GIM (Ghost in the Machine): A DSP-Inspired Accelerator Platform for Exploring Machine-Learning Architectures

Machine-learning (ML) algorithms are finding wide adoption across a rich spectrum of application domains with diverse requirements in terms of performance, power, and cost. Complete ML systems often have a DSP front end, extracting features for the inference engine. Early work in ML drew inspiration from now common DSP algorithms like adaptive filters. Bringing a signal processing focus to an ML accelerator can have benefits from exploring and reasoning about architectural techniques like pipelining, to utilizing new coarse-grained FPGAs containing hundreds and thousands of DSP slices with dedicated local storage. These new coarse-grained architectures allow us to achieve ASIC-like clock rate and reductions in power while rapidly exploring novel and common ML architectures.