A new open-source tool cuts the time it takes to design AI models for tiny, cheap microcontrollers.
Researchers built ENAS, a hardware-aware neural architecture search framework, to find efficient model designs for microcontrollers without needing a GPU. It runs a three-stage search, random sampling, then top-K selection, then mutation, with a feasibility check that filters out designs a chip's memory can't handle. Tested on eight microcontrollers ranging from 20KB to 1MB of SRAM across two tasks, Visual Wake Words (person detection) and Melanoma Cancer (skin-lesion classification), ENAS found comparable models 2.41x and 1.70x faster than the NanoNAS framework on each task respectively, while using less peak RAM at matched accuracy. On one chip, the STM32H743, it reached 79.4% accuracy, 2.6 percentage points above a greedy CPU-only baseline.
TinyML developers building on-device cameras or sensors have mostly relied on either manual tuning or NAS tools that assume GPU clusters most edge teams don't have. Cutting search time on ordinary CPUs while trimming peak activation RAM, the actual constraint that determines whether a model fits on a microcontroller, targets a real deployment bottleneck rather than a leaderboard number. Releasing the framework as open-source lowers the cost of entry for smaller teams who can't afford dedicated ML infrastructure just to pick a model architecture.
Faster search is a genuine engineering win, but it's tested on two benchmarks and eight chips; whether ENAS's picks generalize to messier, real-world sensor data is still an open question.