A Memristor-Based Dual-Domain System for Overcoming Limitations of Traditional Analogue Compute-in-Memory
Analogue compute-in-memory (ACIM) systems leveraging memristor crossbar arrays offer exceptional energy efficiency and parallelism for neural network classification tasks but face critical limitations in regression tasks requiring high dynamic range and floating-point precision. The reliance on low-precision integer computations and susceptibility to noise accumulation restrict their applicability to complex regression problems such as object detection. This work presents the memristor-based AnDi architecture, a dual-domain system integrating ACIM arrays with digital cores to overcome these limitations. The architecture incorporates a dual-domain floating-point processor (DDFP) that enables seamless data conversion between analogue and digital domains. A quantization unit transforms floating-point feature maps into 8-bit integer (INT-8) representations with 15-bit floating-point (FP-15) scaling factors, while a dequantization unit recovers high-precision FP-32 outputs. Experimental validation demonstrates that during 256 × 256 floating-point matrix multiplication, the ACIM array consumes 98.7% of total energy, whereas the DDFP processor accounts for only 1.3%, achieving significant energy efficiency. The hybrid mapper flexibly distributes matrix multiplication across both domains with single row and column granularity, enhancing floating-point compatibility, accuracy, and array spatial utilization compared to pure ACIM systems. This dual-domain approach addresses the fundamental bottlenecks of traditional ACIM in floating-point compatibility, noise accumulation, and array utilization, enabling high-precision regression tasks while maintaining energy efficiency. The system integrates up to 12 memristor arrays and an RISC-V softcore CPU on a Xilinx FPGA, demonstrating a viable path for deploying advanced neural network applications in edge computing environments.