KIOXIA DEMONSTRATES BILLION-SCALE AI VECTOR SEARCH

KUALA LUMPUR, March 17 (Bernama) -- Kioxia Corporation, a leader in memory solutions, has demonstrated high-dimensional vector search scaling to 4.8 billion vectors on a single server using its open-source KIOXIA AiSAQ approximate nearest neighbour search (ANNS) technology.

The company in a statement said the achievement marks a significant advancement for retrieval augmented generation (RAG) applications, which rely on fast and efficient access to large-scale vector databases.

Kioxia Corporation Managing Executive Officer, Vice President, SSD Division, Masashi Yokotsuka said scaling vector databases into billions requires rethinking both memory and compute.

 He added that combining SSD-based vector search with GPU-accelerated indexing enables practical deployment at scale, and Kioxia, as industry innovators, will continue to push the boundaries of artificial intelligence (AI) using flash memory.

In collaboration with NVIDIA, Kioxia also demonstrated substantial reductions in index build time by leveraging graphics processing unit (GPU) acceleration via NVIDIA cuVS. The improvement reached up to 20 times faster index construction for high-dimensional vectors and up to 7.8 times faster end-to-end build times.

This translates to a reduction in index build time from 28.4 days using central processing units (CPUs) to 1.4 days using four NVIDIA Hopper GPUs, while overall processing time fell from 31 days to four days.

Its AiSAQ technology enables scalable storage architectures by allowing vector search directly from solid-state drives (SSDs), reducing reliance on dynamic random-access memory (DRAM). The system is capable of meeting latency requirements for RAG applications using a single query server within a Milvus vector database environment.

​​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.

-- BERNAMA