Skip to main content
Back

UC San Diego's SNAC-Pack Cuts FPGA Neural Network Design From Months to Hours

SNAC-Pack automates FPGA-ready neural architecture search using hardware surrogate models, cutting months of manual tuning to hours

  • SNAC-Pack is an open-source AutoML framework that runs hardware-aware neural architecture search and produces synthesized FPGA firmware end-to-end.
  • A surrogate model predicts LUT, DSP, BRAM, FF, and latency for each candidate architecture, replacing expensive synthesis in the search loop.
  • Multi-objective search via NSGA-II in Optuna produces a Pareto front of accuracy vs. hardware cost tradeoffs, not a single model.
  • On LHC jet classification, the resource-optimized model cuts LUT usage by 78% and achieves 50 ns latency vs. 105 ns for the baseline.
  • On qubit readout, SNAC-Pack reduced design space exploration from months of manual tuning to a few hours of automated search.
  • Code is open-source on GitHub; configured via YAML with an optional AI agent (MCP) frontend.

Deploying neural networks on FPGAs is one of the most unforgiving engineering challenges in applied ML. Unlike CPUs or GPUs, where you can roughly predict cost from floating-point operations, FPGAs impose a five-dimensional resource budget: lookup tables (LUTs), digital signal processors (DSPs), block RAM (BRAM), flip-flops (FFs), and clock cycles. Get any one of those wrong and your design simply does not fit. A team from UC San Diego, ETH Zurich, and Fermi National Accelerator Laboratory has now released SNAC-Pack, an open-source AutoML framework that automates the entire pipeline from architecture search to synthesized FPGA firmware.

The gap that existing NAS tools leave open

Neural architecture search (NAS) has matured a lot, but almost all existing tools optimize for accuracy or use software-level proxies like FLOPs or bit operations (BOPs) as stand-ins for hardware cost. The hls4ml library is open-source software designed to facilitate the deployment of ML models on FPGAs, targeting low-latency and low-power edge applications. But hls4ml is a synthesis tool, not a search system. The missing piece was a framework that could explore thousands of candidate architectures and score them on real FPGA metrics without running a full synthesis pass (which can take minutes to hours per candidate) on every single trial.

Taking a neural network model as input, hls4ml generates C/C++ code designed to be transpiled into FPGA firmware by processing it with a high-level synthesis (HLS) library. That synthesis step is the bottleneck. SNAC-Pack solves this by inserting a fast surrogate model into the search loop that predicts resource utilization and latency without invoking the synthesizer, reserving full synthesis only for the final selected architectures.

How the pipeline works, step by step

SNAC-Pack system architecture diagram showing global search, local compression, and hardware synthesis stages

SNAC-Pack operates in two main stages: a global search and a local compression pass. Here is the full flow:

  1. Configure via YAML (or an AI agent): Users specify the dataset, search space (layer depths, widths, activations, regularization), optimization objectives, and target FPGA board in a YAML file. An optional Model Context Protocol (MCP) frontend lets agentic clients drive the same pipeline through natural language.