Meta and NVIDIA have been collaborating for three years to push vector search performance further on GPUs. The latest result is in Faiss v1.10.0, which now ships GPU-accelerated indexing algorithms from NVIDIA cuVS inside Meta's open source similarity search library. The new implementations target faster index build times and lower query latency for both inverted file (IVF) and graph-based indexes.

Faiss, originally developed by Meta FAIR, has long supported vector search on GPUs, including the GpuIndexFlat, GpuIndexIVFFlat, and GpuIndexIVFPQ indexes introduced around 2016. The newer collaboration brought GpuIndexCagra, a graph index designed for GPUs, first contributed by NVIDIA and now officially part of Faiss 1.10.0. The release also includes a new conda package that lets users choose between the classic Faiss GPU implementations and the newer cuVS algorithms, while retaining the ability to switch between GPU and CPU execution.

Benchmark setup and results

Benchmarks were run with the cuVS-bench tool on an NVIDIA H100 GPU and compared against an Intel Xeon Platinum 8480CL CPU system. Two datasets were exercised: a tall, slender set of 100 million 96-dimensional vectors from the Deep1B image dataset, and a shorter, wider set of 5 million text embeddings generated with the OpenAI text-embedding-ada-002 model. Results are reported at 95% recall for k=10 nearest neighbors, along the pareto frontiers of each index.

Index

Embeddings
100M x 96
(seconds)

Embeddings
5M x 1536
(seconds)
Faiss Classic Faiss cuVS Faiss Classic   Faiss cuVS Faiss Classic Faiss cuVS
IVF Flat IVF Flat 101.4 37.9 (2.7x) 24.4 15.2 (1.6x)
IVF PQ IVF PQ 168.2 72.7 (2.3x) 42.0 9.0 (4.7x)
HNSW (CPU) CAGRA 3322.1 518.5 (6.4x) 1106.1 89.7 (12.3x)

Table 1: Index build times for Faiss-classic and Faiss-cuVS in seconds (with NVIDIA cuVS speedups in parentheses).

For IVF indexing, cuVS delivers up to 4.7x faster build times than classic GPU-accelerated Faiss. Graph-based indexing sees an even larger improvement: the CAGRA index builds up to 12.3x faster than CPU-based Hierarchical Navigable Small World (HNSW) graphs.

Index

Embeddings
100M x 96
(milliseconds)

Embeddings
5M x 1536
(milliseconds)
Faiss Classic Faiss cuVS Faiss Classic Faiss cuVS Faiss Classic Faiss cuVS
IVF Flat IVF Flat 0.75 0.39 (1.9x) 1.98 1.14 (1.7x)
IVF PQ IVF PQ 0.49 0.17 (2.9x) 1.78 0.22 (8.1x)
HNSW (CPU) CAGRA 0.56 0.23 (2.4x) 0.71 0.15 (4.7x)

Table 2: Online (i.e., one at a time) search query latency for Faiss-classic and Faiss-cuVS in milliseconds (with NVIDIA cuVS speedups in parentheses).

Search latency also improves meaningfully. IVF search with cuVS runs up to 8.1x faster than the classical GPU implementation, while CAGRA search is up to 4.7x faster than CPU HNSW. These gains come at the same recall target, putting cuVS ahead on both build and query phases of the index lifecycle for these workloads.

What the integration means going forward

The incorporation of cuVS algorithms into Faiss reflects a broader shift toward GPU-native vector index designs. Graph-based indexes like CAGRA, built specifically for GPU memory hierarchies rather than adapted from CPU layouts, are now a first-class option within Faiss. As newer GPU architectures continue to evolve, the Meta-NVIDIA partnership intends to keep folding state-of-the-art algorithms into the library. More details on cuVS are available from NVIDIA's cuVS page.