AuraTracer智迹闻
中文

EVENT DOSSIER

“A shining star of domestic products! Results of the global AI storage benchmark test MLPerf Storage 3.0 are released, with nine first-place finishes.”

2026-09-02 08:00 Products & Apps 🔥 28.9 heat score
1sources
1days unfolding
28.9heat score
5mentions
SummaryAI generated

On September 1, 2026, MLCommons announced the results of the MLPerf Storage 3.0 benchmark test. This test covered three different I/O load scenarios: Unet3D, RetinaNet, and Llama3 Checkpoint, with 19 manufacturers and 23 products participating. In the parallel file storage category, MeshFS from XSKY星辰天合 performed outstandingly in key metrics such as microsecond-level I/O latency and TB/s-level aggregate bandwidth, thanks to its proprietary Shared-Everything all-flash architecture and utilization of NVMe and RDMA hardware capabilities. It ultimately won nine first-place awards.

Related eventsRELATED EVENTS
Key entitiesKEY ENTITIES
IntelMLCommonsMeshFSStars and Sky AllianceXSKY

Coverage · reports per dayLANGUAGE SPLIT

Entity relations
Intel × MLCommons1Intel × MeshFS1Intel × Stars and Sky A…1Intel × XSKY1MLCommons × MeshFS1MLCommons × Stars and S…1

SignalsSIGNALS

Keyword heat
  • XSKY1
  • Stars and Sky Alliance1
  • MLCommons1
  • MeshFS1
  • Intel1

All reports (1)SOURCES

甲子光年 zh 2026-09-02 08:00

“A shining star of domestic products! Results of the global AI storage benchmark test MLPerf Storage 3.0 are released, with nine first-place awards.”

On September 1, MLCommons released the results of the MLPerf Storage 3.0 benchmark test. MeshFS under XSKY achieved first place in nine categories of the parallel file storage track. The test was organized by MLCommons and covered three different I/O patterns: Unet3D, RetinaNet, and Llama3 Checkpoint. A total of 19 manufacturers and 23 products participated in the evaluation. MeshFS uses a self-developed Shared-Everything all-flash architecture, leveraging NVMe and RDMA hardware capabilities, and performed excellently in terms of microsecond-level I/O latency and TB/s-level aggregate bandwidth. The test used a three-node EC hardware configuration (dual-processor Intel Xeon 6530P, 512GB memory, 48 SSDs). MeshFS performed well in terms of small-file training throughput, while Unet3D and RetinaNet achieved results with 405B Checkpoint…