Does Pytorch benefit from Nvidia GPU direct storage?

Following the guide for installing the Cuda library, step 2.5. Install MLNX_OFED says that if you intend to use GPUDirectStorage (GDS), you must install the CUDA package and MLNX_OFED package.

Searching for GPUDirectStorage (GDS), I found that it accelerates data access for GPU. However, I wonder if Pytorch will benefit from installing this.

1 Like

Yes, you could use GDS e.g. via DALI in PyTorch (and other frameworks).

1 Like