What is the name of Google's custom ASIC designed for machine learning?

Google’s custom ASIC designed for machine learning is the Tensor Processing Unit, commonly abbreviated TPU. It is a specialized neural-network accelerator developed by Google for workloads involving tensor operations, especially those used in machine learning. Unlike a general-purpose CPU or graphics-focused GPU, a TPU is designed around the mathematical patterns common in neural-network computation.

Google began using the first TPUs internally in 2015 and publicly described the technology at Google I/O in May 2016. The original chips were built for inference—the stage where a trained model produces predictions—and were closely associated with TensorFlow. Their architecture used large matrix-multiplication resources, including a systolic-array design, to deliver high throughput with relatively efficient power usage.

TPUs are not synonymous with every machine-learning processor. “ASIC” describes the chip’s category: an application-specific integrated circuit. “GPU” is a different processor type that can also accelerate machine learning, while “FPGA” is reconfigurable hardware. Google later introduced TPU generations for both training and inference, Cloud TPU access for developers, and Edge TPUs for lower-power on-device workloads such as embedded vision.