Scaling NVMs in Event-Driven Architectures for Edge Inference

Abstract

Edge devices running machine learning (ML) inference, such as in automotive systems, drones, and wearables, face constraints in power, form factor, and cost, while needing high-performance computing. Increasing ML model sizes and on-chip SRAM memory requirements present bottlenecks for system efficiency, especially in advanced technology nodes. Emerging memory technologies and event-based AI computing have the potential to overcome these challenges. High-speed MRAM can reduce the power and area overheads of on-chip SRAM for neural network weights storage, and event-driven dataflow can lower inference latency and power consumption. This study models and analyzes the PPA scaling of SRAM and MRAM macros from a 22nm planar CMOS (N22), 12nm FinFET CMOS (N12), down to imec 14A degrees gate-all-around nanosheet CMOS (A14) technology node. We evaluate MRAM versus SRAM for weight storage in an event-driven multicore inference architecture, simulating different workloads. Our results show that STT-MRAM reduces total energy by similar to 25% and leakage energy by similar to 40% per inference in N22 and N12 nodes, though scaling to A14 VGSOT-MRAM increases energy consumption and latency. Despite this, MRAM weight storage reduces area by similar to 40% across all nodes. This work provides insights into facilitating the future development of edge computing technologies using emerging non-volatile memories.