Experimental confirmation of compute-bound scaling on non-H100 accelerators

Establish experimentally whether inference for the WAN2.1-T2V text-to-video diffusion model exhibits a compute-bound regime and the same quadratic (in spatial resolution and frame count) and linear (in denoising steps) scaling laws on hardware accelerators other than the NVIDIA H100 SXM, for realistic token lengths.

Background

The paper validates a compute-bound analytical model and associated scaling laws for WAN2.1-T2V on an NVIDIA H100 SXM, showing quadratic scaling with spatial and temporal dimensions and linear scaling with denoising steps.

Although Appendix results suggest these behaviors should generalize across accelerators, the authors explicitly acknowledge that such generalization has not been experimentally confirmed, identifying a gap that requires multi-accelerator measurement to verify the model’s portability.