Temperature-Based Tuning for In-Memory Matrix Multiplication

Abstract

A technique for performing in-memory matrix multiplication that takes into account temperature variations within the memory is disclosed. In one example, a matrix multiplication memory uses ohmic multiplication and current addition to perform the dot products involved in the matrix multiplication. One drawback of this analog form of multiplication is that temperature affects the accuracy of the result. Therefore, a technique is provided herein that compensates for the effect of temperature increases on the accuracy of the in-memory matrix multiplication. According to this technique, portions of an input matrix are classified as valid or invalid. The valid portions are mapped to a low-temperature region of the in-memory matrix multiplier, and the invalid portions are mapped to a high-temperature region of the in-memory matrix multiplier. Then, the matrix multiplication is performed.