An FP-CIM Macro for Edge AI With a Local Maximum Search Array and a Gated Alignment-Accumulation Architecture

While Floating-Point Compute-in-Memory (FP-CIM) offers wider applicability for edge AI than its integer counterpart, its complex exponent and mantissa handling incurs significant power, performance, and area overheads. This paper presents an FP-CIM macro that overcomes these challenges through two key innovations. For exponent handling, we introduce a Local Maximum Search Array (LMSA) that identifies the maximum value without external comparators. For mantissa processing, an energy-efficient scheme driven by a Gated Accumulation Controller (GAC) and a sparsity aware Zero-Skipping Adder Tree (ZSAT) replaces power-intensive barrel shifters. This integrated design achieves concurrent mantissa alignment and accumulation, significantly reducing computational overhead. Validated on a 28nm CMOS prototype, the proposed macro achieves an energy efficiency of 30.09 TFLOPS/W at 344.8MHz and 700mV, a 1.84× improvement over prior works.