https://dev.to/ingero/124x-slower-what-pytorch-dataloader-actually-does-at-the-kernel-level-3o3a
124x Slower: What PyTorch DataLoader Actually Does at the Kernel Level - DEV Community
TL;DR: PyTorch's DataLoader can be 50-124x slower than direct tensor indexing for in-memory GPU... Tagged with pytorch, gpu, python, cuda.