A Reconfigurable Floating-Point Compute-in-Memory With Analog Exponent Preprocesses

This letter presents a reconfigurable floating-point compute-in-memory (FP-CIM) macro that preprocesses the exponent in the analog domain, enhancing the energy efficiency of edge devices for the floating-point (FP) inference. The presented FP-CIM macro supports FP8 inference, while can be configured to BP16 precision in a segmented computation manner. Furthermore, a time-domain analog-to-digital converter facilitates the analog compute-in-memory (CIM) macro while improving energy efficiency by sharing the counter and quantizing in a coarse-fine structure. Fabricated in a 28-nm CMOS process, the presented FP-CIM macro achieves 314.6-TFLOPS/W energy efficiency and 12.13-TFLOPS/mm2 area efficiency at the FP8 mode.