Tensor Processing Unit (TPU).
Proposed by @kierisi
Card 1 of 259. Answer: Tensor Processing Unit (TPU).
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
Tensor Processing Units (TPUs), custom chips Google designed to run neural networks. AlphaGo, built by Google DeepMind, beat Lee Sedol 4 games to 1 in Seoul from 9 to 15 March 2016, with TPUs evaluating its policy and value networks. Google revealed the chip only at Google I/O in May 2016, when engineer Norm Jouppi wrote that TPUs powered AlphaGo in those matches.
The first-generation TPU is an application-specific integrated circuit (ASIC) built for inference: it runs networks that are already trained, using a matrix unit of 65,536 8-bit multiply-accumulate cells. Google said TPUs had been in its data centers for more than a year. The AlphaGo that beat Fan Hui in October 2015 had run on 1,202 CPUs and 176 GPUs; DeepMind's 2017 AlphaGo Zero paper says the Lee Sedol version used 48 TPUs.
A typical follow-up is how a TPU differs from a GPU: it trades general-purpose flexibility for fast, power-efficient matrix multiplication. The first TPU could not train networks; TPU v2, announced in 2017, could.
