Hybrid Non-Volatile Memory Based IMC Architecture for AI Edge Processors.

Abstract

Additionally, the proposed hybrid architecture leverages the complementary strengths of SRAM and ReRAM to address the limitations of conventional von Neumann systems, where frequent data movement between memory and processing units leads to significant latency and energy overhead. By integrating computation directly within the memory arrays, the in-memory computing (IMC) paradigm minimizes data transfer bottlenecks and enables parallel processing of large-scale neural network workloads. This is particularly beneficial for edge devices that operate under strict power and performance constraints.