KIOXIA achieves 4.8 billion vector search on a single AI server with minimal DRAM

KIOXIA's open-source AiSAQ (All-in-Storage ANNS with Product Quantization) technology is something we've covered in the past, and it has been a game-changer for AI workloads with Retrieval-Augmented Generation (RAG) pipelines by offloading vectorized data from costly DRAM to a more efficient, cost-effective SSD solution.

2

This opens the door to scaling, especially for running more complex AI workloads. And when it comes to large-scale workloads and systems, at NVIDIA GTC 2026, KIOXIA demonstrated that, using the NVIDIA cuVS Library with KIOXIA AiSAQ Technology, it successfully scaled high-dimensional vector search to an impressive 4.8 billion vectors on a single server.

With the NVIDIA cuVS Library offering a "significant reduction in index build time by leveraging GPU acceleration," this breakthrough is a big step forward for retrieval augmented generation (RAG) search solutions. And with that, KIOXIA is already looking toward supporting larger-scale event deployments that go beyond 4.8 billion vectors.

Index build time is a notable pain point for a massive-scale vector database, and with its partnership with NVIDIA, KIOXIA has demonstrated a massive 20X improvement in AiSAQ index build time (for high-dimensional vectors of 1024 dimensions) and up to a 7.8X improvement in end-to-end build times. To put that into perspective, it can mean the difference between a CPU taking upwards of a month to build the index and a handful of NVIDIA Hopper GPUs, using the NVIDIA cuVS Library and KIOXIA AiSAQ Technology, completing the same task in a couple of days.

"Vector databases provide a backbone for applications that need to understand intent, context, and similarity across massive, unstructured datasets in real time," said Jason Hardy, Vice President, Storage Technologies, NVIDIA. "By leveraging GPU-accelerated indexing with the NVIDIA cuVS library, Kioxia supports high-dimensional vector databases that can scale and build indexes with unprecedented efficiency."

For an in-depth look at this milestone and the testing, check out KIOXIA's blog post here.