OpenTPU: An Open-Source Reimplementation of Google's TPU

UC Santa Barbara's ArchLab has released OpenTPU, an open-source re-implementation of Google's Tensor Processing Unit (TPU). Based on details from Google's research paper, but lacking a formal specification, OpenTPU differs in several implementation details from Google's design. Currently supporting matrix multiplication and ReLU/sigmoid activation functions, OpenTPU is missing features like convolution and pooling. Implemented using PyRTL, the project includes hardware and functional simulators, along with a checker for verifying results. While lacking hard synthesis figures for a full 256x256 OpenTPU at this alpha release, its open-source nature offers valuable learning and improvement opportunities for researchers.