EMO-CIM: An Input/Stationary-Data Similarity-Aware Computing-In-Memory Design for Variable Vector-Wise Computation in Edge Multioperator AI Acceleration

We propose an edge multi-operator computing-in-memory design (EMO-CIM) that supports variable vector-wise multiply-and-accumulate (MAC) in CNN, Depthwise-Convolution (DW), and Attention operators. It features: (1) A single EMO-CIM bank excels in variable vector-wise MAC for multi-operators; (2) Merging local input-shared compute units with a decode-unit & adder-tree facilitates input/stationary-data similarity-aware computing to improve energy and area efficiency; (3) Based on a row-wise vector separable accumulation strategy, Attention/CNN task is optimized via multi-vector parallel crossover computation, while DW task is optimized through the partial-convolution-first computation. EMO-CIM achieves peak energy efficiencies of 93.5 and 39.3 TOPS/W for Attention/CNN and DW tasks, respectively, and area efficiencies of 4.971 and 3.728 TOPS/mm2 for the same tasks under INT8 MAC operations.