Lc0 settings for Leela on GPU- Nvidia RTX 2080 Super

[Nordlandia]

Any recommended tweaks for slightly enhanced performance?

CPUCT or similar tweaks.

https://github.com/LeelaChessZero/lc0/wiki/Lc0-options

Based of this thread:
viewtopic.php?f=2&t=70035

[corres]

Any recommended tweaks for slightly enhanced performance?
...

As the power of "Super" is higher only about 10% those parameters what is good for normal 2080 Ti also good for Super.
There are no "ideal" parameters because they depends on the used NET and TC, but in the most cases are good for lc0 ver.0.22.0
Threads = 2
Backend = cudnn-fp16
NNCacheSize = 10000000
any other are Default.

For LTC may be better
Threads = 2
Backend = cudnn-fp16
NNCachesize = 20000000
MiniBatchSize = 512
CPuct = 3.4
CPuctBase = 10000

[Nordlandia]

I usually test casual engine games at 30m+30s time control. And recently two computer match via Fritz engine cloud with ponder enabled.

So 30 minute + 30 second increment in addition to ponder enabled. i7-5960X 4.5GHz and i7-5930K 4.5GHz. So the main hardware is fairly even.

Stockfish on 5960X and Leela on 5930K.

[corres]

I usually test casual engine games at 30m+30s time control. And recently two computer match via Fritz engine cloud with ponder enabled.
So 30 minute + 30 second increment in addition to ponder enabled. i7-5960X 4.5GHz and i7-5930K 4.5GHz. So the main hardware is fairly even.
Stockfish on 5960X and Leela on 5930K.

It depends on the used NET and other factors (Stockfish params, opening book, etc.)

[Ponti]

I usually test casual engine games at 30m+30s time control. And recently two computer match via Fritz engine cloud with ponder enabled.
So 30 minute + 30 second increment in addition to ponder enabled. i7-5960X 4.5GHz and i7-5930K 4.5GHz. So the main hardware is fairly even.
Stockfish on 5960X and Leela on 5930K.

It depends on the used NET and other factors (Stockfish params, opening book, etc.)

Which one is the current best version of the Leela neural network ?
The binary my computer uses (Ryzen 1800X) is lc0-v0.19.1.1-windows-opencl. File NN is 32141.
I don't own a great video-card (using a Radeon Sapphire HD 7790), though I'm planning to buy a new one.

Thanks in advance,
A. Ponti

[corres]

I usually test casual engine games at 30m+30s time control. And recently two computer match via Fritz engine cloud with ponder enabled.
So 30 minute + 30 second increment in addition to ponder enabled. i7-5960X 4.5GHz and i7-5930K 4.5GHz. So the main hardware is fairly even.
Stockfish on 5960X and Leela on 5930K.

It depends on the used NET and other factors (Stockfish params, opening book, etc.)

Which one is the current best version of the Leela neural network ?
The binary my computer uses (Ryzen 1800X) is lc0-v0.19.1.1-windows-opencl. File NN is 32141.
I don't own a great video-card (using a Radeon Sapphire HD 7790), though I'm planning to buy a new one.
Thanks in advance,
A. Ponti

It is your task to find it or ask peoples who use opencl.

[shrapnel]

I don't own a great video-card (using a Radeon Sapphire HD 7790), though I'm planning to buy a new one.

Well, don't make the mistake of buying an AMD Radeon Card again, especially if you want to use AI Chess Engines, like lc0 and Fat Fritz.

[Nordlandia]

I'm using T40B.pb from http://data.lczero.org/files/networks-contrib/

[corres]

I'm using T40B.pb from http://data.lczero.org/files/networks-contrib/

T40B.4-160 or -260 are strong nets for a single Ti.
Maybe 42872 is stronger in some situations.
For 2-4 x RTX 2080 Ti J13B.2-188 is even better.

[Nordlandia]

I'm using T40B.pb from http://data.lczero.org/files/networks-contrib/

T40B.4-160 or -260 are strong nets for a single Ti.
Maybe 42872 is stronger in some situations.
For 2-4 x RTX 2080 Ti J13B.2-188 is even better.

Which of 160 or 260 do you recommend

I usually test setup positions. Often imbalances ones.