GEMMHeal: An Efficient Self-repairing Architecture for Matrix Multiplication

In modern machine learning models like Transformers, matrix multiplication dominates most computation. Specific hardware often uses large-scale PE arrays, such as systolic arrays, to accelerate this process. However, these extensive PE arrays tend to experience high fault rates, with cumulative errors in matrix multiplication potentially impacting final algorithmic outcomes. We propose GEMMHeal, a self-repairing, systolic-array-like architecture optimized for matrix multiplication. This architecture optimizes the matrix multiplication array design and data flow in the buffer while decoupling the error correction (EC) core from the array. The EC core operates as an independent computation unit, maintaining resource efficiency and avoiding performance degradation due to the spatial locality of faults. Additionally, our proposed design incorporates redundancy within the EC core, allowing it to self-repair and thus further improve fault tolerance. GEMMHeal is implemented in a 32nm process to demonstrate its area and physical achievability. A single EC core occupies only 1.8% of the area while effectively repairing up to 6.3% of faulty PEs, all while maintaining overall performance. By accounting for potential faults within the EC core, our redundant design achieves an average improvement of 31% in the EC core normal rate compared to non-redundant designs when the computation cell fault rate is within 15%.