tinyML Talks: CFU Playground: Customize Your ML Processor for Your Specific TinyML Model

tinyML Talks CFU Playground: Customize Your ML Processor for Your Specific TinyML Model Tim Callahan Staff Software Engineer Google CFU (Custom Function Unit) Playground lets you build your own specialized & optimized ML processor based on the RISC-V ISA, implemented on an FPGA using a fully open source stack. This talk isn’t about general ML extensions; it’s about a methodology for building your own extensions specific to your tinyML model. The extensions can range from a few simple new instructions, up through a complex accelerator that interfaces to the CPU via a set of custom instructions; we will show examples of both. CFU Playground bundles the TensorFlow Lite for Microcontrollers library, a number of sample tinyML models, and profiling & test software, so that you can immediately build an initial processor, evaluate inference performance on your FPGA board, and start customizing. CFU Playground is designed for rapid iteration and design space exploration; you