KIOXIA Releases AiSAQ Technology as Open-Source Software to Reduce DRAM Requirements for Generative AI Systems

Kioxia Corporation, a world-leading memory solutions provider, today announced the open-source release of its new All-Storage ANNS Product Quantization (AiSAQ) technology(1). The KIOXIA AiSAQ™ software introduces a novel Approximate Nearest Neighbor Search (ANNS) algorithm optimized for solid-state drives (SSDs), delivering scalable performance for Retrieval-Augmented Generation (RAG) while eliminating the need to store index data in DRAM—instead, searches are performed directly on SSDs.

Generative AI systems require significant computational, memory, and storage resources. While AI has the potential to drive transformative breakthroughs across industries, its deployment often comes with high costs. RAG is a crucial stage in AI development, refining large language models (LLMs) by leveraging application- or enterprise-specific data.

At the core of RAG is a vector database, which accumulates and converts domain-specific data into feature vectors. RAG also relies on ANNS algorithms to identify vectors that enhance the model by evaluating the similarity between accumulated and target vectors. To be effective, RAG must retrieve the most relevant information quickly. Traditionally, ANNS algorithms have been deployed in DRAM to achieve the required high-speed performance.

The KIOXIA AiSAQ technology delivers a scalable and efficient ANNS solution capable of handling billion-scale datasets with minimal memory usage and fast index switching capabilities.