GPU-lco testing

[cma6]

Are there any sites that compile results for various GPUs using lc0--especially the new GPUs from Nvidia & AMD?

I'm looking at a new system build. My last one has Nvidia 4070 Super GPU Super. From start position, after 47 minutes running lc0 v31.2 with BT4-spsa-1740 net, I get 8.9 kN/s.

Thanks in advance to the experts?

[smatovic]

Phoronix has some benchmarks:

https://www.phoronix.com/
https://cse.google.com/cse?cx=partner-p ... earch&ref=

Lc0 website benchmarks look outdated:
https://lczero.org/dev/wiki/benchmarks/

There were some Excel sheets with NPS numbers in the web, but it is impossible to compare different net-architectures on different hardware with different backends with different backend-parameters.

--
Srdja

[Hai]

Are there any sites that compile results for various GPUs using lc0--especially the new GPUs from Nvidia & AMD?

I'm looking at a new system build. My last one has Nvidia 4070 Super GPU Super. From start position, after 47 minutes running lc0 v31.2 with BT4-spsa-1740 net, I get 8.9 kN/s.

Thanks in advance to the experts?

Buy the Apple MacBook Pro 16-inch M4 MAX maxed out.

[cma6]

Are there any sites that compile results for various GPUs using lc0--especially the new GPUs from Nvidia & AMD?

I'm looking at a new system build. My last one has Nvidia 4070 Super GPU Super. From start position, after 47 minutes running lc0 v31.2 with BT4-spsa-1740 net, I get 8.9 kN/s.

Thanks in advance to the experts?

3/26/25 Thanks, Srdja, for very helpful references. I joined Phoronix though could not find their benchmarks.
The lc0 benchmarks are way out of date but perhaps they will update in a couple of months.

Greatly appreciated,
Harmonia

[smatovic]

Michael Larabel from Phoronix runs also OpenBenchmark:

https://openbenchmarking.org/test/pts/l ... 82#metrics

but, GPU chart only up to RTX 3000 series and only OpenCL backend (what is outdated in regard in of network architecture AFAIK).

You can compare RTX 4000 and 5000 series simply by TFLOPS in FP16 vector and matrix compute (Tensor), but I do not know, how this translates in NPS gain for Lc0.

https://en.wikipedia.org/wiki/List_of_N ... _40_series
https://en.wikipedia.org/wiki/List_of_N ... _50_series

--
Srdja

[Werewolf]

Michael Larabel from Phoronix runs also OpenBenchmark:

https://openbenchmarking.org/test/pts/l ... 82#metrics

but, GPU chart only up to RTX 3000 series and only OpenCL backend (what is outdated in regard in of network architecture AFAIK).

You can compare RTX 4000 and 5000 series simply by TFLOPS in FP16 vector and matrix compute (Tensor), but I do not know, how this translates in NPS gain for Lc0.

https://en.wikipedia.org/wiki/List_of_N ... _40_series
https://en.wikipedia.org/wiki/List_of_N ... _50_series

--
Srdja

I heard 5090 was around +25% NPS over 4090

[smatovic]

I heard 5090 was around +25% NPS over 4090

Then we have a 1:1 relation with vector + matrix TFLOPS FP16, +25%.

--
Srdja