Huawei가 AI 추론 워크로드 가속화를 위해 설계된 OceanStor M900 컨텍스트 메모리 스토리지 시스템을 하이퍼스케일 데이터 센터용으로 출시한다.

Huawei has introduced the OceanStor M900 Context Memory Storage, a system designed to accelerate AI inference in hyperscale data centers by breaking through on-chip memory and DRAM capacity bottlenecks. The product delivers PB-scale KV cache for SuperPoDs and provides a fully shared memory space with PB-scale capacity and TB/s-level performance.

The announcement was made by David Wang, deputy chairman and rotating chairman of Huawei, at Huawei Connect 2026 during his keynote "Advancing the Agentic World, Building a Solid Silicon Foundation." The introduction reflects a broader industry shift in AI infrastructure from a compute-centric model toward deeper collaboration among compute, network, and storage systems.

As large models scale to 10 trillion parameters, SuperPoDs have become the optimal infrastructure choice for AI deployment. Mainstream models now support context windows exceeding one million tokens, with multi-turn inference and complex tasks becoming standard. The resulting KV cache data generation has pushed on-chip memory and DRAM beyond capacity and cost-effectiveness limits, creating consensus around building multi-tier storage systems that coordinate on-chip memory, DRAM, and SSDs into a fully shared memory space.

OceanStor M900 employs three key technologies to address these challenges. The high-speed UnifiedBus interconnect network enables PB-scale global multi-tier KV cache pooling with one-hop connections, expanding capacity from on-chip memory and DRAM to SSDs. A single cluster delivers 64PB of capacity, raising available KV cache per NPU from gigabytes to terabytes and significantly boosting KV cache hit ratios.

The system uses an integrated architecture combining CPU, network controller unit, and NAND controller unit to provide native KV semantics and enable one-hop connections between SuperPoD NPUs and SSDs. This eliminates protocol conversion and CPU forwarding, reducing access latency from milliseconds to 60 microseconds—a 90% reduction. A single cluster delivers 40TB/s of aggregate access bandwidth, 1.5 times higher than competing solutions. In typical AI scenarios, this architecture doubles inference cluster token throughput and halves time to first token.

OceanStor M900 incorporates industry-first KV-aware adaptive storage technology that predicts KV cache lifecycles based on data value and intelligently distributes data across storage media. The technology enables up to 24 drive writes per day (DWPD), extending SSD endurance by 16 times and ensuring three-year stability. By reducing media replacement and operations costs, it lowers long-term AI inference infrastructure expenses and accelerates adoption.