LC0 @ low power GPU

[yurikvelo]

There is quite a lot information about running Leela at TDP>120+W GPU (RTX class)

I wonder what is good low-power (TDP<=65...75W) GPU for Leela

Is GTX-1650 best in this sub-75 class (as of 2020-Q2)?

What is SF-dev ratio for LC0@GTX-1650, how many NPS need SF-dev to be roughly equal?

[mehmet123]

What is SF-dev ratio for LC0@GTX-1650, how many NPS need SF-dev to be roughly equal?

Ratio is changed due to Lc0 nets and time controls. Lc0 is more weaker at short time controls

[yurikvelo]

Ratio is changed due to Lc0 nets and time controls. Lc0 is more weaker at short time controls

so how many NPS needed?

[mehmet123]

Let's look at some tests:

Program Elo + - Games Score Av.Op. Draws

1 Stockfish 200320 x64 bmi2[c0] : 2403 22 21 100 51.0 % 2397 90.0 %
2 Lc0 384x30-t60-3200 [Default] : 2397 21 22 100 49.0 % 2403 90.0 %


Individual statistics:

1 Stockfish 200320 [c0] : 2403 100 (+ 6,= 90,- 4), 51.0 %

Lc0 3200 [Default] : 100 (+ 6,= 90,- 4), 51.0 %

2 Lc0 3200 [Default] : 2397 100 (+ 4,= 90,- 6), 49.0 %

Stockfish 200320 [c0] : 100 (+ 4,= 90,- 6), 49.0 %


Arena Gui, 10 min + 2 sec tc, 8 moves pgn opening book,
Stockfish 200320 : 9 threads, contempt:0, Core i7 9750h (6 core /12 threads
Lc0 3200:Gtx 1650

(Stockfish 200320 ~11000 kn/s, Lc0 3200 ~2.5 kn/s)

http://www.mediafire.com/file/01amaumub ... 5.pgn/file


Program Elo + - Games Score Av.Op. Draws

1 Stockfish 050420 [c0] : 2407 9 9 2000 51.9 % 2393 65.1 %
2 Lc0 591226 [Kayra3] : 2402 13 13 1000 49.3 % 2407 64.4 %
3 Lc0 591226 [Default/Kiudee] : 2385 13 13 1000 46.8 % 2407 65.8 %


Individual statistics:

1 Stockfish 050420 [c0] : 2407 2000 (+388,=1302,-310), 52.0 %
Lc0 591226 [Kayra3] : 1000 (+185,=644,-171), 50.7 %
Lc0 591226 [Default/Kiudee] 1000 (+203,=658,-139), 53.2 %

2 Lc0 591226 [Kayra3] : 2402 1000 (+171,=644,-185), 49.3 %
Stockfish 050420 [c0] : 1000 (+171,=644,-185), 49.3 %

http://www.mediafire.com/file/basb9q7rh ... 0.pgn/file

3 Lc0 591226 [Default/Kiudee] : 2385 1000 (+139,=658,-203), 46.8 %
Stockfish 050420 [c0] : 1000 (+139,=658,-203), 46.8 %

Arena Gui, 30'' + 0.5'' sec tc, Lc0 v.24.0 rc-1, Balsa Opening Book (5 moves)
Stockfish: Core i7- 9750h (2 core), Contempt 0
Lc0:Nvdia Gtx 1650
Hash:512 Mb , No Tablebase, Ponder:Off


First Test : Lc0/Stockfish Kn/s : 2.5 / 11000
Second Test:Lc0/Stockfish Kn/s : 24 / 3900

Every Lc0 nets has different speeds :
At my tests : kn/s of 384x 30 net:~ 2.5 kn/s, kn/s of 320x24 net:~ 4 kn/s, kn/s of 256x24 net:~ 6 kn/s, kn/s of 128 x10 net:~ 24 kn/s
Lc0 3200 (384x 30 net), 591226 (128x10 net)

At 10 min + 2 sec time control match the performance of Lc0 384x30-t60-3200 is close to Stockfish dev.
But at short time control matches Stockfish can easily beats Lc0 384x30-t60-3200.
All my test was made with GTX 1650 mobile GPU. The performance of Nvidia GTX 1650 at desktop is `% 10 better than Nvidia GTX 1650 at laptop.

As you can see there isn't a standart ratio for Lc0 vs Stockfish matches. It changes according to the Lc0 nets and time control.

[yurikvelo]

As you can see there isn't a standart ratio for Lc0 vs Stockfish matches. It changes according to the Lc0 nets and time control.

lc0-0.25 + SV-20b-t40-1541 @ GTX 1650 vs 8 MNPS SF-latest (nps at starting position) at TC=10'+2"

[mehmet123]

As you can see there isn't a standart ratio for Lc0 vs Stockfish matches. It changes according to the Lc0 nets and time control.

lc0-0.25 + SV-20b-t40-1541 @ GTX 1650 vs 8 MNPS SF-latest (nps at starting position) at TC=10'+2"

128x 10 nets: Fastest nets but they have less knowledge
256x 24 nets: Slower than 128x 10 nets but they have more knowledge according to 128 x10 nets
320x 24 nets: Slower than 256x 24 nets but they have more knowledge according to 256x 24 nets
384 x30 nets: Slower than 320x 24 nets but they have more knowledge according to 320x 24 nets
512x 40 nets: It has been training by Sergio Vieri, but for now it's not strong enough.

128x 10 nets are more suitable for weak hardwares/ less time control matches
..
..
384 x30 nets are more suitable for strong hardwares (GPU) /long time control matches.


At TC=10'+2" matches (GTX 1650) 320x 24 nets (Lc0 636 xxx nets) and 384 x30 nets (Lc0 3010, Lc0 3180, Lc0 3200) are more stronger than 256 x24 (Lc0 1541 etc…) nets.

[MMarco]

My integrated card Radeon rx vega 10 runs at 10w and the bests 10x128 nets (i.e. 591226) are on par with single threaded Stockfish 11 on my Ryzen 7 3750h at blitz time control. Roughly, Lc0 reaches about 5000 nps (without tree reuse) while Stockfish about 1 500 000.

Code: Select all

Match: Various 10x128 vs. Stockfish 11 1CPU
LC0-version: lc0 v0.25.1 Dx 12
LC0 options: weights=128x10-t60-2-1770.pb.gz, backend=dx12, backend-opts=gpu=1, threads=2, logit-q=false, minibatch-size=256, max-prefetch=32, max-collision-events=32, max-out-of-order-evals-factor=2, syzygy-paths=C:\syzygy
SF options: threads=1, Hash=64
Hardware: Radeon RX Vega 10 + ryzen-7 3750h
Time control: 40/27.5sec repeating (one-fourth of CCRL blitz TC)
Book: Balsa_v500.pgn
Tablebases: 5 mens syzygy
Adjudication: Draw: after 50 moves, 5 moves < 5cp, Win:after 50 movess, 5 moves > 500cp
Software: Cute Chess 1.0.0
Comments: Results are very close.
games: http://s000.tinyupload.com/?file_id=03539289722480849649

Rank Name                        Elo    +    - games score oppo. draws
   1 SF 11 1CPU                   18    8    8  3000   54%    -6   58%
   2 Dx 12 Lc0 25.1 t60-2-1770    -4   13   13  1000   46%    18   56%
   3 Dx 12 Lc0 25.1 702556        -7   13   13  1000   46%    18   56%
   4 Dx 12 Lc0 25.1 591226        -7   13   13  1000   45%    18   60%

   1 SF 11 1CPU                   18 3000.0 (1629.5 : 1370.5)                     
   2 Dx 12 Lc0 25.1 t60-2-1770    -4 1000.0 (460.0 : 540.0)
   3 Dx 12 Lc0 25.1 702556        -7 1000.0 (456.0 : 544.0)
   4 Dx 12 Lc0 25.1 591226        -7 1000.0 (454.5 : 545.5)

[Modern Times]

My integrated card Radeon rx vega 10 runs at 10w and the bests 10x128 nets (i.e. 591226) are on par with single threaded Stockfish 11 on my Ryzen 7 3750h at blitz time control. Roughly, Lc0 reaches about 5000 nps (without tree reuse) while Stockfish about 1 500 000.

That seems pretty impressive to me !

[corres]

There is quite a lot information about running Leela at TDP>120+W GPU (RTX class)
I wonder what is good low-power (TDP<=65...75W) GPU for Leela
Is GTX-1650 best in this sub-75 class (as of 2020-Q2)?
What is SF-dev ratio for LC0@GTX-1650, how many NPS need SF-dev to be roughly equal?

The power of an NN net depend on the hardware on what the asked NN was built too.
Because its power at least on RTX level, so the NN power is on at least RTX level, that is you can use relative effectively with a weak GTX card too.

[MMarco]

My integrated card Radeon rx vega 10 runs at 10w and the bests 10x128 nets (i.e. 591226) are on par with single threaded Stockfish 11 on my Ryzen 7 3750h at blitz time control. Roughly, Lc0 reaches about 5000 nps (without tree reuse) while Stockfish about 1 500 000.

That seems pretty impressive to me !

Indeed! The onboard card is surprisingly strong, at least with 10x128 nets. I'll eventually try larger ones to see what happens then.