Finding the right neural architecture for a microcontroller with a few tens of kilobytes of SRAM is not just a complexity-reduction problem: it is a physical constraint that changes how search itself is designed. ENAS, the hardware-aware Neural Architecture Search (NAS) framework presented by EdgeIntelligenceLab, tackles exactly this point. The project stands out because it does not require GPUs: search can run in resource-limited development environments, a detail that separates it from many existing NAS frameworks designed for GPU acceleration.

ENAS combines a static feasibility check, a cell-based search space with standard, depthwise-separable, and bottleneck blocks, optional skip connections, and a three-stage hybrid strategy: random, top-K, and mutation, with persistent cross-run caching. In practice, the framework can explore a set of valid architectures before training, discarding early those that do not fit within hardware limits.

Tests cover two TinyML benchmarks, Visual Wake Words and Melanoma Cancer, across eight microcontrollers with SRAM from 20 KB to 1 MB and nine input image resolutions. On Visual Wake Words, mean search time accelerates by 2.41 times compared to NanoNAS; on Melanoma Cancer by 1.70 times, while maintaining competitive accuracy. ENAS achieves 79.4% test accuracy on an STM32H743 microcontroller, outperforming a greedy CPU-only baseline by 2.6 percentage points. Moreover, selected models use substantially lower peak activation RAM at matched accuracy, a decisive fact because activation RAM is the binding constraint for microcontroller deployment.

The thesis that emerges from these results is that architecture search for TinyML is becoming a CPU-bound activity, not a data center one. If a NAS framework can achieve meaningful speedups without GPUs, the barrier to entry for embedded developers drops sharply. No cluster is needed to search for models; a workstation is enough, and data can remain in the lab or on the factory floor. This shifts incentives: value no longer lies in raw compute power, but in the ability to embed hardware constraints into the search pipeline. Reducing peak activation RAM is the key point, because on microcontrollers it is the limit that blocks deployment before parameter count does.

For embedded teams, especially in medical, industrial, and automotive sectors, independence from GPUs means faster and more private iteration, with less exposure of sensitive data to cloud services. Microcontroller vendors can benefit from a more automated design flow. Those betting solely on GPU acceleration for NAS may see their competitive edge narrow in this specific segment. The open-source release is not a detail: it allows verifying results and integrating the framework into CPU-only pipelines, including air-gapped environments. Those evaluating on-premise deployment on embedded devices can explore trade-offs with the analytical frameworks available on AI-RADAR.

The open question is whether the same GPU-free approach can scale to larger search spaces or bigger models. For now, ENAS shows that NAS for TinyML does not need a data center: it can live close to the silicon that will run the models.