Video Generation

Video is an even more demanding workload than image generation, not only for number crunching but also for memory, which makes it a natural fit for a 128 GB box. We ran Wan 2.2, a current open text-to-video model, on both machines.


Just like with image generation, the two GPUs couldn't handle the task. Ryzen AI Halo completed the test in a bit under 10 minutes; the DGX Spark was much faster here.


The video output looks pretty decent.

MiniMax H3

For most video models, the Halo is a slower but fully capable machine. The problem is at the bleeding edge, and the limit there is support rather than performance. The newest text-to-video model we tried, MiniMax H3, generates synchronized audio alongside the video, and it ran on the Spark but would not run on the Halo at all. It needs FP8, which RDNA 3.5 does not have, and it relies on NVIDIA-specific kernels.


Here is MiniMax-H3's take on the same fox prompt the Wan clips used—much better, not far from photorealistic.