Amin Vahdat on Google's AI Infrastructure, TPUs, and Co-Design
Google's Chief Technologist for AI Infrastructure Amin Vahdat examines the engineering challenges of managing clusters at 100,000-accelerator scale, where failures occur multiple times per hour. Rather than focusing on theoretical FLOPS, he advocates for "goodput," measuring the useful work delivered through real-world faults. Vahdat details the rationale behind splitting Google's TPU line into 8i and 8t models, while explaining why core TPU primitives have persisted since v1. He highlights Google's close co-design process with DeepMind, which allows teams to intercept chip designs mid-flight before tape-out. Additionally, the conversation explores how long-horizon agents drive unexpected demand for CPUs and storage alongside accelerators, the role of optical circuit switching in rerouting traffic, and the power constraints shaping gigawatt-scale data center planning.
Original sources
Podcast Training Data
Episode:Google's AI Infrastructure Chief, Amin Vahdat, on the Physics & Economics of Frontier AI ↗
YouTube Sequoia Capital
Video:Google's AI Infrastructure Chief, Amin Vahdat, on the Physics & Economics of Frontier AI ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.