Definition
Plain language
Google's custom chip built specifically for running AI workloads, an alternative to NVIDIA's GPUs.
As stated in the literature
Tensor Processing Unit — Google's application-specific accelerator for neural-network training and inference; its arithmetic implementation produces a distinct floating-point rounding signature from GPUs and CPUs.
Also called: TPUs
Why it matters: Its distinct arithmetic gives it a recognizable rounding signature, which matters both for performance choices and for hardware-fingerprinting attacks.
For example, a model trained on Google's hardware runs on chips purpose-built for neural networks rather than the GPUs many others use.
Heard on the show
“An NVIDIA GPU, a Google TPU, an Intel server CPU, Apple silicon — they all run their math through different libraries that schedule these operations differently.”Episode 158 — How Floating-Point Rounding Lets a Model Tell Which Chip It's On — And Misbehave