#622: How NVIDIA Connects 72 GPUs in One Rack
NVIDIA CTO Michael Kagan joins David Bombal to explain AI networking, NVLink, and why GPUs alone are not enough to scale AI data centers. How do you connect 72 GPUs in a rack? What is the difference between scale up, scale out, and scale across? And why do latency and network jitter matter when GPUs need to work together? Michael explains how NVIDIA approaches these challenges using NVLink, InfiniBand, and Spectrum-X Ethernet. He also shares the story of co-founding Mellanox, joining NVIDIA, and building the networking behind increasingly large AI systems. We discuss the path toward million-GPU systems, connecting AI factories...