Can Leela use 2x RTX gpus?

[M ANSARI]

Without really understanding what is going on with the GPU or how it works in multiple configurations. Is it possible that the way things are setup now is giving a false positive? Maybe the higher NPS numbers are not actually giving the proper scaling due to some inefficiency in another sector? Someone should try using SLI to see if that maybe then becomes "1" GPU rather than multiple GPU's and might work better? One thing is for sure is that in the last Sufi the 4 x GPU Lc0 that played had some moves that were inferior to what I was seeing with only 1 GPU 2080 Ti card. I had thought that this might have been a thermal throttling issue, but maybe it is something else and simply that higher nps with multiple GPU's is not really showing an accurate measure of scaling. Has there been a test match up between one 2080 Ti card vs say 2 x 2070 cards (which should have higher nps and thus should score better)?

[corres]

Without really understanding what is going on with the GPU or how it works in multiple configurations.

In general there are rather few information about Lc0 build and its working.
Developers do not like to write simple text and obviously their viewpoints of copy protection is also a factor.

...
One thing is for sure is that in the last Sufi the 4 x GPU Lc0 that played had some moves that were inferior to what I was seeing with only 1 GPU 2080 Ti card.

In this phenomenon there is no any surprising. You can find this in the case of AB engines too. The engines do not work deterministic mode. In their working and in their evaluation the chance plays an important role.

Has there been a test match up between one 2080 Ti card vs say 2 x 2070 cards (which should have higher nps and thus should score better)?

I also should like to see a correct match between a 2080 Ti card and the well working dual cards.

[Werewolf]

No, I am saying it doesn't scale up with more cores at all. In fact, you will lose performance with more cores, not gain.

But for the example I gave - dual 2080 Ti - 3 cores is clearly faster than one. I'm not sure I follow your point.

[Hugo]

Without really understanding what is going on with the GPU or how it works in multiple configurations. Is it possible that the way things are setup now is giving a false positive? Maybe the higher NPS numbers are not actually giving the proper scaling due to some inefficiency in another sector? Someone should try using SLI to see if that maybe then becomes "1" GPU rather than multiple GPU's and might work better? One thing is for sure is that in the last Sufi the 4 x GPU Lc0 that played had some moves that were inferior to what I was seeing with only 1 GPU 2080 Ti card. I had thought that this might have been a thermal throttling issue, but maybe it is something else and simply that higher nps with multiple GPU's is not really showing an accurate measure of scaling. Has there been a test match up between one 2080 Ti card vs say 2 x 2070 cards (which should have higher nps and thus should score better)?

Hello

I had +29Elo with doubling Lc0 speed by adding a second RTX card.
On network 42000 I have relieable Elo estimation. So I could compare the results.
I used backend multiplexing.
C.K.

[Albert Silver]

No, I am saying it doesn't scale up with more cores at all. In fact, you will lose performance with more cores, not gain.

But for the example I gave - dual 2080 Ti - 3 cores is clearly faster than one. I'm not sure I follow your point.

Nor do I understand yours. I told the OP that slow clocks were a problem, and you asked if this could not be overcome with more cores, and I said no. 3 cores is ideal for two 2080tis, and 5, 8, 12 or any number more is worse.

[corres]

I had +29Elo with doubling Lc0 speed by adding a second RTX card.
On network 42000 I have relieable Elo estimation. So I could compare the results.
I used backend multiplexing.
C.K.

Congrats!