KIOXIA Corporation has added a function to its SSD-based generative AI vector search software, KIOXIA AiSAQ, that allows users to adjust the balance between the conflicting elements of response performance and answer accuracy in a RAG (Retrieval Augmented Generation) system. System designers can now flexibly adjust the balance between KIOXIA AiSAQ’s search performance and maximum number of vectors within the installed SSD capacity without changing the hardware configuration.

KIOXIA AiSAQ (All-in-Storage ANNS with Product Quantization) software, announced in January 2025, uses a new SSD-optimized ANNS (Approximate Nearest Neighbor Search) algorithm and places indexed data on SSD instead of DRAM, allowing vector databases to scale without being limited by DRAM capacity.

AI Authority Trend: Kioxia Expands 8th Gen BiCS FLASH SSDs to Boost GPU Use in AI and HPC Workloads

If the capacity of the SSD installed in the RAG system is fixed, in order to improve search performance (number of responses per time), it is necessary to increase the SSD consumption capacity per vector, which results in a smaller maximum number of vectors that can be held, leading to a decrease in the accuracy of the RAG system’s answers. On the other hand, in order to prioritize the accuracy of the RAG system’s answers and increase the number of vectors, it is necessary to reduce the SSD consumption capacity per vector, which leads to a decrease in search performance. With the addition of this new function, it is now possible to appropriately balance these two conflicting conditions according to the workload of each RAG system. In addition, the ability to optimize the balance between search performance and the number of vectors expands the use of KIOXIA AiSAQ not only for RAG applications but also for other vector-intensive applications such as offline semantic search.

AI Authority Trend: Kioxia, AIO Core and Kyocera Announce Development of PCIe 5.0-Compatible Broadband Optical SSD

As the demand for large-scale AI services grows, SSDs, as a practical alternative to DRAM, provide high throughput and low latency for RAG systems. KIOXIA AiSAQ efficiently meets these demands and realizes large-scale generative AI without being limited by DRAM capacity.

By releasing KIOXIA AiSAQ as open source, we will promote the development of AI systems with SSD-centric architecture and contribute to the AI ​​community.

AI Authority Trend: KIOXIA AiSAQ Tech Reduces DRAM Needs in GenAI Systems, Released as Open Source

Source – businesswire

To share your insights, please write to us at sudipto@intentamplify.com