Fast Algorithms for Exact IR Drop De-Embedding in Analog Multiply-Accumulate Computing

IR drop that comes from the line resistance is a well-known issue of the crossbar structure. When the latter is used in the in-memory computing, this issue sets the limit for the analog computing parallelism. In this work, the previously unappreciated linear network modeling is proven to be sufficient for its calibration. Fast algorithms exploiting this simplification are proposed for an exact and efficient de-embedding of the IR drop. Physical pictures are offered for a more intuitive understanding. Scaling in the input and output peripherals is proposed to overcome the dynamic range overflow issue while maintaining the memristor cell precision. An all-in-one-chip MNIST demonstration is presented. Neural network scheduling and gear-like convolution are proposed for real-time processing. Full MNIST test set with the calibration shows a 98.70% accuracy, with 0.45% loss from the software baseline and a 3.58% improvement upon the without-calibration case. Besides inference deployment, these methods may also be generalized to edge learning where the computation resources are limited.