A fault-tolerant PE array based matrix multiplier design

Matrix multiplication comprises a significant portion of computation efforts in many engineering and scientific applications. Realizing matrix multiplication in a two-dimensional PE (processing element) array improves computation efficiency; however, reliability and yield issues arise as the PE array complexity grows. This paper presents three fault tolerant schemes for the Cannon algorithm based two-dimensional PE array matrix multiplier. In the twisted column scheme, free PEs take over the tasks of adjacent faulty PEs; this preserves the computation efficiency. The column replacement scheme reallocates the task of a faulty column to a fault-free one; this degrades the overall performance but significantly enhances fault tolerance. The hybrid approach combines the previous two schemes; it achieves the best fault tolerance and incurs no performance degradation if the number of fault PEs is small. Simulation results and overhead analyses are presented to validate the proposed schemes.