Tiny TPU is a small-scale, FPGA-based implementation of Google's Tensor Processing Unit. The goal of this project was to learn about the end-to-end technicalities of accelerator design from hardware to software, while deciphering the lower level intricacies of Google's proprietary technology. In doing so, we explored the possibility of a small scale, low power TPU.
A more detailed report is located in /docs/report/report.pdf.
The TPU currently doesn't have all of the hardware validated, but is close to having a complete instruction decoder for an instruction set, as well as functional accumulator tables for large matrix multiplication. Right now it can only do up to 16x16 matrix multiplication on the synthesis branch.
The stable version is the latest commit on the synthesis branch, as master is left open to continue progress on the project. This project was synthesized on Quartus 15.0 and programmed onto an Altera DE1-SoC FPGA. No other setup is guaranteed to work. Synthesis requires the necessary Cyclone V libraries to use the IP blocks instantiated in the Verilog.
The Verilog files for synthesis are located in /hardware/synthesis and the Verilog files are located in /hardware/hdl.