NMCu-CNN: A Scalable Near-Memory Computing Co-Processor on a RISC-V MCU in 22 nm FDSOI

This is the authors' abstract. We don't add key points for this paper.

This paper presents NMCu-CNN, a near-memory computing (NMC) co-processor for the hardware acceleration of convolutional neural networks (CNNs) on a low-power, flexible MCU platform. The scalable architecture and the selected platform enable adaptability to the rapidly evolving Edge AIoT landscape, where energy and performance requirements are constrained. Application-tailored NMC units, equipped with an optimized CNN dataflow, are integrated into the shared memory address space of a RISC-V-based MCU. The proposed architecture supports highly flexible runtime configurability, achieving $94 %$ computational efficiency. Fabricated in 22 nm FDSOI technology, NMCu-CNN delivers a performance of 203 GOPS and an energy efficiency of 1716 GOPS/W (1.7 TOPS/W) benchmarked on convolutional layers of a CNN model, outperforming the processing capabilities of other state-of-the-art techniques.