SNAC-Pack: Open-Source Framework for Hardware-Aware Neural Architecture Codesign on FPGA

The Surrogate Neural Architecture Codesign Package (SNAC-Pack) serves as an open-source AutoML framework tailored for hardware-aware neural architecture codesign and comprehensive FPGA deployment. It fills a critical gap by moving beyond existing NAS approaches that focus solely on accuracy or depend on proxy metrics like bit operations (BOPs), which often poorly reflect hardware expenses. This is particularly relevant for FPGA deployment, where costs are influenced by a complex budget involving lookup tables, DSPs, flip-flops, BRAM, and latency. Utilizing Optuna and NSGA-II, SNAC-Pack conducts a multi-objective global search, storing trials in a shared SQLite database to facilitate parallel processing across compute nodes. A hardware surrogate model estimates resources and latency per trial, thus eliminating the high costs associated with synthesis during the search process. Finally, a local search phase incorporates quantization-aware training and fine-tuning, enabling automated model design that adheres to actual hardware limitations.