Surrogate Neural Architecture Codesign Package (SNAC-Pack)

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, extended

Summary

The Surrogate Neural Architecture Codesign Package (SNAC-Pack) is an open-source AutoML framework for hardware-aware neural architecture codesign and end-to-end FPGA deployment. It addresses the challenge of optimizing models for multi-dimensional FPGA resource budgets (LUTs, DSPs, FFs, BRAM, latency) by integrating a hardware surrogate model (rule4ml, wa-hls4ml GNN) into a multi-objective global search using Optuna and NSGA-II. This surrogate model provides fast resource and latency estimates, avoiding the high cost of full synthesis during the search loop. A subsequent local search stage applies quantization-aware training (QAT) and iterative magnitude pruning, with final models synthesized via the hls4ml Python library. Configurable via YAML and an optional agentic frontend, SNAC-Pack demonstrated its effectiveness on jet classification at the Large Hadron Collider and superconducting qubit readout, discovering compact architectures that matched or exceeded baselines while significantly reducing FPGA resource utilization and, for qubit readout, cutting design exploration from months to hours.

Key takeaway

For AI Hardware Engineers or Machine Learning Engineers deploying models to FPGAs, SNAC-Pack provides a critical tool for optimizing architectures under multi-dimensional hardware constraints. If you are struggling with long design cycles or poor correlation between software metrics and actual FPGA costs, this framework can automate hardware-aware neural architecture search, compression, and synthesis. You can achieve compact, high-performance models, potentially reducing exploration time from months to hours, as demonstrated in qubit readout. Consider integrating SNAC-Pack to streamline your development process.

Key insights

SNAC-Pack integrates surrogate hardware models into multi-objective NAS for efficient, hardware-aware FPGA deployment.

Principles

• Hardware-aware NAS needs fast, accurate feedback.
• Multi-dimensional FPGA costs require specific optimization.
• Surrogate models reduce search loop synthesis cost.

Method

SNAC-Pack uses multi-objective global search with surrogate hardware estimates, followed by local QAT and iterative magnitude pruning, then hls4ml synthesis.

In practice

• Automate FPGA model design for resource-constrained edge devices.
• Accelerate jet classification at LHC with sub-microsecond inference.
• Optimize superconducting qubit readout for fidelity and hardware budgets.

Topics

• Neural Architecture Search
• FPGA Deployment
• Hardware-Aware AI
• Surrogate Models
• Quantization-Aware Training
• Model Pruning
• hls4ml

Code references

• fastmachinelearning/nac-opt
• fastmachinelearning/hls4ml

Best for: AI Engineer, Research Scientist, Machine Learning Engineer, AI Hardware Engineer, AI Scientist

Related on AIssential

• SNN-MLIR: An MLIR Dialect for Compiling Neuromorphic SNNs from NIR to Bare-Metal C — SNN-MLIR unifies SNN compilation from NIR to bare-metal C, supporting both floating-point and quantized data.
• Gemma 4 QAT models: Optimizing model compression for mobile and laptop efficiency — Quantization-Aware Training (QAT) significantly reduces LLM memory footprint while preserving quality for on-device deployment.
• A new version of OpenAI’s Codex is powered by a new dedicated chip — Dedicated AI hardware accelerates lightweight models for real-time developer collaboration and rapid iteration.
• S2LC – 100 LoRA adapters in 3.59ms by reconstructing weights in GPU registers, never writing to HBM — S2LC (Shared Spectral Low-Rank Compression) is a novel method that exploits shared spectral structure across neural network modules to efficiently serve numerous LoRA adapters.
• ITP-STDP: An Intrinsic-Timing Power-of-Two Learning Engine for On-Chip SNN Training — ITP-STDP significantly reduces SNN on-chip training overheads via algorithmic and hardware optimizations.
• A Synthesizable RTL Implementation of Predictive Coding Networks — Predictive coding offers a local, hardware-friendly alternative to backpropagation for distributed online learning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.