Kioxia Adds Flexible Capacity-Performance Tuning To Open-Source Software

KUALA LUMPUR, July 3 (Bernama) -- Kioxia Corporation, a world leader in memory solutions, has announced an update to its KIOXIA All-in-Storage ANNS with Product Quantisation (AiSAQ) software.

The update is an ongoing effort to improve the usability of artificial intelligence (AI) vector database searches within retrieval-augmented generation (RAG) systems by optimising the use of solid-state drives (SSDs).

This new open-source release introduces flexible controls allowing system architects to define the balance point between search performance and the number of vectors, which are opposing factors in the fixed capacity of SSD storage in the system.

Kioxia in a statement said the resulting benefit enables architects of RAG systems to fine tune the optimal balance of specific workloads and their requirements without any hardware modifications.

First introduced in January this year, KIOXIA AiSAQ software uses a novel approximate nearest neighbour search (ANNS) algorithm that is optimised for SSDs and eliminates the need to store index data in dynamic random-access memory (DRAM).

By enabling vector searches directly on SSDs and reducing host memory requirements, KIOXIA AiSAQ technology allows vector databases to scale, largely without the restrictions caused by limited DRAM capacity.

This update makes KIOXIA AiSAQ technology a suitable SSD-based ANNS for not only RAG applications but also other vector-hungry applications such as offline semantic searches.

With growing demand for scalable AI services, SSDs offer a practical alternative to DRAM for managing the high throughput and low latency that RAG systems require.

KIOXIA AiSAQ software makes it possible to meet these demands efficiently, enabling large-scale generative AI without being constrained by limited memory resources.

With the release of KIOXIA AiSAQ software as open-source, Kioxia reinforces its commitment to the AI community with the promotion of SSD-centric architectures for scalable AI.

-- BERNAMA