Google described its first data-center TPU, an inference accelerator built around a systolic array for 8-bit matrix multiplication. The study compared it with CPUs and GPUs on deployed neural networks.
Google Tensor Processing Units (TPUs)
Also known as: Google TPU · Cloud TPU · Tensor Processing Unit
Plain English
A Tensor Processing Unit is a Google-designed chip that speeds up the matrix operations used to train and run neural networks. Google connects multiple TPUs so they can work on models that exceed the capacity of a single chip.
Technical Overview
TPUs use systolic matrix multiplication units that move intermediate values directly between neighboring processing elements. Later generations add high-bandwidth memory and inter-chip links for distributed training and inference. Frameworks such as JAX use XLA compilation to map programs and their sharding layouts onto the hardware.
Timeline (5)
2017
Paper2023
PaperThe TPU v4 paper described optical circuit switches that reconfigure links within a 4,096-chip system. The fabric can accommodate different topologies and route around unavailable components.
2025
ReleaseGoogle Cloud made TPU7x, the first Ironwood variant, generally available. Its higher-memory configuration targets both model training and inference.
2026
MilestoneGoogle described two eighth-generation TPU designs. The 8t emphasizes training throughput across large pods, while the 8i adds on-chip SRAM and specialized collectives for inference.
- Milestone
Google announced TPU 8tx for confidential inference and TPU 8i Max for long-context inference. The 8i Max pairs two 8i chips on one package, increasing memory bandwidth and SRAM per package.
See also (6)
Open full graphGoogle Tensor Processing Units (TPUs)