A Hybrid Monolithic 3D Integration of 2T0C DRAM and RRAM Chip for High-Precision In-Memory Point Cloud Acceleration with Ultra-Fine-Grained Dataflow

In this work, we propose the hybrid monolithic 3D (H-M3D) computing-in-memory (CIM) accelerator to address the challenges of frequent data granularity in existing point cloud acceleration on edge devices. By integrating and orchestrating high-precision 2T0C DRAM CIM and RRAM CIM for Euclidean distance (L2D) and matrix-vector multiplication (MVM) operations, we achieve efficient feature computation and refine the dataflow granularity to individual points, enabling high-speed data transmission. The proposed M3D-CIM point cloud system demonstrates <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.51 \times$</tex> speedup and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$2.56 \times$</tex> enhancement in energy efficiency, providing a novel solution for point cloud acceleration on edge devices.