NN Engine GPU Usage

[Nordlandia]

Is ponder an option for twin GPUs and NN
For example Leela selfplay at one GPU each.
My PSU is strong enough.

If you play one GPU against the other GPU the ponder may be O.K.
I am not sure if ponder works on Leela.
If it works you should configure Leela-s well.
I do not know why you stick to ponder. It causes plus heat and decreases the stability of the system.
Ponder is really good for PC against PC match but for engine-engine match is problematic.

Graham Banks said years ago that ponder works fairly well on dual CPU motherboards like Xeon. So i reasoned that twin GPUs may give the same effect for NN engines.

[corres]

Is ponder an option for twin GPUs and NN
For example Leela selfplay at one GPU each.
My PSU is strong enough.

If you play one GPU against the other GPU the ponder may be O.K.
I am not sure if ponder works on Leela.
If it works you should configure Leela-s well.
I do not know why you stick to ponder. It causes plus heat and decreases the stability of the system.
Ponder is really good for PC against PC match but for engine-engine match is problematic.

Graham Banks said years ago that ponder works fairly well on dual CPU motherboards like Xeon. So i reasoned that twin GPUs may give the same effect for NN engines.

You thought well, the situation is the same.
But a good server is different from a home-PC.

[corres]

roundrobin(if you have 2 identical GPUs) for using the 2 GPUs.

Hmm... I didn't know that. I tried it and it seems to be giving faster speeds than multiplexing, so you are right.
Any other tips for multi-gpu users ?

If you consecutive run speed tests some times the result may differ from each others with 5 - 10 %.
Especially if you do not let enough time to cool down the GPU(s). Moreover the speeds what you can get with different but good parameters are in the same range.
So to decide which setup is better is not a simple work.
But it is sure Multiplexing works well in every cases even if you use different GPUs.
Multiplexing can also equalize the different heating of GPUs. If you use your PC for hours this is a
non- negligible issue.

[Stephen Ham]

OK, switching Leela to Multiplexing, since I have two GPUs, and 4 CPUs, but leaving BackendOptions blank, since it crashes when the recommended data is entered (see my other post) I now have Task Manager open exactly as you advised.

Leaving backendoptions blank can never lead to a good end.

And what Corres said, is the first time i see this type of command for Leela.
This is what you want to put in the BackendOptions in Leela parameters the Fritz GUI:
(backend=cudnn-fp16,gpu=0),(backend=cudnn-fp16,gpu=1)

And choose multiplex(if you do NOT have 2 identical GPUs) or roundrobin(if you have 2 identical GPUs) for using the 2 GPUs.

If that doesn't work then.......nah it will work.

Thanks again, George!

While switching from multiplexing to roundrobin does indeed show increased kN/s, the BackendOptions you cited, which match what Corres cited, still cause the engine to crash. So, until this is resolved, I need to leave BackendOptions blank.

Any solution?

All the best,
-Steve-

[corres]

While switching from multiplexing to roundrobin does indeed show increased kN/s, the BackendOptions you cited, which match what Corres cited, still cause the engine to crash. So, until this is resolved, I need to leave BackendOptions blank.

Because Fat Fritz uses the same engine as Leela the given parameters are good.
The problem is you did not optimize well your Fat Fritz.
Find Setup_Fat_Fritz.exe in c:\Program Files\Common Files\ChessBase\InstallUserDocs\GUI17\ library.
Click on the .exe and step by step go thru the process. Mark with x cuda RTX as type of graphic card, if it asks.
If the installation is completed start Fritz17 and choose "Enter & Analyse" then "Engine" etc. and you can use every parameters what is appropriate for Leela on the Load engine - Advanced - etc. panel.
Maybe you should ask from ChessBase a Manual of Fritz17.