Wan 2.2 video server: 4× NVIDIA RTX PRO 6000 Blackwell Server Edition or 4× H200 NVL

Two four-card servers for one video model

For Wan 2.2 at 720p, a server with four NVIDIA RTX PRO 6000 Blackwell Server Edition cards runs four clips at the same time, one per card, and a server with four H200 NVL cards does the same with more compute per card or joins all four over NVLink to finish one clip sooner. By our estimate from NVIDIA's peak BF16 rates, the upper bound on clips per hour is about 1.7 times higher per card on the H200 NVL.

Wan's repository states that its single-GPU 720p command for Wan2.2-T2V-A14B "can run on a GPU with at least 80GB VRAM", with offloading options switched on, so both cards hold one job each: the RTX PRO 6000 has 96 GB, the H200 NVL 141 GB.

This article compares the two four-card servers for a studio, agency or training team that produces many clips. The model list, licences of other video models and a configuration table for evaluation up to a central platform are in our guide to video generation GPU requirements; the general comparison of the two cards for LLMs is in RTX PRO 6000 vs H200 NVL.

NVIDIA product pages and product briefs, read 24 October 2026. BF16 peaks as in our GPU TFLOPS comparison: NVIDIA's RTX PRO Blackwell whitepaper gives 503.8 dense TFLOPS for the Workstation Edition and no dense figure for the Server Edition, which has the same GPU at a lower FP32 rating (120 against 125 TFLOPS), so the Workstation figure bounds it; the H200 NVL's 1,671 TFLOPS with sparsity, halved. FP8 rates are StorageReview's MAMF measurements (sustained dense matrix multiplication) of 14 October 2026, the RTX PRO 6000 figure taken on the Workstation Edition, which StorageReview says profiles identically to the Server Edition; the H200 NVL ran only at 450 W, so StorageReview calls its figure a floor for the 600 W card.

What one Wan 2.2 job needs

Wan 2.2 was released with weights and code on 28 July 2025. The A14B text-to-video and image-to-video models are mixtures of experts: "Each expert model has about 14B parameters, resulting in a total of 27B parameters but only 14B active parameters per step", with a high-noise expert for the early steps and a low-noise expert for the later ones. Wan's code switches from one to the other at a fixed point in the noise schedule.

The repository stores each expert as six safetensors files totalling 57.2 GB, which is 4 bytes per parameter for about 14B parameters, so the files hold 32-bit weights by our reading. The option --convert_model_dtype "converts model parameter types to config.param_dtype", and Wan's shared configuration sets that to BF16. The whole repository is 126 GB on disk, including the 11.4 GB UMT5-XXL text encoder in BF16 and a 508 MB VAE.

Hugging Face file list of Wan-AI/Wan2.2-T2V-A14B and Wan's configuration files on GitHub, read 24 October 2026; BF16 sizes our arithmetic at 2 bytes per parameter.

With --offload_model True, Wan's sampling loop moves the inactive expert to system memory and the T5 encoder to the processor after encoding the prompt, which is how the job fits Wan's 80 GB figure. Wan publishes no figure without offloading. By our arithmetic, all weights in BF16 take 64.4 GiB, which leaves about 31 GiB of the 95.6 GiB the RTX PRO 6000 driver reports and about 76 GiB on the H200 NVL. On the H200 NVL both experts and the encoder can stay on the card; on the RTX PRO 6000, test at your resolution whether the job runs without offloading before you plan on it.

Offloading moves the weights into the host. Each offloaded job keeps about 69 GB of BF16 weights in system memory by our arithmetic, so four jobs need about 276 GB. That sits within the 768 GB we quote for four RTX PRO 6000 cards and the 1,152 GB for four H200 NVL, figures that follow NVIDIA's guide for certified systems of at least twice the GPU memory. Our guide to how much RAM a GPU server needs explains the rule. We size host memory and NVMe together with the cards; tell us how many jobs should run at once, and we propose the memory for that load.

How much compute one clip takes

Wan's defaults for the A14B model are 81 frames at 16 frames per second, about 5 seconds, and 40 sampling steps. The VAE compresses each frame by 8 in both directions and the frame count by 4, and the transformer cuts the latent into patches of 2 × 2, as its configuration states (vae_stride (4, 8, 8), patch_size (1, 2, 2)). At 1280 × 720 that gives 21 latent frames of 90 × 160, or 3,600 tokens per latent frame and 75,600 tokens per clip, by our arithmetic. Every step runs the transformer twice, once with the prompt and once with the negative prompt, because guidance is on by default.

Each video token passes through about 12.0B of the 14B weights. Wan's model code gives every one of the 40 blocks four self-attention projections, a cross-attention whose query and output projections act on the video tokens and whose key and value projections act on the 512 text tokens, and two feed-forward layers: 6 × 5,120² + 2 × 5,120 × 13,824 weights per block on the video path. By the standard count a pass then costs about 2 × 12.0B × 75,600 operations in the linear layers and 4 × 75,600² × 5,120 × 40 in self-attention, 1.8 × 10^15 plus 4.7 × 10^15, so attention takes nearly three quarters of the work at 720p. Eighty passes per clip come to about 5.2 × 10^17 operations. The cross-attention to the text, the text encoder and the VAE are left out, which keeps the figure a floor. At 832 × 480 the sequence has 32,760 tokens and the clip costs about 1.3 × 10^17, a quarter of the 720p figure, because attention grows with the square of the sequence.

Dividing by a card's peak rate gives a floor for the time per clip. Wan runs its transformer in BF16, so we use NVIDIA's dense BF16 Tensor Core peaks, 503.8 TFLOPS for the RTX PRO 6000 and 835.5 TFLOPS for the H200 NVL. Real kernels stay below the peak: StorageReview's sustained FP8 matrix rate on the RTX PRO 6000 is 774 TFLOPS, 77 per cent of NVIDIA's dense FP8 figure of 1,007.6, and softmax, normalisation, the text encoder and the VAE decoder add time that the count leaves out. The figures below are therefore an upper bound on clips per hour by our estimate, useful for comparing configurations and not for promising a delivery rate.

Our estimate of the upper bound in clips per hour, rounded down, four-card rows with one job per card: 5.19 × 10^17 operations per 720p clip and 1.33 × 10^17 per 480p clip, from Wan's default settings, configuration files and model code, divided by NVIDIA's dense BF16 Tensor Core peaks of 503.8 TFLOPS (RTX PRO 6000) and 835.5 TFLOPS (H200 NVL); measured times will be longer.

We found no published minutes per clip for Wan 2.2 on either card from Wan, NVIDIA or an independent lab, and Wan's own efficiency table is published only as an image of results on unnamed GPUs. MLPerf® Inference v6.1 added a text-to-video benchmark on Wan2.2-T2V-A14B with 81 frames at 720 × 1280 and 20 steps, as our video guide describes, and NVIDIA's inference performance page (metadata 19 October 2026) lists results for it only on GB300 NVL72, GB200 NVL72 and DGX B300 systems. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.

One clip per card or one clip on four cards

A studio queue has two numbers: clips per hour and the wait for one clip. One job per card maximises the first, because the cards exchange no data and every card keeps its own copy of the model. Four RTX PRO 6000 cards then work like four separate machines in one chassis, and a failed card takes out a quarter of the capacity.

Sequence parallelism splits one clip across cards. Wan's repository runs its multi-GPU examples with "PyTorch FSDP and DeepSpeed Ulysses", --dit_fsdp --t5_fsdp --ulysses_size 8 on eight GPUs, and its efficiency notes test 4 and 8. Ulysses gives every card a quarter of the 75,600 tokens and, in each of the 40 attention layers, exchanges queries, keys, values and outputs between all cards. By our estimate each card sends about 23 GB per transformer pass in BF16, or about 1.9 TB over the 80 passes of a clip.

Our estimate of lower bounds for one 720p clip: compute floor from the table above divided by four; the wait adds both floors and assumes the exchanges do not overlap with computation; traffic 4 exchanges × 18,900 tokens × 5,120 × 2 bytes × 3/4 per layer, 40 layers, 80 passes; PCIe 5.0 at 64 GB/s and NVLink at 450 GB/s in each direction from NVIDIA's H200 page (128 GB/s and 900 GB/s for both directions together).

--dit_fsdp adds traffic of its own: FSDP keeps a quarter of each expert on every card and gathers the rest for every pass, about 21 GB per card per pass by our arithmetic. By our arithmetic the full BF16 model (64.4 GiB) fits each card on both servers, so each card can hold its own copy without FSDP; test both settings. Over PCIe the exchanges also cross the link between the two processors when the cards sit under different sockets, and peer-to-peer transfers depend on the IOMMU and ACS settings our article on RTX PRO 6000 P2P, IOMMU and ACS covers. Wan publishes no scaling figures for PCIe-only cards, so measure the four-card time on the server before you size for it.

The H200 NVL's 4-way bridge joins all four cards in one NVLink domain at 900 GB/s per GPU. NVIDIA's product brief asks for bridged cards in one CPU domain, and our article on the H200 NVL bridge, 2-way or 4-way shows how the slot layout affects it. For a team where a creative director waits for each draft, that shorter wait is the reason to choose the bridged server. For a pipeline that renders hundreds of variants overnight, one job per card gives more clips for the same power, on either card.

Which four-card server for which studio

Our estimates from the tables above; card features from NVIDIA's product pages and our sibling configuration pages.

The RTX PRO 6000 server is the more flexible box. Each card splits into up to four MIG instances of 24 GB for image models or a small LLM, and the Server Edition is the RTX PRO 6000 edition NVIDIA supports with vGPU, so the same server can carry video jobs by day and other work at night. Its Blackwell cores also compute FP4, which the H200 NVL lacks; we found no published FP4 or FP8 Wan 2.2 pipeline from Wan itself, so we size on BF16. Our page on the four-card RTX PRO 6000 Server Edition server covers chassis, MIG and the power modes.

The H200 NVL server is the faster box for this one model. Its higher BF16 rate shortens every clip, its 141 GB keep both experts and the encoder on the card, and its NVLink bridge makes sequence parallelism worthwhile. Our four H200 NVL server page covers bridges, chassis and host memory, and the four RTX PRO 6000 or four H200 NVL comparison weighs both for LLM serving.

We build both servers to order and size them by the clips your team produces at the busiest hour; send us your resolutions, clip lengths and the number of people submitting jobs, and our team replies with a configuration for each layout.

Wan2.2-Animate-2 and other Wan models on the same server

Wan2.2-Animate-2-14B animates a character from a reference image and a driving video. Its card, under Apache 2.0, gives 7 August 2026 as the release of the inference scripts and the base and distilled weights, and says the defaults are "tuned for 8× A800 GPUs" at 720p, with 480p also tested on two A800. The distilled version runs 10 steps with guidance off, a quarter of the base model's 40 steps and no second pass per step, so the same card delivers several times as many clips with it by the arithmetic above. The card gives no memory figure, so test it on one card before planning a queue; game studios that use it for previs will find our AI server for game studios useful for the rest of the pipeline.

Wan2.2-TI2V-5B, the small model, generates 720p at 24 frames per second and runs on "a GPU with at least 24GB VRAM" with offloading, so one RTX PRO 6000 can hold several copies or MIG instances of it for previews. Wan2.2-S2V-14B, the speech-to-video model, states the same 80 GB as the A14B models. The Wan 2.2 cards we read on 24 October 2026, T2V-A14B and Animate-2-14B, name Apache 2.0, and the repository states that "The models in this repository are licensed under the Apache 2.0 License"; Apache 2.0 grants commercial use without a territory or revenue condition; our video guide compares this with the licences of other open video models. For a marketing or agency team, our AI server for creative agencies puts video next to image generation and rendering.

Power, cooling and the feed

Our estimate, as on our configuration pages: cards at 600 W, two AMD EPYC processors (2 × EPYC 9135 with two cards, 2 × EPYC 9355 with four, 2 × EPYC 9555 with eight) and 15 per cent for memory, drives, fans and conversion losses; a measurement on the delivered system confirms or replaces it. For eight-card nodes Dell's datasheet gives about 8 kW as an indicative figure; plan the feed on the server maker's figure.

Video generation keeps the cards at full load for minutes per clip, so plan the feed on the full figure. At about 3.4 kW, a four-card server is above the 2.94 kW we plan per 230 V 16 A feed at 80 per cent and needs 32 A or three-phase, as our guide to 16 A, 32 A and three-phase feeds explains. The RTX PRO 6000 Server Edition can be capped at 450 W, which lowers the server to about 2.7 kW at a cost in speed; StorageReview measured 2 to 9 per cent on LLM work, and we found no figure for video models. The H200 NVL at 450 W is covered in our article on running the H200 NVL at 450 W.

Plan the storage as well. The A14B repository alone is 126 GB, and a server that keeps several Wan models and their Diffusers versions ready holds several hundred GB on NVMe by our estimate, plus the clips themselves. Store prompt, seed and model revision with every clip so that a result can be reproduced.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the RTX PRO 6000 Server Edition and the H200 NVL with its NVLink bridges and build AI servers to order with two, four or eight cards, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. The operating system, drivers, CUDA and a container runtime are installed on request; quantities and lead time are confirmed in writing with the quote.

FAQ