Short answers, pulled from the story.
A Tensor Processing Unit is an application-specific integrated circuit that Google built for neural network machine learning, supporting frameworks including TensorFlow, Jax, and PyTorch. Google has used TPUs for tasks such as AlphaGo, AlphaZero, Street View text recognition, Google Photos, and the RankBrain search system.
Norman P. Jouppi served as tech lead and principal architect, leading the TPU's design, verification, and deployment to production in 15 months. Amir Salek was recruited in 2013 to build Google's custom silicon development capability, and Jonathan Ross was one of the original TPU engineers before founding the chip company Groq.
Google's 2017 paper reported the first TPU reached 15-30 times the performance of contemporary CPUs and GPUs, and 30-80 times their performance per watt. An April 2023 paper claimed TPU v4 ran 5-87% faster than a Nvidia A100 on machine learning benchmarks.
A Tensor Processing Unit is designed for high-volume, low-precision computation, sometimes as little as 8-bit, without the rasterization or texture-mapping hardware graphics processors carry. TPUs work best for convolutional neural networks, while GPUs suit some fully connected neural networks and CPUs suit recurrent neural networks.
The Edge TPU is Google's smaller, lower-power chip for running machine learning models outside the data center, capable of 4 trillion operations per second on 2 watts. Google made it available to developers in January 2019 under the Coral brand, and it also underlies the Pixel Neural Core in the Pixel 4 and the Google Tensor chip in the Pixel 6.
In 2019, Singular Computing, founded in 2009 by Joseph Bates, sued Google alleging its TPU chips infringed on patents covering a wide dynamic range for floating point numbers. Google agreed to a settlement with undisclosed terms in January 2024, after the number of patents in dispute had grown to eight.