Leela Hardware Question

[Stephen Ham]

Hello All,

I recently upgraded my GPU to a GeForce RTX 5080, but my CPUs are AMD Ryzen 9 3900X 12-cores. Of course, only two cores are dedicated to Leela. I use Leela 32.0 with the BT4-1740 net on Cuda 12.

I'm an ICCF GM but a total dummy regarding hardware tech. An avid gamer buddy thinks that my CPUs might cause an underperformance. I ran my Leela benchmark: 17,340 NPS and submitted that to the Leela website. Unfortunately, no feedback was given regarding whether that score is good/poor/average.

My thinking is that since only two cores are dedicated to Leela, my CPUs should thus only be a minor factor regarding Leela's performance. Leela runs with 3+ cores, but the website still advocates using only two to avoid a data bottleneck. But if my slower CPUs are a drag on performance, then shouldn't assigning three CPUs be a fix?

My question is whether evidence confirms that my six year old CPUs cause reduced performance when mated with my new GPU.

Also, where can I find comparison Leela benchmarks? I found this website: https://openbenchmarking.org/test/pts/l ... 86c408e8e0
but what's displayed there looks odd. Some of it claims to be current data, but elsewhere it shows the data is from June 18, 2021, and referenced Leela 0.26.

Thanks in advance.
-Steve-

[Werewolf]

Hello All,

I recently upgraded my GPU to a GeForce RTX 5080, but my CPUs are AMD Ryzen 9 3900X 12-cores. Of course, only two cores are dedicated to Leela. I use Leela 32.0 with the BT4-1740 net on Cuda 12.

I'm an ICCF GM but a total dummy regarding hardware tech. An avid gamer buddy thinks that my CPUs might cause an underperformance. I ran my Leela benchmark: 17,340 NPS and submitted that to the Leela website. Unfortunately, no feedback was given regarding whether that score is good/poor/average.

My thinking is that since only two cores are dedicated to Leela, my CPUs should thus only be a minor factor regarding Leela's performance. Leela runs with 3+ cores, but the website still advocates using only two to avoid a data bottleneck. But if my slower CPUs are a drag on performance, then shouldn't assigning three CPUs be a fix?

My question is whether evidence confirms that my six year old CPUs cause reduced performance when mated with my new GPU.

Also, where can I find comparison Leela benchmarks? I found this website: https://openbenchmarking.org/test/pts/l ... 86c408e8e0
but what's displayed there looks odd. Some of it claims to be current data, but elsewhere it shows the data is from June 18, 2021, and referenced Leela 0.26.

Thanks in advance.
-Steve-

You're using Zen 2 cores, as am I. In one week I am changing to Zen 5 so I'll let you know the uplift with solid data. I suspect your friend has a grain of truth in his argument, but I'll bet changing the CPU won't be a massive upgrade. I very much doubt going to 3 cores will help and might be a downgrade.

BT4 is a great net but it's big, and your nps are therefore good.

[Modern Times]

You're using Zen 2 cores, as am I. In one week I am changing to Zen 5 so I'll let you know the uplift with solid data. I suspect your friend has a grain of truth in his argument, but I'll bet changing the CPU won't be a massive upgrade.

That will be interesting to see - assuming the GPU stays the same. I'd be surprised if there is a big difference but you never know. Some of the experts on the Discord Lc0 group will probably have a good idea on it.

[lucario6607]

Probably gpu bottlenecked more than anything. When you have 8x5090 like tcec you then become cpu bottlenecked and a faster cpu will help. Not saying you won’t see an uplift but probably not much.

[Ferdy]

I compiled lc0 benchmarks of a few instances from vast.ai.

Your 17,340 is actually higher.

https://lc0-benchmark.vercel.app/

[Stephen Ham]

Thanks to all of you.

Based upon your posts, and especially your benchmark compilation, Ferdy, I now believe that I have no CPU bottleneck and that everything is fine with my GPU upgrade. A buddy who owns multiple GPUs also ran some benchmarks, displaying results similar to yours, Ferdy.

My conclusion is that the two assigned CPUs play very little role in Leela's performance. It's instead 99% GPU performance that matters.

However, I'm curious how much performance improvement Cuda 12 offers over Cuda 11.

All the best,
-Steve-

[Ferdy]

However, I'm curious how much performance improvement Cuda 12 offers over Cuda 11.

At least for rtx 3090, there is not much. Test results below are from vast.ai.

Code: Select all

AMD EPYC 7402P 24-Core Processor
./lc0 benchmark

Code: Select all

Loading weights file from: ./weights_BT4-spsa-1740.pb.gz
Weights file has multihead format, updating format flag
Switching to [cuda-fp16]...
CUDA Runtime version: 11.8.0
Latest version of CUDA supported by the driver: 12.8.0
GPU: NVIDIA GeForce RTX 3090
GPU memory: 23.5695 Gb
GPU clock frequency: 1695 MHz
GPU compute capability: 8.6
L2 cache capacity: 6291456
Nodes/second    : 8749

Code: Select all

Loading weights file from: ./weights_BT4-spsa-1740.pb.gz
Weights file has multihead format, updating format flag
Switching to [cuda-fp16]...
CUDA Runtime version: 12.8.0
Latest version of CUDA supported by the driver: 12.8.0
GPU: NVIDIA GeForce RTX 3090
GPU memory: 23.5695 Gb
GPU clock frequency: 1695 MHz
GPU compute capability: 8.6
L2 cache capacity: 6291456
Nodes/second    : 8752