KIOXIA updates its groundbreaking AiSAQ software for SSD-based scalable AI systems

KIOXIA AiSAQ (All-in-Storage ANNS with Product Quantization) is a groundbreaking software solution for large AI RAG systems that leverages the flexible nature of high-speed SSD storage, rather than expensive DRAM. It's an open-source solution that has just been updated with new, flexible controls.

2

With large AI models covering a wide range of datasets, efficiently indexing data is extremely important. However, as speed is a critical component, these are usually stored in memory, which is extremely expensive. Solutions that leverage disk-based storage exist, but what makes AiSAQ so impressive is that its 'all-in-storage' approximate nearest neighbor search (ANNS) algorithm is fast, efficient, and scalable.

This week's KIOXIA AiSAQ update is all about enhancing the control users and organizations have over scaling by offering system architects the ability to "define the balance point between search performance and the number of vectors."

In a system with a fixed SSD capacity, increasing search performance (defined as queries per second) requires more SSD capacity per vector. This means fewer vectors. On the other hand, increasing the number of vectors results in reduced SSD capacity and lower performance. The latest update allows users to fine-tune this balance and find the optimal solution with new flexible configuration tools and options.

"With the latest version of KIOXIA AiSAQ software, we're giving developers and system architects the tools to fine-tune both performance and capacity," said Neville Ichhaporia, senior vice president and general manager of the SSD business unit at KIOXIA America, Inc. "This level of flexibility is critical to building scalable, RAG systems - powered by SSD storage. By open-sourcing our technology, we're reinforcing our commitment to the AI community with solutions that are both powerful and accessible to everyone."

The latest version of KIOXIA AiSAQ open-source software can be downloaded here.