faiss.knn_gpu vs torch.topk

[algoriddle]

Summary

torch.topk outperforms faiss.knn_gpu - both classic and raft

Platform

OS: Linux, Tesla V100-SXM2-16GB

Faiss version:

faiss-gpu-raft 1.8.0 py3.12_h4c7d538_114_cuda12.1.1_nightly pytorch/label/nightly
libfaiss 1.8.0 hb0f4bcb_114_cuda12.1.1_raft_nightly pytorch/label/nightly

Installed from: conda

Faiss compilation options: OPTIMIZE AVX2 GPU NVIDIA_RAFT

Running on:

• CPU
• GPU

Interface:

• C++
• Python

Reproduction instructions

See notebook: https://gist.github.com/algoriddle/bae7ebaf4cee6a63b218ce24f0558cf0

@cjnolet, can you tag the appropriate people on your side, please?

[cjnolet]

cc @achirkin @tfeher @divyegala @tarang-jain

[stepelu]

See #3045

[mfoerste4]

The recent changes in both cuvs and the faiss/cuvs-PR #3549 reduce launch overhead and improve performance, especially for smaller dimensions (also see cuvs issue). As an example, here are results for a local build on a L40: