A Novel Self-Repair Mechanism for Tiled Matrix Multiplication Unit

General Matrix Multiplications (GEMMs) are widely used in scientific and machine learning applications. Convolutional Neural Networks (CNNs) rely primarily on convolution operation, which is often expressed as a GEMM problem. To speed up GEMM operations, modern CPUs include hardware accelerators specifically designed for this purpose. One such hard-ware accelerator is the Tile Multiplication unit (TMUL), which is supported by the Advanced Matrix Multiplication (AMX®) instruction set in the Intel Sapphire Rapids microarchitecture. However, like other frequently used hardware components, the TMUL is susceptible to permanent or persistent faults due to aging, which can lead to Silent Data Corruption (SDC). Google and Meta have recently reported recurring SDCs. Algorithm-based error detection (ABED) is a highly effective method for detecting persistent faults in GEMM operations. Our previous work involved a novel hardware implementation of ABED in TMUL, which could locate faults but only identify the of Fending matrix column, not the exact cell. In this paper, we present a self-repair technique for persistent faults in GEMM based on the aforementioned ABED approach. Our solution hinges on software-based column avoidance. When the hardware-based ABED detects an error caused by a faulty column in the TMUL, our software redirects computations to circumvent the faulty column. This eliminates the need for additional hardware modifications to ABED. We demonstrate that our proposed approach can perform self-repair for single faults for all matrix sizes, from those smaller than the TMUL dimension to larger matrices. Furthermore, we show that our solution can even correct multiple faults with low-performance overhead, and we report the performance overhead for multiple fault repair using a clustered fault model. Unlike prior solutions, our approach performs online detection and repair, enabling real-time correction of both persistent and intermittent faults.