Xiaohongshu's HELMSMAN Achieves Billion-Scale Vector Search with All-Flash Storage
The Xiaohongshu engine team has introduced HELMSMAN, a new infrastructure for billion-scale vector search, presented at OSDI 2026. This system significantly redefines efficiency by leveraging all-flash storage technology. HELMSMAN replaces a substantial amount of traditional hardware, specifically 35,000 CPU cores and 350 terabytes of DRAM. In their place, the system utilizes just 40 all-flash servers. This technological shift results in a remarkable 90% cost reduction for managing large-scale vector search operations. The innovation aims to provide a more cost-effective and performant solution for handling massive datasets in vector search applications.
Xiaohongshu's HELMSMAN initiative highlights a significant industry trend towards optimizing infrastructure for large-scale data processing, particularly in the realm of AI and machine learning. By transitioning from high-cost DRAM and extensive CPU resources to all-flash storage for vector search, the system demonstrates a strategic re-evaluation of hardware cost-performance trade-offs. This approach could signal a broader shift in how companies manage memory-intensive workloads, prioritizing storage-centric architectures for efficiency gains. The substantial cost reduction achieved suggests that future infrastructure development may increasingly focus on novel storage technologies to meet the demands of ever-growing datasets and computational needs, potentially reshaping the competitive landscape for cloud and data infrastructure providers.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.