Bespoke

Bespoke is a system for synthesizing FPGA stream processors from 8-bit quantized neural network specifications. For edge ML applications, power efficiency and speed are paramount. The system described presents an alternative to run- ning intensive ML routines on microcontrollers, allowing for edge systems to use cheaper and more power efficient microcontrollers alongside a bespoke FPGA that can be dynamically reconfigured to suit any ML stream-processing task. The model could even be modified or tuned on the edge by the microcontroller. Suitable applications include sensor data processing, state estimation and control, forecasting, and decision making