Mentatcurated
▸ Concept

AI chips

Processors purpose-built for the matrix math that training and running neural networks requires — trading general-purpose flexibility for raw throughput on that one workload.

In a nutshell

A CPU executes arbitrary instructions; an AI chip does one thing — multiply large arrays of numbers together — and does it at a scale a CPU cannot match. GPUs were the first press-ganged for the job; custom ASICs (Google's TPU, Cerebras, Groq) go further, stripping out everything except that arithmetic and the memory bandwidth to feed it. The hard part is not the chip itself but the software stack: a processor is useless without compilers and frameworks that know how to tile and schedule tensor operations across thousands of cores. That full-stack dependency is why switching chips is expensive and why vertical integration is a structural bet.

Where it came from

Year2016
SourceGoogle — "In-Datacenter Performance Analysis of a Tensor Processing Unit" (ISCA 2017)
Why it matteredFirst public disclosure of a custom AI ASIC deployed at scale, establishing the template for purpose-built inference silicon.

How this connects

Tap a node to open it