Most AI companies shout about GPUs. But here’s the truth: Google TPUs are the real cheat code. It’s not just about raw FLOPs or benchmark chasing, it’s about pipelines. When you design your machine learning stack TPU-first, GPU-second, you unlock massive gains in price and performance. This talk starts at the roots, by building an understanding of machine mentality. We’ll explore foundational concepts like Instructions Per Cycle (IPC) and vector processing, not just as theoretical metrics but as tools to reshape how you prepare, encode, and move data through your pipeline.


Even if you're new to data encoding, Arrow, Parquet, JAX, or TPUs, you’ll walk away with clear mental models and practical insights on how these pieces connect to build performant ML systems.

Full Title - Google TPU-First, GPU-Second: Understanding Data Encoding—Arrow and Parquet’s Impact on Training Pipelines