Kioxia Company, a worldwide chief in reminiscence options, has introduced an replace to its KIOXIA AiSAQ (All-in-Storage ANNS with Product Quantization) software program, which is a part of an ongoing effort to optimize the usage of solid-state drives (SSDs) and make AI vector database searches inside retrieval-augmented technology (RAG) methods extra usable.
With the assistance of the brand new open-source launch’s configurable controls, system architects can now decide the best steadiness between search pace and vector depend, two parameters which are mutually unique given the system’s set SSD storage capability. The following benefit permits RAG system architects to exactly steadiness explicit workloads and their wants with out requiring any adjustments to the {hardware}.
The brand new approximate closest neighbor search (ANNS) algorithm utilized by KIOXIA AiSAQ software program, which was first launched in January 2025, is optimized for SSDs and doesn’t require index knowledge to be saved in DRAM. The constraints imposed by restricted DRAM capability are considerably eradicated by KIOXIA AiSAQ expertise, which permits vector searches immediately on SSDs and lowers host reminiscence necessities.
Growing search efficiency (queries per second) necessitates utilizing extra SSD capability per vector whereas the system’s put in SSD capability is mounted. Because of this, there are fewer vectors. However, lowering SSD capability utilization per vector is important to maximise the variety of vectors, which ends up in decreased efficiency. Relying on the workload, there are a lot of methods to finest steadiness these two conflicting circumstances.
The KIOXIA AiSAQ software program provides intensive setup decisions to assist discover the proper steadiness. With this most up-to-date improve, directors can select the most effective steadiness for a spread of disparate workloads throughout the RAG system. With this model, KIOXIA AiSAQ expertise is now an SSD-based ANNS which may be used for offline semantic searches and different vector-hungry purposes along with RAG purposes.
As the necessity for scalable AI providers grows, SSDs present a workable substitute for DRAM in dealing with the excessive throughput and low latency wanted by RAG methods. These expectations might be successfully met by KIOXIA AiSAQ software program, which permits for large-scale generative AI with out being restricted by reminiscence sources.
Kioxia demonstrates its dedication to the AI neighborhood by selling SSD-centric designs for scalable AI by means of the open-source launch of KIOXIA AiSAQ software program.
