GIM (Ghost In the Machine): A Coarse-Grained Reconfigurable Compute-In-Memory Platform for Exploring Machine-Learning Architectures
Maya Borowicz, James Ding, Winnie Fan, Zhongqi Gao, Davis Jackson, Ares Lu, Sophia Rohlfsen, Ray Simar · 2024
Machine-Iearning (ML) algorithms are finding wide adoption across a rich spectrum of application domains with diverse requirements in terms of performance, power, and cost. These diverse requirements are making it necessary to explore a large space of ML architectures and reexamine fundamental computational structures, a process of exploration that is very expensive. To get around the costly computations associated with large data sets and long training times, there have been increasing investments in specialized fixed-function hardware. However, this specialized hardware is expensive and hard to generalize to address the spectrum of applications. For our experiments, we focus on a novel, highly parallel, superset ML architecture, and use it to test the capabilities of 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 exploring novel and common ML architectures.