wenqijiang/fast-vector-similarity-search-on-fpga

Fast-Vector-Similarity-Search-on-FPGA is a research project that speeds up a common task in modern computing: finding the most similar items in a massive collection of data. This is the kind of work that powers recommendation engines, image search, and other AI applications. Instead of relying on standard computer processors, this project uses specialized hardware to do the math much faster, aiming to deliver quicker results for systems dealing with large-scale data. At a high level, the project is built for an FPGA, which is a type of computer chip that can be physically reprogrammed to excel at a specific job. In this case, the chip is configured to handle Approximate Nearest Neighbor Search, a technique for quickly estimating which items in a dataset are most similar to a query. The repository provides two versions of this setup: one designed to process requests over a network using standard internet protocols, and a local version that operates without the networking component. The primary audience for this project includes researchers and engineers working on high-performance search infrastructure. For example, a team building a large-scale image retrieval system might use this approach to search through millions of images in a fraction of the time it would take standard hardware. Because configuring an FPGA from scratch is a time-consuming process, the project includes pre-built bitstreams, the finished, ready-to-load hardware configurations, so users can get the system running without spending hours building it themselves. The repository also includes materials related to its origin as an academic submission, featuring baseline experiments using a well-known search library and scripts for generating performance plots. While the core code is written in C++, the overall focus is on demonstrating a hardware-level performance advantage rather than providing a simple software library. The README doesn't go into detail about the specific setup requirements or benchmarks, but the included materials point to a project designed for specialized, high-speed data processing.