Investigating Energy Bounds of Analog Compute-in-Memory with Local Normalization

Feb 8, 2026 by Brian Rojkov, Shubham Ranjan, Derek Wright, Manoj Sachdev

We show how a cheap local normalization scheme—Gain-Ranging MACs—lets analog compute-in-memory keep the heavy lifting in low-precision analog while extending input dynamic range by ~4 bits and cutting ADC requirements, so you get LLM-friendly floating formats without killing energy efficiency. If you care about pushing analog CIM into real-world edge AI, this gives a practical energy-scaling trick that beats the usual shared-scale pain points.

source S2