Leela ver.0.25.0

[corres]

I wait for lc0 ver.0.25.0 but without bugs in optionsparser.cc and other minor issues.

[Modern Times]

Yes I've found rc1 pretty unstable, lots of pop-up error messages.

[BeyondCritics]

Try to drill it down as much as possible. Then head over to https://github.com/LeelaChessZero/lc0/i ... ed-descand report the issue. Thank you in advance.

[corres]

Try to drill it down as much as possible. Then head over to https://github.com/LeelaChessZero/lc0/i ... ed-descand report the issue. Thank you in advance.

There are some developers of Leela among the members of TC maybe they would frequently read TC posts to get more information from us.

[Modern Times]

Try to drill it down as much as possible. Then head over to https://github.com/LeelaChessZero/lc0/i ... ed-descand report the issue. Thank you in advance.

Yes I should have taken some screenshots, will do so if I encounter it again. The particular usage scenario was not a common one for me - Winboard and polyglot, ponder on, tournament play. I have not encountered any issues with cutechess and ponder off engine matches. Perfectly stable there.

[corres]

I did not find notable difference in gaming power and in speed between Leela ver.0.23.3 and Leela ver.0.25
For test I used RTX 2080 Ti and J13B.2-188 with the same parameters.
I think it is not a good trick to make displaying the net "name" in hexadecimal numbers. Because it is disturbing.
It would be better displaying the natural name (for eg. J13B.2-188) of the used net. Or displaying nothing.

[corres]

I did not find notable difference in gaming power and in speed between Leela ver.0.23.3 and Leela ver.0.25
For test I used RTX 2080 Ti and J13B.2-188 with the same parameters.
I think it is not a good trick to make displaying the net "name" in hexadecimal numbers. Because it is disturbing.
It would be better displaying the natural name (for eg. J13B.2-188) of the used net. Or displaying nothing.

Using more than one GPU it seems to me Leela ver.0.25.0 utilize the multiplexing backend more effective.
Net J13B.2-188, old (not kiudee) parameters and GPUs 1 x RTX 2080Ti + 2 x RTX 2060 (my TRIAD) go nodes 10000000 test results:
Leela ver.0.23.3
max.speed = 46.9 kn/sec (Depth=20, Time=160sec, Nodes=7246 kn)
Leela ver.0.25.0
max.speed = 52.7 kn/sec (Depth=20, Time=130sec, Nodes= 6640 kn)
The enhancement is in max.speed ~13 %.

Note
There are some parameters in Leela ver.0.25.0 what are not in ver.0.23.3 so they may also influence the result.

[corres]

To check the enhancement I made for my DUAL (= 2 x RTX 2060, Backend=Multiplexing, Threads=4, NNCachesize=20000000, others are old default) two version of Leela: ver.0.23.3 and ver.0.25.0.
The net was JHorthos T40B.4-160 (20x256).
Results (after go nodes 10000000)
DUAL with Leela ver.0.25.0
max.speed = 55.1 kn/sec (Depth=21, Time=183 sec, Nodes=9705 kn)
DUAL with Leela ver.0.23.3
max.speed = 49.8 kn/sec (Depth=19, Time=123 sec Nodes=5980 kn))
The enhancement is ~10 %.
For the interest I made test with my RTX 2080 Ti OC TURBO Threads=2, net was T40B.4-160, using Leela ver.0.25.0
max.speed = 50.0 kn/sec (Depth=18, Time=100sec, Nodes=4970 kn)

Note
The version 0.25.0 is a good Leela, it works well and its compilation was flawless.

[zullil]

For the interest I made test with my RTX 2080 Ti OC TURBO Threads=2, net was T40B.4-160, using Leela ver.0.25.0
max.speed = 50.0 kn/sec (Depth=18, Time=100sec, Nodes=4970 kn)

Note
The version 0.25.0 is a good Leela, it works well and its compilation was flawless.

Did you mean JHorthos T40B.4-106?

What do you get for the built-in Lc0 benchmark, which now runs 34 positions instead of just the starting position?

Code: Select all

$ ./lc0 benchmark
       _
|   _ | |
|_ |_ |_| v0.26.0-dev+git.ad4b5f2 built Apr 30 2020
Found pb network file: ./T40B.1-106
Creating backend [cudnn-auto]...
Switching to [cudnn-fp16]...
CUDA Runtime version: 10.2.0
Cudnn version: 7.6.5
Latest version of CUDA supported by the driver: 10.2.0
GPU: GeForce RTX 2080 Ti
GPU memory: 10.7534 Gb
GPU clock frequency: 1635 MHz
GPU compute capability: 7.5

Position: 1/34 rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1
Benchmark time 26ms, 3 nodes, 375 nps, move e2e4
Benchmark time 30ms, 5 nodes, 454 nps, move e2e4
Benchmark time 34ms, 9 nodes, 600 nps, move e2e4
Benchmark time 37ms, 18 nodes, 1000 nps, move e2e4
Benchmark time 41ms, 27 nodes, 1227 nps, move d2d4
Benchmark time 45ms, 43 nodes, 1653 nps, move d2d4
Benchmark time 49ms, 75 nodes, 2500 nps, move e2e4
Benchmark time 56ms, 96 nodes, 2594 nps, move e2e4
Benchmark time 65ms, 125 nodes, 2717 nps, move e2e4
Benchmark time 68ms, 143 nodes, 2918 nps, move e2e4
Benchmark time 72ms, 177 nodes, 3339 nps, move e2e4
Benchmark time 79ms, 245 nodes, 4016 nps, move e2e4
Benchmark time 87ms, 329 nodes, 4838 nps, move e2e4
Benchmark time 98ms, 434 nodes, 5493 nps, move e2e4
Benchmark time 105ms, 474 nodes, 5511 nps, move e2e4
Benchmark time 113ms, 507 nodes, 5393 nps, move e2e4
Benchmark time 120ms, 563 nodes, 5574 nps, move e2e4
Benchmark time 127ms, 587 nodes, 5435 nps, move e2e4
Benchmark time 131ms, 592 nodes, 5238 nps, move e2e4
Benchmark time 135ms, 647 nodes, 5577 nps, move e2e4
Benchmark time 143ms, 724 nodes, 5838 nps, move e2e4
Benchmark time 150ms, 767 nodes, 5854 nps, move e2e4
Benchmark time 161ms, 923 nodes, 6500 nps, move e2e4
Benchmark time 172ms, 1083 nodes, 7078 nps, move e2e4
Benchmark time 187ms, 1341 nodes, 7982 nps, move e2e4
Benchmark time 201ms, 1635 nodes, 8983 nps, move e2e4
Benchmark time 212ms, 1938 nodes, 10041 nps, move e2e4
Benchmark time 215ms, 2043 nodes, 10423 nps, move e2e4
Benchmark time 222ms, 2245 nodes, 11059 nps, move e2e4
Benchmark time 233ms, 2518 nodes, 11766 nps, move e2e4
Benchmark time 243ms, 2784 nodes, 12428 nps, move e2e4
Benchmark time 248ms, 2965 nodes, 12947 nps, move e2e4
Benchmark time 260ms, 3319 nodes, 13771 nps, move e2e4
Benchmark time 271ms, 3701 nodes, 14686 nps, move e2e4
Benchmark time 273ms, 3796 nodes, 14944 nps, move e2e4
Benchmark time 282ms, 3991 nodes, 15174 nps, move e2e4
Benchmark time 292ms, 4314 nodes, 15802 nps, move e2e4
Benchmark time 302ms, 4608 nodes, 16282 nps, move e2e4
Benchmark time 316ms, 5141 nodes, 17251 nps, move e2e4
Benchmark time 334ms, 5788 nodes, 18374 nps, move e2e4
Benchmark time 353ms, 6419 nodes, 19218 nps, move e2e4
Benchmark time 361ms, 6701 nodes, 19593 nps, move e2e4
Benchmark time 372ms, 7098 nodes, 20107 nps, move e2e4
Benchmark time 437ms, 9258 nodes, 22148 nps, move e2e4
Benchmark time 457ms, 9885 nodes, 22517 nps, move e2e4
Benchmark time 475ms, 10443 nodes, 22901 nps, move e2e4
Benchmark time 499ms, 11083 nodes, 23041 nps, move e2e4
Benchmark time 507ms, 11412 nodes, 23337 nps, move e2e4
Benchmark time 529ms, 12265 nodes, 24049 nps, move e2e4
Benchmark time 552ms, 13131 nodes, 24636 nps, move e2e4
Benchmark time 585ms, 14409 nodes, 25457 nps, move e2e4
Benchmark time 618ms, 15751 nodes, 26295 nps, move e2e4
Benchmark time 703ms, 19262 nodes, 28119 nps, move e2e4
Benchmark time 749ms, 20524 nodes, 28115 nps, move e2e4
Benchmark time 773ms, 21353 nodes, 28319 nps, move e2e4
Benchmark time 1090ms, 35522 nodes, 33167 nps, move e2e4
Benchmark time 1166ms, 38897 nodes, 33911 nps, move e2e4
Benchmark time 1426ms, 50723 nodes, 36050 nps, move e2e4
Benchmark time 2029ms, 79591 nodes, 39597 nps, move e2e4
Benchmark time 2591ms, 106907 nodes, 41565 nps, move e2e4
Benchmark time 3168ms, 134763 nodes, 42795 nps, move e2e4
Benchmark time 3348ms, 142730 nodes, 42874 nps, move e2e4
Benchmark time 5896ms, 274908 nodes, 46776 nps, move e2e4
Benchmark time 6444ms, 300847 nodes, 46824 nps, move e2e4
Benchmark time 9937ms, 474516 nodes, 47843 nps, move e2e4
Benchmark time 10000ms, 477184 nodes, 47809 nps, move e2e4
bestmove e2e4

[33 additional positions snipped]

Code: Select all

===========================
Total time (ms) : 340967
Nodes searched  : 16865944
Nodes/second    : 49465

[corres]

For the interest I made test with my RTX 2080 Ti OC TURBO Threads=2, net was T40B.4-160, using Leela ver.0.25.0
max.speed = 50.0 kn/sec (Depth=18, Time=100sec, Nodes=4970 kn)

Note
The version 0.25.0 is a good Leela, it works well and its compilation was flawless.

Did you mean JHorthos T40B.4-106?

No.
I used T40B.4-160 net (TCEC16)
Leela bench is a new possibility, so I can not compare the bench result to my earlier ones.
If you want, use it. I do not need it.