How much precision does an edge classifier actually need?
Quality-configurable arithmetic units let an embedded neural network spend bits where they matter and save them where they do not. A short tour of what changes when accuracy is a knob.
Most embedded neural classifiers are over-provisioned. The arithmetic is fixed-precision because the silicon was fixed-precision, not because the workload demanded it. Make the arithmetic precision a runtime parameter and several things become possible at once: voltage scaling becomes safe, batteries last longer, and the same accelerator can serve a wider product line.
Our 2025 TVLSI paper introduces a multiply-accumulate unit whose precision can be reconfigured per layer, per inference, and even per channel. The trick is in the bias correction: lowering the precision of a MAC introduces a systematic error that, left uncompensated, accumulates across layers and destroys the classifier. We characterise this error analytically and pre-compute a correction term per layer that adds zero runtime cost.
The headline result, measured on a small CNN deployed for visual wake-words: 38 % energy reduction at the iso-accuracy operating point, with the option to trade further accuracy for a further 2.4× energy reduction when the device is on battery. The quality knob is a knob the application gets to turn, not a fixed design decision.
Comments & corrections
If you have questions about this note, or you spot something we got wrong, please write to the author directly. We post addenda to articles when a correction is warranted.