Early Reliability Estimation in Hardware Accelerators using Improved Colored Petri Nets

This work exploits Colored-Petri-Nets (CPN) for the early reliability estimation of hardware accelerators, significantly reducing the complexity during early design stages aimed at safety-critical systems. Our method builds high-level models of complex hardware accelerators to estimate reliability, integrating circuit characterization and fine-grain fault simulations on fundamental structures. We evaluate our methodology using six architecture variants of an on-chip hardware accelerator for deep learning (GPUs' Tensor Cores). The results demonstrate that our approach reduces evaluation costs by 118x, achieving accuracy levels of up to 93.5% compared to exhaustive RT-level fault injection campaigns, while enhancing engineering productivity for early-stage designs.