Open diffusion models on the QuietBox 2 and the DGX Spark, all on Talos Linux and Kubernetes, each with the server its vendor's stack provides. FLUX.1-schnell images, one at a time at 1024×1024 and 4 steps: —. Wan2.2 video, five-second clips at 832×480: —.
← Best Personal AI Hardware Talos Linux vs Ubuntu Image data (CSV) Video data (CSV) Code and data on GitHub
32 images, one request at a time, after two discarded warm-up images. Higher is better.
Every timed image (a dot each) from request to the finished image back at the client; the bar is the median. The ring is the generation time the server itself reported. Lower is better.
The first pass over the prompt set, same prompt and seed on both. The two stacks seed their noise differently, so the same seed does not give the same picture; shown at 512×512 (generated at 1024×1024).
Wan2.2-T2V-A14B, 81 frames (about five seconds at 16 fps) at 832×480, 40 steps with guidance, one clip at a time: a dot per clip, the bar is the mean. Lower is better.
The same prompt, seed and settings on each system (the stacks seed their noise differently, so the clips differ).
tt-media-inference-server 0.18.0, the
FLUX.1-schnell build marked complete for p300x2 in tt-inference-server's catalog) across all four Blackhole
chips, with the settings from Tenstorrent's Helm chart, firmware 19.15.0. The DGX Spark ran
vLLM-Omni 0.32.0 (the vLLM project's server for
diffusion models; plain vLLM serves text only) on its GB10 with NVIDIA driver 595, in BF16.kubectl port-forward from a workstation on the same network for both
machines, which adds a little to every request on both.tt-media-inference-server 0.17.0, marked complete in the catalog), which serves a fixed
832×480, 81-frame shape with guidance 4.0 then 3.0; the Sparks were sent the same shape, guidance, step count
(40), seed and Wan's standard negative prompt, so every system ran the same amount of guided denoising.