Self-Calibrating Analog Circuitry for Softmax-Scaled Function With Analog Computing-In-Memory

Mar 1, 2026 by Linjun Jiang, Yitong Zhou, He Zhang, Wang Kang (IEEE Transactions on Very Large Scale Integration (VLSI) Systems)

DOI 10.1109/TVLSI.2026.3651307

We built a self-calibrating analog front-end that natively implements softmax-like scaling in computing-in-memory by mapping exponentiation and division into successive subthreshold transistor circuits, then dynamically tunes biases to kill mismatch and keep accuracy—no expensive digital exponent/divide blocks needed. The result is a compact, robust ACIM that pushes energy and area efficiency for softmax-enabled inference.

source S2, crossref