tpu

Project creating a General Matrix Multiply (GEMM) unit and corresponding verification framework. A future attempt could be to extend this into a full-out TPU. The project takes inspiration from Google's TPU.

See also the Acknowledgements section below.

Overview

src

Here, you can find the hardware designs. They are described using Verilog. The goal is to enable efficient matrix multiplication using systolic arrays. The top-level module is SystolicMatrixMultiplier.v.

You can find more through explanations in the respective files.

tests

Here, you can find the verification of the hardware designs. To this end, cocotb was employed.

Tests can be run via the following interface:

python test_runner.py <path_to_dir_with_verilog_designs> <name_of_module_to_be_tested> (optional:) <parameter_name>=<parameter_value>

For example:

python test_runner.py ../src SystolicMatrixMultiplier or

python test_runner.py ../src SystolicMatrixMultiplier N=16 OP_WIDTH=4

docs

Here, you can find documentation or additional information regarding the project.

Acknowledgments

This project was conducted as part of the “Creation of Deep Learning Methods“ Practical Course at the Technical University Munich. It is loosely based on prior work from the same course. Special thanks to Vladimir Golkov and Dirk Stober who provided guidance and supervision throughout the project.