Indonesian Political, Business & Finance News

KIOXIA Achieves 4.8 Billion High-Dimensional Vector Search Database on Single Server, with 7.8x Index Creation Time Acceleration via GPU

| Source: ANTARA_ID Translated from Indonesian | Technology
KIOXIA Achieves 4.8 Billion High-Dimensional Vector Search Database on Single Server, with 7.8x Index Creation Time Acceleration via GPU
Image: ANTARA_ID

Tokyo—Kioxia Corporation has announced a significant achievement in high-dimensional vector search technology by successfully demonstrating the ability to handle 4.8 billion vectors on a single server using its AiSAQ™ open-source Approximate Nearest Neighbor Search (ANNS) technology. In collaboration with NVIDIA, Kioxia has also demonstrated substantial improvements in index creation time through GPU acceleration via NVIDIA’s cuVS library. These developments represent major advances for Retrieval Augmented Generation (RAG) search solutions, with continuous development underway to support even larger-scale implementations exceeding 4.8 billion vectors.

Index creation time in large-scale vector databases has been a primary bottleneck for the industry. Through its partnership with NVIDIA, Kioxia demonstrated a 20-fold speed increase in AiSAQ index creation for high-dimensional vectors with 1024 dimensions, and an overall 7.8-fold speedup. The 20-fold acceleration is based on reducing index creation time from 28.4 days using CPU processing to just 1.4 days using four NVIDIA Hopper GPUs, whilst overall testing time decreased from 31 days to 4 days.

Modern AI applications can now reliably handle larger volumes of vectorised information, reaching tens of billions of vectors and beyond stored on SSDs, whereas DRAM alone becomes impractical at the billion-scale level. Kioxia’s AiSAQ technology enables highly scalable storage architecture for billion-scale searches, exceeding the latency requirements of RAG applications using a single query server in a Milvus vectorDB environment supported by GPU acceleration during index creation, making large-scale deployment practical.

“Vector databases have become essential for applications that need to understand intent, context and similarity across vast collections of unstructured data instantly,” said Jason Hardy, Vice President of Storage Technologies at NVIDIA. “By leveraging GPU-accelerated indexing with NVIDIA’s cuVS library, Kioxia supports high-dimensional, scalable vector databases and delivers indexing efficiency previously unseen.”

First announced last year, KIOXIA’s AiSAQ open-source software technology addresses RAG scalability challenges by enabling vector search directly from SSD, with significantly lower DRAM usage. AiSAQ technology offers exceptional scalability, making it well-suited for multi-tenant environments and large-scale monolithic index implementations. The technology leverages an innovative Global Index algorithm that combines hybrid clustering and graph search to deliver efficient vector search at extreme scales. With flexible tuning options to balance performance and scalability for high-volume vectors, KIOXIA’s AiSAQ software makes large-scale implementations more accessible and expandable.

“Scaling vector databases to billions requires rethinking both memory and computing,” said Masashi Yokotsuka, Managing Executive Officer and Vice President of the SSD Division at Kioxia Corporation. “Through the combination of KIOXIA’s SSD-based vector search AiSAQ and NVIDIA’s GPU acceleration for index creation, we provide practical index creation for large-scale implementations. As an industry innovator, we will continue pushing the boundaries of AI using flash memory.”

Kioxia remains committed to advancing AI-based storage solutions that support intelligent data processing at massive scale and continues developing AiSAQ towards trillion-scale vector implementations.

The open-source AiSAQ KIOXIA software can be downloaded at: https://github.com/kioxia-jp/aisaq-diskann.

About Kioxia

Kioxia is a global leader in memory solutions, dedicated to the development, production and sale of flash memory and solid-state drives (SSDs). In April 2017, its predecessor, Toshiba Memory, was spun off from Toshiba Corporation, the company that invented NAND flash memory in 1987. Kioxia is committed to elevating the world with “memory” by offering a diverse range of products, services and systems that create choices for customers and memory-based value for society. KIOXIA’s innovative 3D flash memory technology, BiCS FLASH™, brings the future of storage to high-density applications including advanced smartphones, PCs, automotive systems, data centres and generative AI systems.

View JSON | Print