TPU (hardware)
Tensor Processing Unit — Google's custom-built chip (ASIC) for machine learning, optimized for the massive matrix multiplications neural networks require. The driving force behind Google's AI and an alternative to Nvidia's GPUs in the cloud.
Idea: instead of a general GPU, Google builds a chip whose heart is a "systolic array" — a grid of multipliers that streams data and does matrix multiplication extremely efficiently per watt. Used for both training and inference of large models; available only via Google Cloud (not sold standalone). Generations from TPU v1 (2015, inference) to today's training pods linked by the thousands. Trade-off: enormously efficient for what it's built for, but less flexible than a GPU and locked into Google's ecosystem/software stack (JAX/TF). Part of a broader trend of AI-specific ASICs (compare AWS Trainium, Microsoft Maia). Related to NPU and tensor cores.