TPU
A TPU is Google's custom chip built for a single job: the matrix multiplications at the heart of neural networks. You do not buy it; you rent it by the hour inside Google Cloud, where it often beats general-purpose GPUs on training cost.
No jargon, no math
In plain words, what is it?
Like a GPU, a TPU crunches matrices in parallel, but the hardware is a fixed pipeline shaped around one operation: multiply numbers and add them up, at scale. You rent TPU chips inside Google Cloud, wired into pods of thousands with a custom interconnect, so many chips behave like one enormous brain. Before your model runs, Google's software compiles it into a static map of that pipeline. What maps cleanly runs extremely fast per watt and per euro; exotic operations that do not fit the pipeline run slower or wait for software support. Where it breaks: flexibility. New or unusual model architectures may compile poorly or not at all, and you are tied to Google's cloud, Google's pricing and Google's release schedule.
Terms that live nearby.
GPU
A GPU is a processor built to run thousands of small calculations at the same time. Designed for graphics, it turned out to be the perfect engine for AI, because running a neural network is mostly one enormous pile of matrix multiplications.
ReadAI Accelerator
An AI accelerator is a processor built for one job: the matrix multiplications that neural networks are made of. GPUs, TPUs and NPUs are all accelerators; they trade the flexibility of a CPU for far higher throughput per watt on AI workloads.
ReadTraining
Training is the process that turns an untrained model into a working AI: you feed it thousands or millions of examples, compare its guesses with the correct answers, and nudge its internal settings until the guesses get good.
Read§ 03
Want this running in your stack?
We design and ship AI systems with permissions, logs and human approval. It starts with a 30-minute scoping call.