Retrospective on "In-Datacenter Performance Analysis of a Tensor Processing Unit" (AI column)

Interesting. note Two pages on how and why Google’s TPU came to be past-post Iron for ML/AI. TPU proved to be a great solution and supported product deep learning initiatives within Google. 10 years in succession, starting from the beginning 2015 The year he came in. Tomorrow there will be a bigger note about the entire history of the evolution of TPU, and for tomorrow I recommend reading this mini whitepaper, where there is such an explanation for the start of the project.

A key signal soon afterward was that matrix multiplication exceeded 1% of CPU fleet cycles in Google Wide Profiling. Another signal was the analysis by Jeff Dean (a Google Fellow, now the Chief Scientist) that processing a few minutes of speech or video by 100M users would require doubling or tripling the size of the CPU fleet. Other options were clearly required.

#AI #Infrastructure #Engineering #Architecture