BTI Aging Influence in SRAM-based In-Memory Computing Schemes and its Mitigation

In-Memory Computing (IMC) schemes are constantly gaining interest in the area of Neural Networks (NN) for performing computations directly in memory, aiming to overcome the data transfer constrains of traditional Von-Neumann architectures. Static Random Access Memory (SRAM) based IMC schemes are commonly used for NN performance boost. However, Bias Temperature Instability (BTI) aging affects the reliability of SRAM-based IMC topologies. In this paper, we explore the BTI aging influence on differential inputs SRAM-based IMC architectures, and we present a BTI aging mitigation technique to enhance their reliability and extend their lifetime. According to the proposed technique, aging mitigation is achieved by periodically reconfiguring the active current paths in the IMC cells, without any silicon area cost and with negligible performance degradation. We show that the proposed aging mitigation technique prolongs by higher than 100% the lifetime of SRAM-based IMC architectures with differential inputs.