Leela: Single and Dual RTX 2060 - Stockfish dev: 4 threads

[corres]

I made some tests in which Leela ver.0.21.1 run on single and dual RTX 2060 against Stockfish dev. (190331) run on 4 threads.
Parameters of tests:
CPU: Ryzen 7 1800x 8X4000 MHz
GPU: 1xRTX 2060 (SINGLE GPU) and 2xRTX 2060 (DUAL GPU)
During tests GPUs used 2 Threads (SINGLE GPU) and 2x2 Threads (DUAL GPU) and Stockfish used 4 Threads in every cases,
Thinking time: 1min + 2sec/move
For tests I used my opening set (50 opening position)
The games were played with reversed color too.
No syzygy or other TB was used.
GPUs used its dafault parameters, excluding backend=cudnn-fp16 instead of backend=cudnn.
NET file: 41800

Test 1
SINGLE GPU: 1xRTX 2060 against Stockfish dev.190331
Leela : Stockfish dev = 15 : 8 (77 draw), 100 games

Test 2
DUAL GPU: 2xRTX 2060 against Stockfish dev.190331
Leela : Stockfish dev = 19 : 5 (76 draw), 100 games

[Hugo]

thx for this test. Quite intresting match.
what is the difference in Elo please?
Would be intresting if you give it more time. Leela often switches the main line in the first 20 seconds.
So it could be stronger with more time ( e.g. 4+2 )

C.K.

[corres]

thx for this test. Quite intresting match.
what is the difference in Elo please?
Would be intresting if you give it more time. Leela often switches the main line in the first 20 seconds.
So it could be stronger with more time ( e.g. 4+2 )

C.K.

Next time I will run tests with enhanced parameters of Leela.
After then I will use longer TC with the better parameters.
I also should like your test results.
And what motherboard and CPU is in your PC?

[Hugo]

And what motherboard and CPU is in your PC?

Its a bit complicated.
the main machine is a supermicro server with two E5-2697A v4 cpus and 256GB RAM on board.
For my tests , I remove one cpu and some RAM and put it on an Asus Mainboard, so I get a 16 core machine with 64 GB RAM and 1xRTX 2060
plus
A 16 core machine with 128 GB RAM and 1xRTX 2060.

C.K.

[corres]

And what motherboard and CPU is in your PC?

Its a bit complicated.
the main machine is a supermicro server with two E5-2697A v4 cpus and 256GB RAM on board.
For my tests , I remove one cpu and some RAM and put it on an Asus Mainboard, so I get a 16 core machine with 64 GB RAM and 1xRTX 2060
plus
A 16 core machine with 128 GB RAM and 1xRTX 2060.

C.K.

Thaks for the answer.
Where is RTX 2070?

[Hugo]

Thaks for the answer.
Where is RTX 2070?

i7 6700K machine with 32Mb RAm.

its my machine for online chess-gaming.
I am passionate engine player @playchess and infinitychess.

C.K.

[corres]

Thaks for the answer.
Where is RTX 2070?

i7 6700K machine with 32Mb RAm.

its my machine for online chess-gaming.
I am passionate engine player @playchess and infinitychess.

C.K.

And what a success you use Leela and Stockfish?
Maybe you utilize the other two machines for CC games?

[corres]

In this test I made a competition between Single RTX 2060 and Dual RTX 2060
System: Ryzen7 1800x 8x4000 MHz, 2xRTX 2060
Syzygy 6 men database
Played 50 opening position with reversed colors
Time control: 1 min + 2 sec / move
Lc0 version 0.21.1
Lc0 options:
Backend=cudnn-fp16 (Multiplexing for Dual)
NNCacheSize=2000000
Other parameters are Default
Result:
Lc0 DUAL GPU v0.21.1 : Lc0-fp16 v0.21.1 = 15 : 0 (85 draw) 100 games
The Elo difference is about +50 for DUAL GPU.
This is very normal value in the case of AB engines also if the speed one engine is 2 times higher than the other.
You can download the pgn:
http://wikisend.com/download
File ID: 411772
Password: leela0

[corres]

In this test I made a new competition between the Dual RTX 2060 and Stockfish 190331
System: Ryzen7 1800x 8x4000 MHz, 2xRTX 2060, Windows 10 (2019 oct)
Endgame databases: Syzygy 6 men + Nalimov 5 men database
The Nalimov databases used by the GUI only (Fritz 11)
Played 50 opening position with reversed colors
Time control: 1 min + 2 sec / move
Used Threads by Stockfish: 4
Net file: 42102
Lc0 version 0.21.1
Lc0 options:
Backend: Multiplexing
Backendoptions: backend=cudnn-fp16,(gpu=0),(gpu=1)
NNCacheSize: 2000000
Threads: 4
Other parameters are Default
Result:
Lc0 DUAL GPU v0.21.1 : Stockfish190331 = 30 : 4 (66 draw) 100 games
The Elo difference is about +90 for DUAL GPU.

You can download the pgn:
http://wikisend.com/download
File ID: 998268
Password: leela0

[corres]

In this test I made a new competition between the Dual RTX 2060 and Stockfish 190331
System: Ryzen7 1800x 8x4000 MHz, 2xRTX 2060, Windows 10 (2019 oct)
Endgame databases: Syzygy 6 men + Nalimov 5 men database
The Nalimov databases used by the GUI only (Fritz 11)
Played 50 opening position with reversed colors
Time control: 1 min + 2 sec / move
Used Threads by Stockfish: 4
Net file: 42102
Lc0 version 0.21.1
Lc0 options:
Backend: Multiplexing
Backendoptions: backend=cudnn-fp16,(gpu=0),(gpu=1)
NNCacheSize: 2000000
Threads: 4
Other parameters are Default
Result:
Lc0 DUAL GPU v0.21.1 : Stockfish190331 = 30 : 4 (66 draw) 100 games
The Elo difference is about +90 for DUAL GPU.

You can download the pgn:
http://wikisend.com/download
File ID: 998268
Password: leela0

I repeated the above test with enhanced Time Control:
Time Control= 1 min + 20 sec / move
All other parameters are the above
Result:
Lc0 DUAL GPU v.0.21.1 : Stockfish190331 = 23 : 2 (75 draw) 100 games.
The Elo difference is about +75 for DUAL GPU.
So the difference decreased with about 15 Elo. This loss is too small and to get an exact value we would made much more games.
There is an another result: Enhancing the Time Control we can see lesser decided games.

You can download the pgn:
http://wikisend.com/download
File ID: 499520
Password: leela0