Aging Mitigation in Systolic Array Accelerators: Balancing PE Loads for Enhanced Reliability

The systolic-array-based AI accelerator has been considered as one of the most promising architectures for data-intensive computation. However, the fixed systolic array size can lead to notable workload imbalance, with PEs utilization rate varying greatly, potentially reaching up to 14x disparity between maximum and minimum usages. The aging effects such as Hot Carriers Injection (HCI) and Negative Bias Temperature Instability (NBTI) can lead to timing errors or functional failures as the circuit operates for a certain while, especially in heavily utilized PEs. Consequently, the errors PE produced would propagate throughout the CNN computation, rendering the systolic array unusable and shortening its lifetime. Our paper proposes an aging-aware task deployment framework to increase utilization and mitigate aging effects. Experimental results demonstrate that our framework can outperform a conventional systolic-array-based AI accelerator 74.75% on prediction accuracy after 10 years of aging. In the meantime, our method can improve utilization up to 2.4x.