HDANNS: In-Memory Hyperdimensional Computing for Billion-Scale Approximate Nearest Neighbour Search Acceleration

Data retrieval from large-scale database has laid the foundation for numerous crucial applications, including search engines and recommendation systems. Recently, the significance of Retrieval-Augmented Generation (RAG) in enhancing the generalization and professionalism of Large Language Model (LLM) has become increasingly evident, with data retrieval being its cornerstone. To enhance data retrieval efficiency, Approximate Nearest Neighbour Search (ANNS) significantly reduces the computational complexity compared to brute-force accurate search, while maintaining satisfactory retrieval outcomes. However, for the existing mainstream schemes which use CPU/GPU for ANNS deployment, it is difficult to further optimize the retrieval efficiency. The extensive data interaction between CPU(GPU) and memory(video memory) not only constrains retrieval speed but also results in substantial energy consumption. In this work, an ANNS method based on hyperdimensional computing (HDC) is proposed for the first time, which simplifies the similarity calculation in retrieval to logical operations. To match the HDC-based ANNS (HDANNS) for efficient and high-precision retrieval, we propose a time-domain circuit based on Resistance RAM (RRAM) array, which uniquely integrates data storage and similarity computation. In addition, a comprehensive hardware architecture supporting retrieval on billion-scale dataset is also proposed. Experimental results show that, compared to other most advanced designs, HDANNS and the proposed hardware architecture achieves 4-7.58$\boldsymbol{\times}$ throughput.