Optimizing PyTorch on CPU-GPU Coherent Platforms - Matthias Jouanneaux, Nvidia

This talk covers the optimization of PyTorch on Nvidia's GB200 coherent platform, highlighting the benefits and challenges presented by recent advancements in CPU-GPU architecture. The speaker, Matthias Jouanneaux from Nvidia, discusses techniques for leveraging features such as high CPU-GPU interconnect bandwidth and unified memory. Additionally, the session explores the implications of sharing system memory between CPU and GPU to enhance application performance in PyTorch.