An energy-efficient floating-point compute SRAM with pipelined in-memory bit-parallel exponent and bitwise mantissa processing

The promise of compute-in-memory (CIM) for energy-efficient deep neural network (DNN) tasks has been demonstrated. However, most previous CIM works typically focus on low-precision DNN computing. To enable high-precision DNN computing, this work presents a novel SRAM-CIM design that fully supports half-precision floating-point (FP16) MAC operations. To maximize the energy efficiency, an efficient in-memory bit-parallel approach for conducting exponent operations and the bitwise in-memory booth encoder for reducing mantissa multiplication latency are proposed. Moreover, by enabling the pipeline of exponent and mantissa processing, the hardware utilization is improved with high throughput achieved. The proposed design is analyzed in 40 nm CMOS technology. The evaluation shows that the SRAM-CIM achieves a frequency of 714 MHz and a peak energy efficiency of 1.53 TFLOPS/W.