Image and video generation

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: —.

Speed

Images per minute

32 images, one request at a time, after two discarded warm-up images. Higher is better.

Seconds per image

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.

Every number

The images

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).

Video: Wan2.2

Minutes per clip

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.

Every number

The clips

The same prompt, seed and settings on each system (the stacks seed their noise differently, so the clips differ).

Reading these results

  • The QuietBox 2 ran Tenstorrent's media server (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.
  • Both serve one image at a time (batch size 1), so there is no concurrent-user test: a second request waits for the first.
  • Most of the gap is generation itself: about 3.2 s on the QuietBox 2 against 5.1 s on the Spark. The Spark also spends about 0.75 s per image after generation, mostly encoding and sending a PNG about twice the size the QuietBox 2 sends; the QuietBox 2's overhead is a few hundredths of a second.
  • Requests went through kubectl port-forward from a workstation on the same network for both machines, which adds a little to every request on both.
  • Not measured yet for images: startup time (the QuietBox 2's first boot took about 8 minutes to download the weights, compile and warm up), wall power, larger images, and two Sparks.
  • Video. Wan2.2-T2V-A14B (two 14B experts, Apache-2.0). The QuietBox 2 ran Tenstorrent's build for p300x2 (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.
  • The 2x DGX Spark rows run one vLLM-Omni server across both Sparks: a Ray cluster over the QSFP cable, with CFG parallelism putting the positive and negative halves of each guided step on separate GPUs (NCCL over the RoCE rails). That split moves little data per step, and it ran 1.99× as fast as one Spark.
  • Each Spark drew about 200 W at the wall with its GPU at 96% utilization and 80–86 °C for the whole run. One Spark powered itself off 15 minutes into its first clip; after a power cycle it ran the full test, and the pair after it, without a repeat.
  • The QuietBox 2's first Wan2.2 boot took about 30 minutes (a 126 GB download plus compiling); clips are shown re-encoded smaller than the servers produced them.

Method. tools/image_bench.py sends 16 fixed prompts twice (seeds 42 and 43) to each server's /v1/images/generations, one request at a time, after two warm-up requests that are not counted, and times each from request to decoded response. Images per minute is 32 divided by the wall time of the timed requests.

tools/video_bench.py sends four fixed prompts (seed 42) one at a time after one uncounted warm-up clip, and times each from submitting the job to downloading the finished MP4.

Data and code: images.csv, videos.csv, and every image and clip in results/*/flux.1-schnell/ and results/*/wan2.2-t2v-a14b/ at github.com/rothgar/personal-ai-hardware. Built 2026-11-21.