Graphics card

[AaronB]

What are the best graphic cards for baduk computing? Give a few options at different price ranges.

Thanks for any input.

[iopq]

Data point: my RTX 570 only gets 170 nps, I got it for $130 plus tax, but I will try to sell the games for about $50

[Mike Novack]

You should define that question better. Power consumption and heat play a role in choice. So what's best for a desktop/workstation not the same as for a laptop/workstation << I have "workstation" there because whether desktop or laptop, you are presumably talking about a device more powerful that the typical home computer.

[Aram]

If you are a Leela Zero user, then a RTX series card from Nvidia is probably a good idea as Leela supports the tensor cores.
Leela Chess Zero also supports tensor cores (in case you are interested in chess as well).

So in case your power supply can handle it:

NVIDIA Geforce RTX 2060 (320e to 380e price range)
NVIDIA Geforce RTX 2070 (500e to 600e price range)
NVIDIA Geforce RTX 2080 (650e to 850 price range)
NVIDIA Geforce RTX 2080 Ti (1200e plus price range)

[iopq]

Nvidia 1060 is 200 n/s, you could have gotten it for $170, maybe after AMD cards come out you can

[AaronB]

Thanks for the input guys. I appreciate it.

[spook]

I collected some stats from various posts on LeeLa Zero's github page.

• GTX 760 2GB : 7489 ms/move
  GTX 760: 2912 ms/move
  K80 on google cloud: 2100ms/move
  GTX 780: 1720 ms/move
  GTX 1050 ti: 1333 ms/move
  GTX 1060 3GB: 1230 ms/move
  Titan V: 923 ms/move
  RX 480: 876ms/move
  GTX 1070 8GB: 845ms/move
  GTX 1080Ti : 792 ms/move
  GTX 1080 8GB: 733 ms/move
  GTX 1080: 701 ms / move (9% overclocked)
  GTX 1080 TI: 622 ms/move
  1060 6GB next: 570 ms/move
  2xGTX 1080 : 532 ms/move or 438ms/move
  1060 6GB winograd: 350 ms/move
  2xGTX 1080 : 260 ms/move
  GTX 1070 winograd: 255 ms/move
  2xGTX 1080 winograd: 202 ms/move ---> BEST

However, there is a lot of noise in this list.
And we have to be careful when taking conclusions.
(Aside from all the above, give your system enough RAM.)

Needless to say, I do think that ZBaduk can offer you better value for money.
If you haven't already, create a free account !

[explo]

I collected some stats from various posts on LeeLa Zero's github page.

• GTX 760 2GB : 7489 ms/move
  GTX 760: 2912 ms/move
  K80 on google cloud: 2100ms/move
  GTX 780: 1720 ms/move
  GTX 1050 ti: 1333 ms/move
  GTX 1060 3GB: 1230 ms/move
  Titan V: 923 ms/move
  RX 480: 876ms/move
  GTX 1070 8GB: 845ms/move
  GTX 1080Ti : 792 ms/move
  GTX 1080 8GB: 733 ms/move
  GTX 1080: 701 ms / move (9% overclocked)
  GTX 1080 TI: 622 ms/move
  1060 6GB next: 570 ms/move
  2xGTX 1080 : 532 ms/move or 438ms/move
  1060 6GB winograd: 350 ms/move
  2xGTX 1080 : 260 ms/move
  GTX 1070 winograd: 255 ms/move
  2xGTX 1080 winograd: 202 ms/move ---> BEST

However, there is a lot of noise in this list.
And we have to be careful when taking conclusions.
(Aside from all the above, give your system enough RAM.)

Needless to say, I do think that ZBaduk can offer you better value for money.
If you haven't already, create a free account !

What are those ms/move values?

[Ferran]

What are those ms/move values?

At a guess, miliseconds per move; the lower, the better.

Take care.

[explo]

What are those ms/move values?

At a guess, miliseconds per move; the lower, the better.

Take care.

What are they a measure of?

[Ferran]

What are they a measure of?

Speed. It's telling you how long it takes to evaluate a single move.You're going to have your computer calculate those by the hundreds of thousands. IIRC, Leela Z can reach 50000 move evaluations for every actual real move. 200 moves in a game is not that uncommon. So, 10 million move evaluations per game (not quite: endgame has less positions, you're going to be doing half of those yourself... Etc)

Take care

[explo]

What are they a measure of?

Speed. It's telling you how long it takes to evaluate a single move.You're going to have your computer calculate those by the hundreds of thousands. IIRC, Leela Z can reach 50000 move evaluations for every actual real move. 200 moves in a game is not that uncommon. So, 10 million move evaluations per game (not quite: endgame has less positions, you're going to be doing half of those yourself... Etc)

Take care

You're completly misunderstanding my question. I know they are measure of speed. But the speed of what? Is it the average speed when running autogtp? Is it the average speed when playing LZ on lizzie asking for a billion visits? Those would bring very different speed on the same hardware.
I would assume those are autogtp values, but in this case I probably have 3 poorly set versions of autogtp at home.

[Ferran]

You're completly misunderstanding my question. I know they are measure of speed.

Ah, sorry.

Take care.

[Aram]

Yepp, you should just ignore the table posted above with the ms/move, because it does not give any actual information.

The speed of which Leela generates a move depends on several things:

- You can run it on time settings, in which case it will generate moves according to time control, regardless of hardware strength
- You can run it at a set amount of visits, in which case it will generate a move once it has reached the amount of visits
- You can run it at play outs, in which case it will generate a move once it has reached the amount of play outs

The speed at which a graphics card is able to do one visit or play out depends on what network size you are running at (larger networks are slower),
and if you are running at FP16 or FP32. It also depends on if your graphics card has extra hardware for handling neural networks, such as tensor cores in the RTX series by NVIDIA.

Additionally it is good to note that most projects are heavily optimized for Nvidia cards, as those are usually the ones used in network training due to the very good Nvidia Tensorflow ecosystem and CUDA. So if you want a safe card which is more likely to work instantly with new projects (such as KataGo), get a Nvidia card.

You should note that Leela uses the tensor cores you can find on the Nvidia RTX cards, but that also, the GTX 1660 and GTX 1660 Ti cards support FP16 precision at a 2:1 speedup.
(While for example the older GTX 10xx cards only do FP16 at 1:2 speed, meaning half speed). This means that while the GTX 1660 and GTX 1660 Ti cards do not offer more speed with tensor cores, they do offer the option of running at a faster FP16 precision for a speedup.

[spook]

There is a more recent benchmark here:

http://www.flygo.net/bbs/forum.php?mod= ... tid=110194

The force is strong with RTX cards it seems.