Kioxia AiSAQ Software-Optimized RDMA SSD Solution

The Kioxia AiSAQ (AI Storage Acceleration with Queue) solution is a software-optimized RDMA SSD architecture purpose-built for large-scale AI inference, Retrieval-Augmented Generation (RAG), and vector database workloads. At its core, AiSAQ replaces the conventional requirement of holding entire vector indices in DRAM by enabling direct RDMA-based access to NVMe SSDs, allowing the Approximate Nearest Neighbor (ANN) search engine to operate efficiently against indices resident on flash storage. This architectural shift reduces per-node DRAM consumption by up to 90% for billion-scale vector datasets while sustaining query-per-second (QPS) throughput competitive with fully in-memory deployments, fundamentally altering the economics of AI infrastructure at scale.

AiSAQ integrates at the software layer, providing compatibility with leading open-source vector database frameworks including DiskANN and HNSW-derived engines, and works in conjunction with Kioxia's enterprise-class PCIe Gen 4 and Gen 5 NVMe SSDs. The solution leverages RDMA over Converged Ethernet (RoCE) and InfiniBand transport fabrics to enable disaggregated storage access with microsecond-class latency, allowing compute nodes to query remote SSD pools without burdening host CPUs with data movement. The AiSAQ software stack includes a custom I/O scheduler optimized for the random small-block read patterns characteristic of ANN graph traversal, plus prefetch heuristics that pipeline SSD accesses ahead of algorithmic demand.

Launched in Q1 2025, AiSAQ is particularly notable for enabling organizations to scale vector knowledge bases into the tens of billions of embedding vectors without proportionally scaling DRAM capacity, directly addressing a critical infrastructure bottleneck in enterprise LLM deployments. By decoupling memory capacity requirements from compute node specifications, AiSAQ allows AI infrastructure architects to right-size CPU and GPU clusters while leveraging high-density SSD storage arrays for the bulk of the vector index. This approach is validated for production RAG pipelines, semantic search platforms, recommendation engines, and multi-modal AI systems where DRAM-resident indices become cost-prohibitive beyond the hundred-million-vector threshold.

Ideal for

• Enterprise RAG pipelines: Hosting billion-scale document embedding indices on NVMe SSDs to augment LLM inference without expanding DRAM per inference node
• Semantic and vector search at scale: Powering e-commerce, media, and knowledge management platforms requiring sub-100ms ANN query latency over datasets exceeding 10 billion vectors
• AI recommendation systems: Offloading user and item embedding graphs to SSD-backed storage tiers, enabling recommendation engines to scale catalog coverage without DRAM exhaustion
• Multi-model LLM serving infrastructure: Freeing DRAM capacity on GPU inference servers by relocating retrieval indices to networked NVMe pools, maximizing memory available for model weights and KV cache
• Disaggregated AI storage clusters: Deploying RDMA-connected SSD enclosures as shared vector index repositories accessible by multiple compute nodes simultaneously over RoCE or InfiniBand fabrics
• Healthcare and scientific data retrieval: Enabling similarity search across genomic, imaging, or sensor embedding databases at petabyte scale within latency budgets suitable for real-time clinical decision support

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.