MAFIA: Machine Learning Acceleration on FPGAs for IoT Applications
- Nikhil Pratap Ghanathe ,
- Vivek Seshadri ,
- Rahul Sharma ,
- Steve Wilton ,
- Aayan Kumar
FPL |
Published by IEEE
Recent breakthroughs in ML have produced new classes of models that allow ML inference to run directly on milliwatt-powered IoT devices. On one hand, existing ML-to-FPGA compilers are designed for deep neural-networks on large FPGAs. On the other hand, general-purpose HLS tools fail to exploit properties specific to ML inference, thereby resulting in suboptimal performance. We propose MAFIA, a tool to compile ML inference on small form-factor FPGAs for IoT applications. MAFIA provides native support for linear algebra operations and can express a variety of ML algorithms, including state-of-the-art models. We show that MAFIA-generated programs outperform best-performing variant of a commercial HLS compiler by 2.5x on average.