Skip to content

Hardware requirements

Jump to bottom
Tomasz Sobczyk edited this page Jun 30, 2022 · 2 revisions

Requirements

Minimal hardware requirements for running a single instance of the nnue-pytorch trainer:

  • NVIDIA GPU (CUDA requirement)
  • 4GB VRAM
  • 2GB RAM

Reasonable hardware for running a single instance of the nnue-pytorch trainer:

  • Nvidia GTX 1080ti
  • 6GB VRAM
  • 4GB RAM
  • 8 CPU threads
  • 20GB of disk space

Recommended hardware for running a single instance of the nnue-pytorch trainer:

  • Nvidia RTX 2080
  • 8GB VRAM
  • 8GB RAM
  • 12 CPU threads
  • 50GB of disk space

Reasonable hardware for running network testing in real time alongside training:

  • 400MB RAM / game concurrency
  • 16 CPU threads

Recommended hardware for running network testing in real time alongside training:

  • 500MB RAM / game concurrency
  • 32 CPU threads

Additional notes

It is possible to run multiple trainer instance in parallel. This is often done to fully saturate high-end GPUs and reduce impact of variance between runs. High-end GPUs usually require 2 runs in parallel on each GPU to saturate it.

Testing the networks as they are produced is a heavy task, and usually will impact training speed even when not utilizing virtual cores due to memory bandwidth.

Benchmarks

TODO: A trainer benchmark script.

Clone this wiki locally