Trident-CIM: A LUT-Based Compute-in-Memory Macro With Trident Read Bit-Line and Partial Product Pruning

With the massive data transfer of deep neural network (DNN) models hindering their employment in energy-stringent scenarios, compute-in-memory (CIM) architectures boost energy efficiency with memory-logic coupling. The read-out computing logic and adder tree are the main contributors to overall CIM power consumption. This paper proposes a Trident-CIM macro, which implements LUT-based multiply-and-accumulate (MAC) computation. Assisted by read-out cells featuring trident read bit-lines, read-out power is shared among several memory rows. Moreover, with distribution-aware partial product pruning implemented with leading-one-bit detection and truncation, overall MAC power reduction is achieved. A 16kb Trident-CIM macro is implemented in 28nm CMOS technology for energy efficiency validation. Simulation results show that the proposed LUT-based MAC scheme with partial product pruning achieves an energy efficiency of 40.6 TOPS/W (@200MHz) and 37.9 TOPS/W (@400MHz) featuring INT8 data precision.