faissgpu

faissgpu, also known as faiss-gpu, is the GPU-accelerated component of the FAISS (Facebook AI Similarity Search) library. It provides CUDA-based implementations for building and querying dense vector indices, enabling scalable nearest-neighbor search on large datasets. Faissgpu complements the CPU version by moving computation to NVIDIA GPUs to accelerate both index construction and search operations.

The GPU version supports many FAISS index types that are commonly used for scalable similarity search, notably inverted-file (IVF) based indices and product quantization approaches such as PQ and OPQ. These GPU implementations are designed to leverage parallelism in CUDA to speed up distance computations, quantization, and memory management. Faissgpu also offers mechanisms to transfer indexes and data between CPU and GPU memory and to run on single or multiple GPUs, enabling either straightforward acceleration or distributed processing across devices.

Interfaces for faissgpu include Python and C++ APIs that align with the core FAISS interfaces. Users can create or load a CPU index, move it to the GPU, train quantizers on a subset of vectors, add data, and perform similarity searches. The API supports common workflows such as building large indices, performing batched queries, and combining GPU acceleration with quantization to reduce memory usage and improve throughput, particularly for very large collections of vectors.

Requirements for faissgpu include a CUDA-capable NVIDIA GPU, appropriate drivers, and the CUDA toolkit. The Python package faiss-gpu can be installed via conda or pip. While offering substantial speedups, GPU-specific limitations and hardware constraints may apply, and some CPU-only features may not be available on the GPU version.