Independent validation
Beijing Information Science and Technology University · Huawei Ascend Atlas 910B · leading on all 7 metrics.
SETUP
A reproducible test setup
Objective and checkable: an independent third party, a stated platform, a 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-DPU link | NVMe-oF over RDMA / RoCE (2×200GbE, ~50 GB/s line rate) |
| Metrics | Inference load/service, training I/O, token efficiency — 7 in total |
INFERENCE
Inference: load and service speedup
Bring-up and switching go from minutes to seconds.
| Model | ZK-DPU 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
Training: weights and checkpoint I/O
The bigger the model and the more frequent the checkpoints, the more idle time you save.
| Test | ZK-DPU | 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% |
THROUGHPUT
Token throughput (= effective GPU utilization)
The more frequent the switching, the wider the gap.
| Switch frequency | ZK-DPU 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-DPU WS5000 reached ~85× peak model-load speedup, 5–12× training I/O speedup and up to +357% token efficiency; median latency reduction across 7 metrics was 90.9% — reproducible and verifiable.S38
Benchmark it on your own workload
2 live demo units are ready for immediate PoC. Let the data do the talking.