STAR-PIM: Self-Test and Repair Structure for Processing-in-Memory With Adder Tree-Based MAC

Le résumé fourni par la source

Processing-in-memory (PIM) architectures alleviate memory bottlenecks and improve latency and energy efficiency for AI and ML workloads by accelerating general matrix-vector multiplication (GEMV) operations in DNNs. However, permanent faults in arithmetic units (AUs) within processing units (PUs) critically impact yield and inference accuracy. Although the hybrid built-in self-test (HBIST) method has been proposed, it has limited capabilities in diagnosing and repairing faulty AUs within PUs. In this study, a novel Self-Test And Repair structure for PIM (STAR-PIM) is proposed to enable both fault diagnosis and repair by incorporating a bypass mechanism. A scan-path-like approach enables the testing and precise localization of faulty AUs, while faulty adders are bypassed using a redundant adder structure integrated within the memory die. Furthermore, faulty multipliers are masked using the weight-swapping logic. Experimental results demonstrate that STAR-PIM achieves high AU-level test coverage, ranging from 98.89% to 100% with reasonable area overhead. Recovery experiments show that STAR-PIM maintains low relative errors under fault rates up to 1% for GPT-2 and preserves inference accuracy under fault rates up to 3% for MNIST-MLP. Power measurements on GDDR6-AiM indicate an average overhead of 7.39% with only a 0.08% latency increase. Consequently, STAR-PIM significantly enhances the yield and reliability of PIM while reducing test costs, making it a highly practical solution.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.