ACTION: Automated Hardware-Software Codesign Framework for Low-precision Numerical Format SelecTION in TinyML

Hamed Fatemi Langroudi, Vedant Karia, Tej Pandit, Becky Mashaido, Dhireesha Kudithipudi

Abstract: In this paper, a new low-precision hardware-software codesign framework is presented, to optimally select the numerical formats and bit-precision for TinyML models and benchmarks. The selection is performed by integer linear programming using constraints mandated by tiny edge devices. Practitioners can use the proposed framework to reduce design costs in the early stages of designing accelerators for TinyML models. The efficacy of various numerical formats is studied within a new low-precision framework, ACTION. Results assert that generalized posit and tapered fixed are suitable numerical formats for TinyML when the trade-off between accuracy and hardware complexity is desired.