DGX Spark video generation: Wan 2.2, LTX-2.5 and MiniMax H3 times, and how an RTX 5090 compares

On a DGX Spark, a 5-second 768x512 LTX-2.5 clip with audio takes about 70 seconds end to end in one owner's NVFP4 setup, and a 5-second 1280x704 Wan 2.2 TI2V-5B clip at 30 steps takes 14.1 minutes warm in ComfyUI (reports from June to September 2026). On identical workloads an RTX 5090 finishes 2.4 to 5.5 times sooner, because video diffusion is limited by compute, not by memory bandwidth. The Spark's real advantage is 128 GB of memory: pipelines that peak above 100 GB run on it, and not on a 32 GB card. GPUwerk has not measured video generation itself; every figure below is someone else's, with a link.

Seconds per finished clip. The ai-muninn Spark runs were on an ASUS Ascent GX10 (same GB10 chip) with the GPU clock capped at 2,200 MHz, which the author puts at roughly 20% per run; an uncapped Spark would land somewhat closer to the 5090. The Sol page gives durations, not frame counts. Ratios are my arithmetic.

How long does a video clip take on a DGX Spark?

Between about one minute and a quarter of an hour for a 5-second clip, depending mostly on the model family. Distilled LTX models run about 8 steps per stage and finish in tens of seconds at 768x512. Wan 2.2 at 30 steps takes many minutes. These are the Spark-only reports I could verify; none comes with an RTX 5090 run of the same settings.

Read 7 November 2026. The Wan times are ComfyUI execution times from its /history endpoint; weight loading added about 87 s on top, and six other containers were resident on that machine. The FastVideo pull request was still open, not merged, when I read it. Earlier in the same Wan thread, a native Diffusers pipeline took "30+ minutes" for a 5-second clip, which other posters attributed to the pipeline, not the hardware.

The only Wan 2.2 figure on an RTX 5090 that I found is from RunPod's October 2026 GPU comparison: 1,297 s for the A14B model at 1280x704, 81 frames and 20 steps in fp16. RunPod notes that its 32 GB and 24 GB cards stream weights from system memory on that model, so the time reflects offloading as well as GPU speed. It is a different model and step count from the Spark's TI2V-5B run, so I do not put the two side by side.

Why compute, not bandwidth, sets the pace

A diffusion step runs the whole transformer over thousands of latent tokens at once, so the arithmetic dwarfs the time spent reading weights. LLM decode is the opposite: one token per step, every active weight read once, little math per byte. That is why the Spark's 273 GB/s matters for chat and barely matters here.

A worked check with the FastVideo numbers. The base stage of LTX2.3 at 1088x1920 took 61.97 s for 5 steps, so 12.4 s per step. The model is about 22B parameters (ai-muninn describes LTX-2 as a ~22B DiT), which is about 44 GB in BF16. Reading 44 GB at 273 GB/s takes 0.16 s. By my calculation, weight traffic is about 1.3% of a step; the other 98.7% is compute.

The cross-machine numbers agree. The RTX 5090 has 1,792 GB/s, 6.6 times the Spark's bandwidth (see the Spark spec sheet and the Spark vs RTX 5090 comparison). If bandwidth set the speed, every row in the first table would show a gap near 6.6x. The measured gaps are 2.4x to 5.5x, and the cleanest same-software comparison, NVIDIA's own Sol engine, shows 2.6x and 3.4x. For the Spark's compute, the most direct number I found is a 15-second fp16 matrix burn that Petronella ran on eight GB10 units: 88.6 to 91.4 TFLOPS before a September 2026 firmware update and 77.6 to 81.3 after it. The article does not say whether that test was dense or sparse.

What this means for NVFP4

On an LLM, 4-bit weights speed up decode because fewer bytes cross the bus. On a video model they help only if the matmul itself runs in FP4. ai-muninn found weight-only NVFP4 cut an LTX-based model from 29.2 GB to 19.5 GB on a Spark and made it 6% slower (18.0 to 19.1 s), because the weights are unpacked to BF16 before each matmul. The FastVideo pull request runs real FP4 matmuls through flashinfer on GB10 and cuts total denoise time from 171.3 s to 130.8 s (23.7%) at 1080p; end to end the saving is 16.9%, because the VAE decode stays in BF16. The LTX-2.5 setup repo adds one more trap: upstream NVFP4 kernels are built for sm_100a, sm_110a and sm_120a, GB10 is sm_121, and without a patch the quantize kernel falls back to software emulation about 4x slower. Format background is in the NVFP4 vs MXFP4 vs FP8 explainer.

What fits in 128 GB that does not fit in 32 GB?

Multi-model pipelines and the 14B-class Wan models without offloading. Single distilled LTX models fit both machines.

• MiniMax H3 plus LTX-2.5 refine: 33.5 GiB and 20.1 GiB of weights, 104 GiB measured peak on a 121 GiB GB10 (ai-muninn). The author notes the 5090's 31.84 GiB cannot hold the H3 stage alone, so (my inference) its times in the first table include spilling weights to system RAM.
• NVIDIA's ComfyUI video workflow: peak GPU memory of about 80 GB, 100 GB and 120 GB for its three model tiers, and 70 GB to 230 GB of disk (NVIDIA playbook).
• Wan 2.2 A14B: a mixture of two 14B experts, 27B total and 14B active per step. The Wan 2.2 model card says the single-GPU command needs 24 GB with offloading and that 80 GB or more lets you drop the offload flags.
• LTX-2.5 NVFP4 distilled: 22.5 to 26 GiB peak on the Spark per the setup repo, about 29 GiB system-wide. This one fits a 5090.

Video weights and outputs are large, and a rented Spark has 1 TB of NVMe with no backups kept, so read the storage guide before a long batch and copy finished clips off as you go.

Which open video models can you run in the EU?

Wan 2.2 and LTX-2.5, yes. HunyuanVideo 1.5 and MiniMax H3, not under their published licenses. This matters for anyone renting in Europe: GPUwerk's Sparks sit in Hradec Králové, Czech Republic. I read the license texts on 7 November 2026; this is not legal advice.

The H3 license also requires prior authorization above $20M yearly revenue and a visible "MiniMax H3" label on commercial products. On Hugging Face, the Wan-AI organization lists 27 repositories, all Wan 2.1 or 2.2 (newest: Wan2.2-Animate-2-14B, July and August 2026); there are no Wan 2.5, 2.6, 2.7 or 3.0 weights. The newest Lightricks base model is LTX-2.5 (23 July 2026). The broader license picture is in the open-weight license guide.

ComfyUI on ARM64: what breaks and the fixes people report

ComfyUI runs on the Spark through NVIDIA's own playbook, which the ComfyUI image-generation guide covers along with tunnelling to port 8188. Video adds its own problems, all reported by owners:

• Older NGC PyTorch images have no sm_121 kernels, and a plain pip install in a venv replaces the tuned PyTorch build; Brian Baldock used --system-site-packages to keep it.
• SageAttention has no aarch64 build, so ComfyUI falls back to SDPA, which Baldock puts at about 20% slower. FlashAttention is also unavailable on sm_121 according to the FastVideo pull request.
• torch.compile failed in the sampler with a Triton error about missing CUDA headers in Baldock's setup; he disconnected the compile node.
• Versions matter. ai-muninn got silently broken output from ComfyUI 0.30.0 and correct output from 0.34.0 with comfy-kitchen 0.2.31 and comfy-aimdo 0.4.15.
• nvidia-smi reports memory as N/A on unified memory; use free -h or ComfyUI's /system_stats instead.
• An untiled 1080p VAE decode runs out of memory on GB10 (FastVideo), so keep VAE tiling on even with 128 GB.

DGX Spark or RTX 5090 for video?

For clips from models that fit in 32 GB, the RTX 5090. For pipelines that do not fit, the Spark.

The cost per clip makes the same point. At $0.79 per hour, the 573.8-second 4K LTX-2.5 clip costs about $0.13 of Spark time (573.8 / 3,600 x 0.79). At the RTX 5090 rental rate of $0.89 per hour, the same clip at 170.3 s costs about $0.04. A warm 848-second Wan 2.2 TI2V-5B clip costs about $0.19 on the Spark. By that arithmetic the 5090 is roughly three times cheaper per LTX clip, before model download and warm-up time, which you pay on either machine.

My recommendation: use the 5090 for LTX-2.5 and Wan 2.2 TI2V-5B work, where both fit, and rent a Spark when you need Wan 2.2 A14B without offloading, several models resident at once, or a video model next to an LLM. On the Spark, start with LTX-2.5 in NVFP4 using a build that runs FP4 matmuls, because the step-count advantage of distilled LTX models is larger than any hardware difference here. Do not plan an EU project around HunyuanVideo 1.5 or MiniMax H3 without legal advice, whatever the speed. For still images, the open models, their licenses and VRAM are compared in local image generation models.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• Image to Video Generation using the Spark vs. PC, NVIDIA Developer Forums, 2026-09-24
• [Video] MiniMax H3 Super Acceleration on a DGX Spark: 2x native cold, 3.14x warm, ai-muninn, 2026-09-14
• Accelerating LTX-2.5 with the Sol Video Inference Engine, NVIDIA Labs, 2026-08-13
• [perf] LTX2 FP4-linear on DGX Spark (GB10/sm_121): ~24% faster denoise at 1080p, hao-ai-lab/FastVideo on GitHub, 2026-07-13
• Best GPU for AI video: What the benchmarks show in 2026, RunPod, 2026-10-12
• [Benchmark] NVFP4 shrinks a video model 33% on a DGX Spark, with zero speed gain, ai-muninn, 2026-06-01
• LTX-2.5_Video_DGX_Spark_Setup, GitHub (Hitheshkaranth)
• A $4k box, an open-source video model, and one stubborn easter egg, Brian Baldock, 2026-06-29
• Generate Images and Videos with ComfyUI, NVIDIA build.nvidia.com, 2026-10-13
• Wan-AI/Wan2.2-TI2V-5B model card, Hugging Face
• Lightricks/LTX-2.5 model card, Hugging Face
• MiniMax H3 Community License Agreement, Hugging Face (MiniMaxAI)
• Tencent Hunyuan Community License (HunyuanVideo-1.5), Hugging Face (tencent)
• DGX Spark firmware update: what we learned on 8 GB10 units, Petronella Technology Group, 2026-09-26