Slower Training Performance with MinkowskiEngine on PyTorch 2.4 + CUDA 12.x

[alibalapour]

Hi,

I am running the same training code in two environments and consistently see slower performance when using PyTorch 2.4.0 + CUDA 12.1 compared to PyTorch 1.9.0 + CUDA 11.x. Both environments use MinkowskiEngine 0.5.4, the same dataset, the same model, and identical configuration.

Environment 1 (faster)

• PyTorch 1.9.0 (CUDA 11.x)
• MinkowskiEngine 0.5.4
• KNN-CUDA 0.2

Environment 2 (slower)

• PyTorch 2.4.0 + cu121
• CUDA 12.1
• MinkowskiEngine 0.5.4
• KNN-CUDA 0.2

On the CUDA 12 / PyTorch 2.4 setup, sparse convolutions and coordinate map operations take noticeably longer, and overall training time per epoch increases by a factor of 2–4×. The hardware is identical across both tests.

I’m trying to determine whether MinkowskiEngine is expected to run slower on the CUDA 12 toolchain or with PyTorch 2.x, or if this is a compatibility issue. Any clarification on recommended PyTorch/CUDA versions for optimal performance would be helpful.

Thank you.