PyTorch Conference Europe 2026

Fp8 Training From Hopper To Blackwell - Luca Wehrstedt, Meta

25:24 · 07 Apr 2026 – 08 Apr 2026 · YouTube

About this talk

This talk by Luca Wehrstedt from Meta explores the evolution of training with low-precision float8 data types using NVIDIA's Hopper to Blackwell generations of GPUs. The speaker discusses how TensorCore acceleration revolutionizes training efficiency while navigating crucial decisions related to accuracy versus efficiency and precision versus range. The session highlights advancements such as the DeepSeek release and micro-scaling formats in Blackwell, providing a comprehensive comparison of these approaches to assist researchers in optimizing their training strategies.