A heterogeneous microprocessor for energy-scalable sensor inference using genetic programming

Abstract

We present a heterogeneous microprocessor for IoE sensor-inference applications, which achieves programmability required for feature extraction strictly using application data. Acceleration, though key for energy efficiency, poses substantial programmability challenges. These are overcome by exploiting genetic programming (GP) for automatic program synthesis. GP yields highly structured models of computation, enabling: (1) high degree of specialization; (2) systematic mapping of programs to the accelerator; and (3) energy scalability via user-controllable approximation. The microprocessor (130nm) achieves 325×/156× energy reduction, and farther 20x/9x energy scalability, for programmable feature extraction in two medical-sensor applications (seizure/arrhythmia-detection) vs. GP-model execution on CPU. The energy efficiency is 220 GOPS/W, near that of fixed-function accelerators, exceeding typical programmable accelerators.