CV-CIM: A Hybrid Domain Xor-Derived Similarity-Aware Computation-in-Memory Supporting Cost-Volume Construction

Cost-volume construction, a cornerstone of stereo vision processing, has been widely utilized across diverse domains such as robotics, autonomous vehicles, 3-D reconstruction, and augmented/virtual reality applications. However, this process necessitates substantial parameter sizes and continuous memory data accesses, imposing considerable demands on memory bandwidth. Computation-in-memory (CIM) emerges as a promising technology to alleviate this memory bottleneck by its ability to execute parallel computations within structured arrays. Nevertheless, the direct deployment of cost-volume construction on CIM platforms faces several challenges, including functional limitations, reduced compute margins, and structural rigidity. To overcome these obstacles, we present a novel CIM architecture, named cost-volume CIM (CV-CIM), which incorporates the following key features. First, the xor-derived logic integrates fundamental arithmetic functions and heterogeneous CIM array (CA) function together to complete requisite cost computation. Second, the pixel-wise similarity is discovered to enlarge the bit sparsity. Then, sparse-mode CIM leverages the resulting sparsity to enlarge the xor-based computing margin. Third, dual degrees of freedom peripherals collectively contribute to latency reduction through dynamic row computation skipping (row peripheral) and the overlapping of data loading and computing (column peripheral). A test chip is fabricated in 28-nm technology and operates within a voltage range of 0.6–0.9 V and a frequency range of 50–286 MHz. Notably, CV-CIM demonstrates support for L0/L1/L2 cost computations and achieves a remarkable peak energy efficiency of 1158 TOPs/W.