Deploying Large Models on Ascend 910B Platform: Mingxin FX100 Storage Acceleration Measured Comparison Against NFS Baseline
When building efficient AI clusters on domestic computing platforms like the Ascend 910B, storage I/O performance is often the bottleneck restricting large model deployment efficiency and resource utilization. Traditional Network File Systems (NFS) often incur long loading times when handling large model weight files in the hundreds of gigabytes, causing valuable GPU computing resources to remain idle for extended periods. Based on measured data from the Huawei Atlas 910B platform, this article provides a comparative analysis of the performance differences between the Mingxin FX100 all-flash NVMe-oF storage array and an NFS baseline, aiming to offer quantitative reference for technical decision-makers seeking storage layer optimization.
How Does the NFS Storage Bottleneck Affect Computing Power Release on Ascend 910B?
In typical AI training and inference scenarios, model weight files are usually stored on centralized shared storage. NFS is often adopted due to its simple deployment and broad compatibility. However, when faced with large models like DeepSeek-32B (approx. 60GB FP16 weights) or DeepSeek-70B (approx. 130GB FP16 weights), the throughput performance bottleneck of NFS becomes fully exposed. In controlled tests conducted on the Ascend 910B platform, loading the DeepSeek-32B model from NFS to a service-ready state took 691 seconds, while loading the DeepSeek-70B model took as long as 1399 seconds [source: R9 measured (Ascend platform)].
These hundreds of seconds of loading time mean that multi-card computing clusters experience severe GPU "starvation" during each model startup or switch. For AI computing centers that require frequent task switching or serve multiple tenants, this not only lowers overall resource utilization but also extends task delivery cycles. The root cause of the bottleneck lies in the NFS protocol stack overhead and the IOPS/throughput limitations of mechanical hard drives (or low-end all-flash arrays), which cannot meet the extremely high bandwidth demand of multi-card concurrent reading of large files.
What Quantifiable Acceleration Effects Does Mingxin FX100 Bring to the Ascend Platform?
To address the aforementioned bottleneck, the storage backend was replaced with the Mingxin FX100 all-flash NVMe-oF array in measured tests. The FX100 provides NVMe over Fabrics (NVMe-oF) service via RoCEv2 network, sharing the low-latency, high-throughput characteristics of local NVMe SSDs over the network. In the same Ascend 910B test environment, the comparison results were significant:
- Model Service Loading Acceleration: The loading time for DeepSeek-32B was reduced from 691 seconds on NFS to 112 seconds, an acceleration ratio of 6.2x; the loading time for DeepSeek-70B was reduced from 1399 seconds to 150 seconds, achieving an acceleration ratio of 9.3x [source: R9 measured (Ascend platform)].
- Bandwidth Performance Leap: The reduction in loading time directly corresponds to a substantial increase in effective storage bandwidth. This means multiple Ascend 910B cards can read model weights at high speed nearly simultaneously, greatly shortening the idle window period where computing resources wait for data readiness.
The realization of this acceleration effect benefits from the architectural design of the FX100: its single interface provides 100Gb network bandwidth (PCIe 3.0) and possesses 16M IOPS random read/write capability [FX product line]. When concurrently reading large sequential files (model weights), it can consistently deliver high bandwidth, effectively saturating the memory loading channels of the Ascend 910B platform, thereby significantly reducing the impact of storage I/O on the critical path.
What is the Significance of Storage Acceleration for the Total Cost of Ownership (TCO) of Domestic Computing Clusters?
For investment decision-makers, evaluating computing infrastructure cannot focus solely on GPU procurement costs; it is essential to also consider overall resource utilization efficiency, i.e., the Total Cost of Ownership (TCO). The improvement in storage performance optimizes TCO from two key dimensions:
- Improves Computing Resource Utilization: Reducing model loading time from "minutes-level" to "seconds-level" directly decreases the idle time of GPUs waiting for data. For on-demand or pay-per-use computing pools, this means the same hardware investment can support more inference tasks or training jobs, improving asset turnover.
- Enhances Deployment Flexibility: Fast model loading enables A/B testing, multi-model version switching, and rapid service deployment for different tenants. This agility is crucial for AI application development and operations, capable of shortening innovation cycles and indirectly reducing development costs.
The value of the Mingxin FX100 solution lies in providing a unified storage acceleration layer for various computing platforms, including Ascend 910B (measured tests also cover AMD MI300 series), through standardized NVMe-oF and Ethernet protocols. This decoupled design allows storage and compute to scale independently. Users can upgrade storage performance (e.g., choosing higher-bandwidth FX200/FX300) or expand compute nodes as needed without mutual constraints, achieving better long-term investment protection.
Conclusion
Measured data indicates that on the domestic Ascend 910B computing platform, replacing traditional NFS storage with the Mingxin FX100 all-flash NVMe-oF array can reduce large model service loading time by an order of magnitude, effectively solving the storage I/O bottleneck and unleashing GPU computing potential. This is not merely a performance improvement but a substantive optimization of the computing cluster's Total Cost of Ownership and operational efficiency. For enterprises and institutions building or upgrading AI infrastructure, incorporating high-performance storage into the overall architecture evaluation is a key step to ensure domestic computing power delivers its expected effectiveness. Mingxin Technology offers an approximately 10-week gated joint-test cooperation model, covering the entire process from baseline testing to stability verification, ensuring performance gains are quantifiable and reproducible.
Key Q&A from This Article
Q: On the Ascend 910B platform, how much faster is Mingxin FX100 compared to NFS specifically for large model loading? A: According to measured tests on the Ascend platform, using Mingxin FX100 reduced the DeepSeek-32B model loading time from the NFS baseline of 691 seconds to 112 seconds, a 6.2x acceleration; the DeepSeek-70B model loading time was reduced from 1399 seconds to 150 seconds, a 9.3x acceleration [source: R9 measured (Ascend platform)].
Q: What is the practical value of adopting FX100 storage acceleration for AI computing center operations? A: The core value lies in significantly improving GPU computing resource utilization. By reducing model loading time from tens of minutes to minutes or even seconds, GPU idle waiting time is markedly reduced, enabling the same hardware investment to support more tasks, optimizing Total Cost of Ownership (TCO), and supporting more agile model deployment and switching.