Double Column Error Correction in RRAM Matrix Multiplication Using Weighted Checksums

Resistive RAM (RRAM) technology has emerged as a viable candidate for artificial intelligence and machine learning applications due to its matrix multiplication capability through in-memory computation. However, RRAM technology is characterized by a high soft error rate, which compromises the reliability and accuracy of neural networks. In this paper, weighted checksums are used to devise an error correction scheme to correct any number of errors contained within any two columns of an RRAM array. The error correction mechanism operates directly on the output of the RRAM array, allowing detection and correction of multiple errors within these columns. Experimental results show that the proposed method achieves 96% coverage. Due to the high coverage, low power consumption, and low decoding latency, these codes are ideal for the high throughput and accuracy needs during training and inference of AI models.