Ascend 910B Storage Acceleration: Mingxin FX100 vs NFS
On domestic computing platforms for model inference and training, storage I/O bottlenecks often become performance weak points
On domestic computing platforms for model inference and training, storage I/O bottlenecks often become performance weak points. Measured results of the Mingxin FX100 all-flash NVMe-oF array on the Huawei Ascend 910B platform show that, compared to a traditional NFS baseline, model inference service loading time can be reduced by a factor of 6.2 to 9.3, significantly improving cluster deployment efficiency. This result is based on the measured report R9, providing a quantifiable reference for optimizing storage architecture in computing centers.
Measured Background: Why Focus on Storage Acceleration for the Ascend 910B Platform?
The Huawei Ascend 910B, as a representative domestic GPU product, is being gradually deployed at scale in AI training and inference scenarios. However, its storage subsystem typically relies on NFS (Network File System) for shared storage, which can easily become a performance bottleneck in multi-node clusters. Operations such as model loading and checkpoint saving are sensitive to storage latency and bandwidth, and the performance degradation of NFS under high concurrency can slow down the overall workflow.
The Mingxin FX100 uses the NVMe-oF (NVMe over Fabrics) protocol, providing low-latency, high-bandwidth storage access over a RoCEv2 network. This test was conducted on the Ascend 910B platform, using NFS as the baseline to compare the FX100's performance in model loading scenarios, aiming to verify the acceleration effect of domestic storage solutions on domestic computing platforms.
Test Configuration and Baseline Setup
The test platform was a Huawei Atlas 910B server equipped with 8 Ascend 910B accelerator cards. For storage:
- Baseline: Standard NFS shared storage, typical data center deployment configuration.
- Tested: Mingxin FX100 all-flash NVMe-oF array, 4-disk RAID0 configuration (14 TB, XFS), single-port 100 GbE, RoCEv2 network.
The test models were DeepSeek-32B (approximately 65.6 GB weights) and DeepSeek-70B (approximately 140 GB weights), simulating loading scenarios for medium-scale and large-scale models, respectively. Loading time was measured from the initiation of read requests until the model was fully ready for computation, excluding non-storage factors such as network handshakes.
Measured Results: Model Loading Acceleration of 6.2-9.3x
According to the measured report R9, the FX100's model loading performance on the Ascend 910B platform significantly outperformed the NFS baseline:
- DeepSeek-32B: NFS baseline loading time was 691 seconds, FX100 was 112 seconds, acceleration ratio 6.2x.
- DeepSeek-70B: NFS baseline loading time was 1399 seconds, FX100 was 150 seconds, acceleration ratio 9.3x.
The acceleration ratio increased with model size. This is because the FX100's NVMe-oF architecture can sustain stable throughput under high bandwidth demands, while NFS, when transferring large files, suffers more from protocol overhead and network latency, leading to more pronounced performance degradation. This difference was amplified with the 70B model, indicating that the FX100 is better suited for the storage acceleration needs of large-scale domestic computing clusters.
Technical Attribution of Performance Improvement
The FX100's acceleration effect stems from differences in its underlying technical architecture:
- Protocol Efficiency: NVMe-oF directly maps NVMe commands to the network, reducing the file system protocol stack overhead of traditional NFS, lowering CPU usage and latency.
- Parallel I/O Capability: The FX100 supports multi-channel concurrency; in measured results, sustained read bandwidth can reach several times that of NFS, avoiding single-point bottlenecks.
- RoCEv2 Network: Based on RDMA transport mechanisms, it reduces network interrupts and copy operations, improving throughput efficiency.
On the Ascend 910B platform, these features work in coordination with Huawei's CANN (heterogeneous computing architecture), ensuring that storage acceleration does not become a bottleneck for operator execution. For training scenarios, faster model loading means reduced cluster idle waiting time and improved overall utilization.
Implications for Domestic Computing Cluster Deployment
The domestic computing ecosystem (e.g., Ascend, ROCm platforms) is maturing rapidly, but the alignment of storage infrastructure is often underestimated. NFS, as a traditional solution, may be acceptable in small to medium-scale deployments, but when facing frequent model switching and checkpoint saving operations in thousand-card clusters, its performance shortcomings can be amplified. The 6-9x loading acceleration provided by the FX100 directly translates to:
- Reduced service startup time: During model updates or failure recovery, the cluster can return to working state faster.
- Support for more frequent model iteration: During training, quickly loading new model versions reduces experiment cycle times.
- Lower operational complexity: Centralized management of NVMe-oF is easier to maintain compared to distributed NFS.
For technical decision-makers in computing centers, storage solutions should be chosen based on the I/O characteristics of actual workloads. If operations like model loading and checkpoint saving constitute a significant portion, dedicated storage accelerators like the FX100 can offer quantifiable return on investment.
Conclusion
The measured results of the Mingxin FX100 on the Ascend 910B platform demonstrate the acceleration potential of domestic storage solutions for domestic computing platforms. The 6.2-9.3x improvement in model loading provides a concrete path for optimizing deployment efficiency in computing centers. To evaluate the FX100's performance in your own environment, please contact the Mingxin technical team to arrange a gated joint test (approximately 10-week process, including verification of key metrics such as TTFT reduction ≥25%), ensuring results are reproducible and measurable.
Key Points Q&A
Q: How much acceleration does the Mingxin FX100 provide on the Ascend 910B platform compared to NFS? A: Measured results show model loading time is reduced by a factor of 6.2x (DeepSeek-32B) to 9.3x (DeepSeek-70B), with the acceleration ratio increasing with model size (measured, report R9).
Q: Why is the FX100's acceleration effect superior to NFS? A: The FX100 uses the NVMe-oF protocol and RoCEv2 network, reducing file system protocol overhead compared to NFS, supporting higher parallel I/O, and providing significantly better sustained bandwidth than traditional NFS.
Q: What is the practical significance of this acceleration for domestic computing clusters? A: It can reduce service startup time, support more frequent model iteration, and lower operational complexity, making it especially suitable for frequent model switching and checkpoint saving scenarios in thousand-card clusters.