Researchers shrink FastGRNN to 566 bytes for real-time AI on Arduino and MSP430 microcontrollers

The dominant trend in machine learning is scaling up models and hardware, but a new paper from Emre Can Kizilates takes the opposite approach: compressing neural networks to run on the world's most constrained microcontrollers. The work reproduces and extends FastGRNN, a compact gated recurrent cell, deploying it on bare-metal 8-bit Arduino (ATmega328P, 2KB SRAM) and 16-bit MSP430 (512B SRAM, no hardware multiplier). The compression pipeline combines low-rank weight factorization, iterative hard-thresholding sparsity, and per-tensor Q15 post-training quantization with explicit activation calibration.

The resulting model occupies just 566 bytes of weights and achieves macro F1 = 0.918 on the HAPT activity recognition dataset—matching a PyTorch reference with 100% prediction agreement across 3,399 test windows. Both platforms sustain real-time 50Hz streaming inference (9.21ms per sample on Arduino, 13ms on MSP430). A key innovation is a 256-entry sigmoid/tanh lookup table that delivers a 30.5x speedup on the multiplier-less MSP430. The paper also characterizes recurrent warm-up latency (median 74 samples, 1.48s) and hardware energy: 17.7mW active power, <0.09mW idle, and 96.7% energy reduction with the LUT. This shows that capable recurrent networks can run on ultra-low-power edge devices without any cloud dependency.