Reproduce “TinyML Hits the Memory Wall: A Benchmark of Ten Sub-Watt Edge Devices” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Pietro Bonazzi, Luigi Capogrosso, Ahmet Celik, Rosanna Manzo, Marco Giordano, Szymon Ruciński, Jonathan Hungerbuehler, Michele Magno
The authors compare processor latency to test whether CNNs are limited by computation or data movement. Across four accelerators and 64×64 CNNs, latency tracks estimated weight and activation bytes more closely than multiply-accumulate counts. On Alif E7, SqueezeNet uses 8.5× as many operations as MobileNetV3 yet runs 1.8× faster. This supports memory-bound behavior for tested CNNs, not larger models or resolutions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.