Redundant Number-based NBTI Stress Reduction for Lifetime Resilience Enhancement of Neural Processing Engines

Nowadays, deep neural networks (DNNs) are being applied for safety-critical applications such as automotive and aerospace while emphasizing the high significance of lifetime resilience. Conversely, hardware accelerators have been employed for the efficient execution of complex DNNs on edge devices with resource constraints. Meanwhile, enhancing computation efficiency through redundancy elimination, coupled with technology feature size scaling, can elevate the vulnerability to aging, reducing the lifetime resilience of DNN accelerators. Previously, designers relied on conservative guard bands to prolong the lifetime of CMOS devices, albeit at the expense of performance loss. This paper proposes the serial processing approach for neural networks based on the redundant number system, which improves lifetime resilience without losing accuracy or performance but, with a few area and power overheads. We explore the number-system effect on BTI (Bias Temperature Instability) stress and consequent aging degradation. Experimental results of DNNs execution illustrate that the proposed computing approach mitigates stress by 36% on average leading to a 35.5% higher lifetime over the baseline.