Model Loading and Training Storage Acceleration for Ascend 910B Clusters
The common bottleneck in Ascend 910B clusters is NFS storage: model loads take tens of minutes. Mingxin FX measured DeepSeek-32B/70B serving loads 6.2×/9.3× faster, with all three training workloads improved (R9).
The storage status quo on 910B
Ascend clusters often arrive as full racks with storage defaulting to general-purpose NFS — training data, weights and checkpoints all share one NAS path. R9's measurements show that path is a clear bottleneck under AI loads: DeepSeek-32B serving load takes 691 s and 70B takes 1399 s, GPUs idling throughout.
The R9 three-group comparison
Platform: Huawei Atlas 910B ×8 (Kunpeng-920). Inference loading group: DeepSeek-32B 691 s → 112 s (6.2×), DeepSeek-70B 1399 s → 150 s (9.3×). Training weight/checkpoint group: significant acceleration on Qwen-7B I/O. Training data group: YOLOv8/COCO reads accelerated. All against an NFS baseline; report available on request.
Labeled honestly: R9's platform is Ascend 910B; the 480B KV tiering figures (+29–40%) come from the MI308X platform (R2/R3) — the two sets are never mixed. Ascend-platform KV tiering measurements are on the schedule.
FAQ
How do we choose against Huawei's own OceanStor A series?
All-Huawei stacks may evaluate the first-party solution first; customers with mixed compute or reproducible joint-test requirements are welcome to test FX side by side (see the comparison page, where both sides' strengths are stated honestly).
Is the CANN/MindSpore stack supported?
The R9 measurements were completed on the Ascend software stack (inference loading and training workloads); vLLM-Ascend KV tiering adaptation is on the schedule and will be backfilled from measurements.
What if we have domestic-content requirements?
The FX series is a domestic vendor's product and supports domestic-substitution deliveries; a component-level domestic-content list is provided at the business stage.
Data sources (verifiable)
Related reading
- Storage Acceleration for AMD MI308X Inference Clusters
- Model-Load Acceleration: A Measured Path from 23 Minutes to 2.5 Minutes
- Checkpoint and Data-Read Acceleration for Training Clusters
Joint test first, decisions second: gate-based acceptance with built-in stop-loss
The full costing model is provided as reproducible Python after NDA — customers can rerun it with their own parameters. Every key figure on this site carries a report ID and is open to third-party verification.
This site presents business-cooperation information and constitutes neither an investment offer nor any promise of returns. Measured data come from signed / official test reports (see the Evidence Library); vendor specs, public sources and estimates are labeled as such.