CIM-SECDED: A 40nm 64Kb Compute In-Memory RRAM Macro with ECC Enabling Reliable Operation

Abstract

Resistive RAM (RRAM) is a promising candidate for compute in-memory (CIM) applications owing to its natural multiply-And-Accumulate structure in a 1T-1R bitcell, high-bit density, non-volatility, and voltage and process compatibility. These properties seek to advance applications such as AI with higher throughput and bit-density. However, due to process, temperature, and write-To-write variations the resistive state of each RRAM undergoes both spatial and temporal variations. Significant effort has been made to reduce the impact of device variation using iterative write verify (IWV) or training-Aware approaches [1]. Unfortunately, traditional ECC is not compatible with CIM when multiple cells are read simultaneously on the same bitline. To address this issue at the circuit level, this paper presents a 64Kb RRAM macro in 40nm CMOS supporting SECDED (single error correction, double error detection) scheme compatible with CIM for any number of parallel row accesses. Compared to prior work, our results indicate that CIM-SECDED (1) improves bit error rate (BER) by up to 69.2 \times for compute in-memory (2) relaxes the constraints on resistance variations and directly lowers IWV and write voltages. As a result, when applied to AI workloads we achieve (1) 24.4% (29.9%) accuracy improvement on the CIFAR10 (ImageNet) dataset (2) and consequently, improved endurance though lowering write voltage requirements [2]. © 2021 IEEE.