What happens when you quantize a NAS-discovered Pareto front to INT4?
A new three-stage pipeline maps neural architectures from search to reconfigurable hardware without baking quantization into the search loop. The study shows that an FP32 zero-shot surrogate covers the INT4 Pareto space better than a surrogate trained on quantized data.
Neural architecture search for edge devices folds quantization into the search itself. That coupling balloons the search space and locks architecture choices to a specific bit-width. The paper accepted to the 39th IEEE International System-on-Chip Conference in Heidelberg takes the opposite route: search first in full precision, then quantize, then map to hardware.
The pipeline has three stages: a hardware-agnostic Pareto rank surrogate on NAS-Bench-201, a quantization bridge with Pareto-aware filtering and feedback control, and an evolutionary domain space exploration backend targeting CGRA4ML for final hardware mapping.
The sources this article was written from
- "arXiv:2608.13293v1
https://arxiv.org/abs/2608.13293v1"