Research result · Sep 1, 2026
Yanrong Technology's F9000X Tops MLPerf Storage v3.0 and Launches YRCache ContextBox for Shared KV Cache
On the MLPerf Storage v3.0 list released by MLCommons on September 1, Yanrong Technology's all-flash AI storage system F9000X posted first-place results in a three-node cluster: 544 GiB/s aggregate bandwidth for 3D U-Net training, and 539 GiB/s read / 314 GiB/s write for Checkpoint 70B, all new public records.
The company also introduced YRCache, an inference storage system, and the YRCache ContextBox shared KV cache appliance. It adds a G3.5 shared cache layer; a single unit recorded 120 GB/s, supports eight concurrent inference nodes, and scales to petabyte-level cache. In tests at a leading AI enterprise, first-token latency fell 89–94% and throughput rose 3–6×, the company says.
Original sources (Chinese)
刷新全球权威基准测试纪录！这家国产厂商在AI存储布局有多深？