Validation Whitepaper (Web edition)
Beijing Information Science and Technology University · Huawei Ascend Atlas 910B · NFS baseline · 7 reproducible metrics.
Abstract
Beijing Information Science and Technology University (independent third party) benchmarked ZK-Storage WS5000 on Huawei Ascend Atlas 910B against an NFS baseline, covering inference load/service, training I/O and token throughput — 7 key metrics in total. WS5000 led across the board with a ~90.9% median reduction, reproducible and verifiable (S38). This is the web edition of the whitepaper for search and citation; the full version is linked as a PDF below.
Why are these results credible?
Because they come from an independent third party under a stated platform and baseline, and are reproducible. The device under test is ZK-Storage WS5000 (disaggregated all-flash, 300 GB/s aggregate bandwidth, ~20 μs latency; vendor spec S9), with NFS network storage as the baseline; test and data conventions are itemized below.
Method
The same dataset and workload were used, switching only the storage link (NFS baseline vs ZK-Storage NVMe-oF over RoCE) while holding other conditions constant, to isolate storage's impact on end-to-end performance; representative values are reported.
Test setup (reproducible)
Independent third party, stated platform, stated baseline.
| Item | Detail |
|---|---|
| Tester | Beijing Information Science and Technology University (independent third party) |
| Platform | Huawei Ascend Atlas 910B |
| Baseline | NFS network storage (NFS over TCP, 10GbE, ~1.25 GB/s) |
| ZK-Storage link | NVMe-oF over RDMA / RoCE (2x200GbE, ~50 GB/s line rate) |
| Metrics | Inference load/service, training I/O, token efficiency — 7 in total |
Inference: load and service speedup
| Model | ZK-Storage load | NFS load | Load speedup | Latency cut | Service speedup |
|---|---|---|---|---|---|
| DeepSeek-32B | 6.62 s | 563.85 s | 85.17× | 98.83% | 6.17× |
| DeepSeek-70B | 35.38 s | 1284.66 s | 36.31× | 97.25% | 9.33× |
Training: weights and checkpoint I/O
| Test | ZK-Storage | NFS baseline | Speedup | Latency cut |
|---|---|---|---|---|
| 模型加载 | 12.72 s | 140.23 s | 11.02× | 90.93% |
| 模型保存 | 31.16 s | 165.87 s | 5.32× | 81.21% |
| Checkpoint 加载 | 10.55 s | 131.37 s | 12.45× | 91.97% |
| Checkpoint 保存 | 81.94 s | 451.14 s | 5.51× | 81.84% |
Token throughput (= effective GPU utilization)
| Switch frequency | ZK-Storage util. | NFS util. | Relative gain |
|---|---|---|---|
| 10/day | 99.8% | 80.4% | +24.1% |
| 20/day | 99.5% | 60.8% | +63.6% |
| 40/day | 99.1% | 21.7% | +356.9% |
Conclusion
In Beijing Information Science and Technology University's independent test, ZK-Storage WS5000 reached ~85× peak model-load speedup, 5–12× training I/O speedup and up to +357% token efficiency; median reduction across 7 metrics was 90.9% — reproducible and verifiable.S38
Reproducibility
Figures on this page are rendered by Python from the site's single source (business_plan/outputs/results.json), shared with the Validation page; any update refreshes both, preventing drift.
Last updated：