The benchmark was conducted using the open-source Step-Video-T2V 30B model to evaluate SP (Sequence Parallelism) and TP (Tensor Parallelism) performance. We applied ulysses_degree as sp_degree.
Implementation reference:
https://github.com/stepfun-ai/Step-Video-T2V/tree/main#multi-gpu-parallel-deployment
| GPUs | Parallel Type | Configuration | Latency | Speedup Ratio | Memory Usage |
|---|---|---|---|---|---|
| 1 | Baseline | TP1 SP1 |
213.60s | 1.00x | 92,170M |
| 2 | TP | TP2 |
108.97s | 0.98x | 57,458M ▼37.7% |
| 2 | SP | SP2 |
108.13s | 0.99x | 86,258M ▼6.4% |
| 4 | TP | TP4 |
57.61s | 0.93x | 36,566M ▼60.3% |
| 4 | SP | SP4 |
57.01s | 0.94x | 78,226M ▼15.1% |
| 8 | TP | TP8 |
30.40s | 0.88x | 30,028M ▼67.4% |
| 8 | SP | SP8 |
30.10s | 0.89x | 79,684M ▼13.5% |
-
Hardware Compatibility:
- Consumer GPUs (5090/5090D): Full training support on 32GB×8 configuration
- Inference Accelerators (L20/L40): Full parameter inference on 48GB×4 configuration
-
Efficiency Metrics:
- TP8 achieves 67.4% memory optimization (53.9% higher than SP8)
- Mixed-parallel latency trend remains within <12% deviation from theoretical expectation
-
Scalability:
- Multi-dimensional parameter slicing enables near-linear scaling efficiency
- Layered communication optimization reduces cross-node synchronization overhead by 75%