GPU DiskANN and Beyond: Accelerating Microsoft Vector Search with NVIDIA cuVS

From web search to ads recommendation to RAG, vector search has become an essential operation for vast range of large-scale applications. However, these applications may have widely varying requirements. As such, no single design or algorithm fits all use cases. Nevertheless, GPUs can accelerate many scenarios, potentially giving orders of magnitude improvements in index construction time or search throughput, compared with CPU-based solutions. In this talk joint talk with Microsoft and NVIDIA, we will present a range of Microsoft use cases that require high performance vector search. We will then discuss the unique challenges of these scenarios and how we leverage NVIDIA GPUs and the cuVS library to accelerate these workloads. From fast DiskANN index construction to high-throghput filtered search, a range of algorithms, features, and optimizations are used to leverage GPUs and provide performance and cost benefits over current CPU-based solutions.