected 1 OpenCL platforms.
Platform version: OpenCL 1.2 CUDA 9.1.75
Platform profile: FULL_PROFILE
Platform name:    NVIDIA CUDA
Platform vendor:  NVIDIA Corporation
Device ID:     0
Device name:   GeForce GT 730
Device type:   GPU
Device vendor: NVIDIA Corporation
Device driver: 388.13
Device speed:  1400 MHz
Device cores:  2 CU
Device score:  1112
Selected platform: NVIDIA CUDA
Selected device: GeForce GT 730
with OpenCL 1.2 capability.
Loaded existing SGEMM tuning.
Wavefront/Warp size: 32
Max workgroup size: 1024
Max workgroup dimensions: 1024 1024 64
BLAS Core: Haswell

New GPU LCZero shows the following (identical):

Code: Select all

Generated 1924 moves
Detecting residual layers...v1...64 channels...6 blocks.
Initializing OpenCL.
Detected 1 OpenCL platforms.
Platform version: OpenCL 1.2 CUDA 9.1.75
Platform profile: FULL_PROFILE
Platform name:    NVIDIA CUDA
Platform vendor:  NVIDIA Corporation
Device ID:     0
Device name:   GeForce GT 730
Device type:   GPU
Device vendor: NVIDIA Corporation
Device driver: 388.13
Device speed:  1400 MHz
Device cores:  2 CU
Device score:  1112
Selected platform: NVIDIA CUDA
Selected device: GeForce GT 730
with OpenCL 1.2 capability.
Loaded existing SGEMM tuning.
Wavefront/Warp size: 32
Max workgroup size: 1024
Max workgroup dimensions: 1024 1024 64
BLAS Core: Haswell

New CPU version shows:

Code: Select all

Generated 1924 moves
Detecting residual layers...v1...64 channels...6 blocks.
BLAS Core: Haswell

Using the old LCZero (actually the one from 1-2 days ago), and new CPU and GPU Windows binaries, I got the following:

Code: Select all

Games Completed = 300 of 300 (Avg game length = 97.535 sec)
Settings = RR/64MB/1000ms per move/M 500cp for 3 moves, D 140 moves/EPD:C:\LittleBlitzer\2moves_v1.epd(32000)
Time = 8427 sec elapsed, 0 sec remaining
 1.  LCZero Old               	53.0/200	45-139-16  	(L: m=139 t=0 i=0 a=0)	(D: r=6 i=9 f=1 s=0 a=0)	(tpm=972.7 d=11.29 nps=138)
 2.  LCZero GPU               	54.0/200	45-137-18  	(L: m=137 t=0 i=0 a=0)	(D: r=9 i=8 f=1 s=0 a=0)	(tpm=972.9 d=11.29 nps=98)
 3.  LCZero CPU               	193.0/200	190-4-6  	(L: m=4 t=0 i=0 a=0)	(D: r=3 i=3 f=0 s=0 a=0)	(tpm=960.7 d=14.66 nps=498)





   # PLAYER        : RATING  ERROR    POINTS  PLAYED     (%)   CFS(next)
   1 LCZero CPU    :  387.5   81.1     193.0     200    96.5     100    
   2 LCZero GPU    : -192.1   51.5      54.0     200    27.0      54    
   3 LCZero Old    : -195.4   52.6      53.0     200    26.5     ---    

White advantage = 0.00 +/- 28.99
Draw rate (equal opponents) = 12.69 % +/- 2.87

CPU LCZero has some 5 times higher NPS than GPU LCZero, and is much stronger. The old and the new Windows GPU version seem to perform comparably. With the old LCZero, I got

LCZero Old versus Zurichess Appenzeller (Elo 1821 CCRL):
1s/move:
9.0 - 91.0



Now, with the CPU version, I get a much better result:

LCZero CPU (one core) versus Zurichess Appenzeller (Elo 1821 CCRL):
1s/move:
33.0 - 67.0

Which gives a rating of about 1700 CCRL Elo points for CPU version. I have to revise my plot for progress and extrapolation, using stronger on my PC CPU version. From 27th to 28th of March, LCZero advanced very little, maybe it is a sign of plateauing? If not, the extrapolation on one core (for CPU version on my Haswell) would look as following:

This is toatally normal since your GPU is really crap (pardon my French ).
GT 730 is like 15 times slower than 1080 in terms of dot product ops, and 1080 is just slightly faster than 5960X with all 8 cores so I guess 4 cores Haswell (4770K is it?) being 5 times faster than 730 is really spot on .

[lkaufman]

do you know the elo of the strongest ai chess software today ?

http://www.computerchess.org.uk/ccrl/40 ... t_all.html

Look at 1 core performances, as LCZero is on one core too. Stockfish 9 on one core is about 3500 CCRL Elo.

I meant software which uses neural networks.is it giraffe ?

I think Texel Gi is the strongest that can run without a GPU/TPU. In my tests it is around 2820 on the CCRL 40/4 rating scale, so about 400 elo above Giraffe.

I'd like to ask a somewhat different question, what is the elo of the strongest non-private chess software (using CPU, not GPU) today that uses Monte-Carlo Tree Search, regardless of its evaluation function?

[CMCanavessi]

do you know the elo of the strongest ai chess software today ?

http://www.computerchess.org.uk/ccrl/40 ... t_all.html

Look at 1 core performances, as LCZero is on one core too. Stockfish 9 on one core is about 3500 CCRL Elo.

I meant software which uses neural networks.is it giraffe ?

I think Texel Gi is the strongest that can run without a GPU/TPU. In my tests it is around 2820 on the CCRL 40/4 rating scale, so about 400 elo above Giraffe.

I'd like to ask a somewhat different question, what is the elo of the strongest non-private chess software (using CPU, not GPU) today that uses Monte-Carlo Tree Search, regardless of its evaluation function?

Scorpio maybe?