Skip to main content

Hardware Support

Currently neutrino supports:

Hardware PlatformSupport Status
NVIDIA/CUDA/PTX✅ Supported
AMD/ROCm/GCNAsm🏗️ Supported on CDNA
Intel/OneAPI/VISA🚀 Planning
Apple/Metal/AIR🚀 Planning

Multi-Card Support​

neutrino supports analysis of kernel(s) launched on multiple cards or instances via multi-processing like MPI and solutions on top like PyTorch distribtued.

To use, please place neutrino before launch command like torchrun like:

neutrino -p <prbes> torchrun ...

Additional Notes:

  1. After the run, traces of different process (under MPI) will be placed in separated folder.
  2. Neutrino supports most selecting card options like CUDA_VISIBLE_DEDVICES or torch.device API.
  3. Neutrino's support on multi-threading is still under testing.

Notes for Platforms​

NVIDIA/CUDA/PTX​

  • NVIDIA propietary libraries, cuBLAS/cuDNN, are not supported.

AMD/ROCm/GCNAsm​

  • Support for PyTorch experiences a problem in registerFatbinary, will fix soon.
  • Currently, only CDNA series, MI100x/MI200x/MI300x GPUs, are supported. RDNA GPUs like RX7000x/RX9000x GPUs, are not supported.