Tensor Processing Unit
Google's custom ASIC for neural network machine learning.
The Tensor Processing Unit (TPU) is a neural processing unit (NPU) application-specific integrated circuit (ASIC) developed by Google for neural network machine learning. It is designed for high-volume, low-precision computation with more input/output operations per joule than a graphics processing unit, and lacks hardware for rasterization or texture mapping. TPUs are well suited for convolutional neural networks and are often used for inference in transformer-based neural networks for large language models.
Quick Facts
- Designer
- Introduced
- 2015
- Version
- 8
Facts from the source article.
Lore & Background
Google began using TPUs internally in 2015, and in 2018 made them available for third-party use, both as part of its cloud infrastructure and by offering a smaller version of the chip for sale. The TPU was announced in May 2016 at Google I/O, with the company stating it had been used inside their data centers for over a year. The chip was specifically designed for Google's TensorFlow framework, though as of 2017 Google still used CPUs and GPUs for other types of machine learning. The TPU shares a genealogy with systolic array systems including the WARP, and three separate groups at Google were developing AI accelerators, with the TPU's systolic array design being the one ultimately selected. Broadcom is a co-developer, translating Google's architecture into manufacturable silicon and managing fabrication through foundries like TSMC.
Reader's Guide
The Tensor Processing Unit established a foundational platform for neural network inference at scale across Google's production services. A 2017 paper demonstrated that the TPU achieved 15–30× higher performance and 30–80× higher performance-per-watt than contemporary CPUs and GPUs. Google's TPUs are proprietary; some models are commercially available, and access is offered through Google Cloud Platform, Kaggle, and Colaboratory. They were used in the AlphaGo versus Lee Sedol series, the AlphaZero system, Google Street View text processing (finding all text in the database in less than five days), Google Photos (processing over 100 million photos per day per TPU), and RankBrain for search results. As of 2025, Google Cloud generates product revenues primarily from the sale of TPU systems, and Google has been in talks with neoclouds and Meta about deploying TPUs in their data centers.
Did You Know?
- The first-generation TPU is an 8-bit matrix multiplication engine driven by CISC instructions over PCIe 3.0.
- The second-generation TPU introduced the bfloat16 format and can calculate in floating point, making it useful for both training and inference.
- A single TPU v4 pod contains 4,096 chips and has 10x the interconnect bandwidth per chip at scale compared to any other networking technology.
- The seventh-generation TPU, called Ironwood, was unveiled in April 2025 and has a peak computational performance of 4,614 TFLOP/s.
More in Microprocessors 1-24
Spotted an error? Know more?
Reader corrections go straight into our review queue. Suggest an edit · How this site is sourced