TOKYO -- Kioxia Corporation announced the successful demonstration of achieving high-dimensional vector search scaling to 4.8 billion vectors on a single server with its open-source KIOXIA AiSAQ™ approximate nearest neighbor search (ANNS) technology. Additionally, Kioxia demonstrated a significant reduction in index build time by leveraging GPU acceleration through NVIDIA cuVS. These two achievements mark a significant advancement for retrieval augmented generation (RAG) search solutions. Continued development is underway to support larger-scale deployments beyond 4.8 billion vectors.

Index build time on a massive-scale vector database is a crucial pain point for the industry. In collaboration with NVIDIA, Kioxia demonstrated up to 20x improvement in KIOXIA AiSAQ index build time for high-dimensional vectors of 1024 dimensions, and up to 7.8x improvement in end-to-end build times. This 20x improvement represents a reduction from 28.4 days using CPU to 1.4 days using four NVIDIA Hopper GPUs to build the index, and a reduction from 31 days to 4 days in end-to-end testing.[1]

AI applications may now rely on larger volumes of vectorized information reaching tens of billions of vectors and beyond stored on SSDs, while DRAM alone becomes impractical even at a billion scale. Kioxia enables a highly scalable storage architecture with KIOXIA AiSAQ technology by achieving billion-scale search, exceeding RAG application latency requirements using a single query server in a Milvus vectorDB environment powered by GPU acceleration on index builds that make large scale deployments practical.

“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.”

First announced last year, KIOXIA AiSAQ open-source software technology addresses RAG scalability challenges by enabling vector search directly from SSDs, with reduced DRAM usage. KIOXIA AiSAQ technology provides high scalability, making it well-suited for both multi-tenant environments and large-scale monolithic index deployments. The technology leverages an innovative Global Index algorithm that combines hybrid clustering and graph search to deliver efficient vector search at extreme scale. With flexible tuning options to balance performance and high-volume vector scalability, KIOXIA AiSAQ software makes large-scale deployments more accessible and easier to expand.

“Scaling vector databases into the billions requires rethinking both memory and compute,” said Masashi Yokotsuka, Managing Executive Officer, Vice President, SSD Division, Kioxia Corporation. “By combining KIOXIA AiSAQ SSD-based vector search with NVIDIA GPU acceleration for index construction, we provide practical index build at high scale deployments. As industry innovators, we will continue to push the boundaries of AI using flash memory.”

​​Kioxia remains committed to advancing storage-driven AI solutions that support intelligent data processing at scale and continues to evolve KIOXIA AiSAQ toward trillion-vector deployments.