Reiner Pope – Chip design from the bottom up
Summary
Dwarkesh Patel interviews Rainer Pope, CEO of AI chip startup Maddox, in a technical deep dive into chip architecture. They explain how logic gates perform multiply-accumulate operations, the quadratic area costs of precision, and how systolic arrays optimize matrix multiplication by minimizing data movement. The discussion highlights the critical trade-off between compute density and communication bandwidth in modern AI hardware design.
Topics discussed
Introduction and fundamentals of Multiply-Accumulate (MAC) units
Logic gates, full adders, and the cost of data movement
Systolic arrays and matrix multiplication in AI chips
Sponsor segment: Cursor AI coding assistant
Clock synchronization, registers, and chip design constraints
Jane Street interview on AI and high-frequency trading
FPGA architecture, lookup tables, and deterministic latency
CPU vs GPU architecture: caches, branch prediction, and parallelism
Energy efficiency, clock speeds, and GPU vs TPU design trade-offs
Listen ad-free on Castria