CRAM-Based Acceleration for Intermittent Computing of Parallelizable Tasks

Abstract

There is an emerging requirement for performing data-intensive parallel computations, e.g., machine-learning inference, locally on batteryless sensors. These devices are resource-constrained and operate intermittently due to the irregular energy availability in the environment. Intermittent execution might lead to several side effects that might prevent the correct execution of computational tasks. Even though recent studies proposed methods to cope with these side effects and execute these tasks correctly, they overlooked the efficient intermittent execution of parallelizable data-intensive machine-learning tasks. In this article, we present PiMCo—a novel programmable CRAM-based in-memory coprocessor that exploits the Processing In-Memory (PIM) paradigm and facilitates the power-failure resilient execution of parallelizable computational loads. Contrary to existing PIM solutions for intermittent computing, PiMCo promotes better programmability to accelerate a variety of parallelizable tasks. Our performance evaluation demonstrates that PiMCo improves the performance of existing low-power accelerators for intermittent computing by up to 8× and energy efficiency by up to 150×.