Video generation is the most compute-hungry thing most teams will ever serve. A text model writes one token per step. A video diffusion model rebuilds every frame of the clip on every denoising step, dozens of times per request. That makes the GPU choice for video different from the one for language models. Bandwidth matters less, compute and memory for activations matter more, and the cheapest GPU per clip is often not the cheapest per hour, or the fastest.
Why video is compute-bound
A video diffusion transformer turns the clip into a grid of latent tokens and runs attention over all of them at each step. HunyuanVideo 1.5, for example, compresses video 16 times in each spatial dimension and 4 times in time before the transformer sees it.1 The token count grows with resolution × frame count, and attention cost grows with the square of the token count.
Here is what that means for the same 5-second clip at different resolutions:
| Resolution | Pixels per frame | Tokens vs 480p | Attention cost vs 480p |
|---|---|---|---|
| 480p (848 × 480) | 407,040 | 1× | 1× |
| 720p (1280 × 720) | 921,600 | ~2.3× | ~5.1× |
| 1080p (1920 × 1080) | 2,073,600 | ~5.1× | ~26× |
Doubling clip length doubles tokens and roughly quadruples attention cost again. That is why many open models, including HunyuanVideo 1.5, generate at 480p or 720p and then upscale. HunyuanVideo 1.5 ships a separate super-resolution step to reach 1080p.1
Thousands of tokens per step keep the Tensor Cores busy, so per-step speed tracks compute more than memory bandwidth. Memory still sets the floor: the weights, the text encoder, the VAE and the activations for the whole clip must fit on the GPU, or across the GPUs in one runner.
Step 1: find the smallest GPU that fits
Weights take about 2 GB per billion parameters at BF16. Three open models show the range:
| Model | Parameters | BF16 weights | Notes from the model card |
|---|---|---|---|
| Wan 2.2 TI2V-5B | 5B | ~10 GB | 720p at 24 fps; runs on a single consumer graphics card2 |
| HunyuanVideo 1.5 | 8.3B | ~17 GB | 480p and 720p, with 1080p super-resolution1 |
| Wan 2.2 A14B | 27B total, 14B active per step | ~54 GB for both experts | 720p on one GPU needs at least 80 GB, with offloading3 |
The Wan 2.2 A14B row is the important one. It is a mixture-of-experts model: one 14B expert handles the early, noisy steps and another the later ones.3 Only one runs at a time, but both should stay in memory, or the runner reloads weights mid-clip. An 80 GB GPU can run it at 720p only by offloading parts to CPU memory, which costs time on every request. A GPU with 141 GB or more should keep both experts resident, with room for activations at typical clip lengths. Check peak memory on your own settings.
Matching models to GPUs on fal:
| What you run | Start with | Why |
|---|---|---|
| Models up to ~10B at 480p or 720p | RTX PRO 6000 (96 GB) | Enough memory, lowest rate, hardware video encoders |
| 14B-class models at 720p, one clip at a time | H100 (80 GB) or RTX PRO 6000 | Fits at BF16; check activation memory at your clip length |
| Mixture-of-experts video models, longer clips, or two clips per GPU | H200 (141 GB) | Both experts resident, no offloading |
| 1080p, long clips, or chained models on one GPU | B200 (192 GB) | High compute per GPU at BF16 and FP8 (same as B300), and self-serve per second |
| Several large models resident at once | B300 (288 GB) | Most memory per GPU; capacity set up with fal's team |
Step 2: price per clip, not per hour
You pay for the seconds a GPU holds a request. So the useful number is how much faster a pricier GPU must render a clip to cost the same. Against the RTX PRO 6000, the cheapest GPU on fal's pricing page:4
| GPU | List price | Must render faster than RTX PRO 6000 by (list) | "As low as" | Must render faster by ("as low as") |
|---|---|---|---|---|
| RTX PRO 6000 | $4.00/h | 1× | $1.99/h | 1× |
| H100 | $4.50/h | 1.13× | $2.49/h | 1.25× |
| H200 | $6.00/h | 1.50× | $2.99/h | 1.50× |
| B200 | $7.99/h | 2.00× | $5.49/h | 2.76× |
| B300 | $12.99/h | 3.25× | $5.99/h | 3.01× |
If a B200 renders your clip in less than half the RTX PRO 6000's time, it is cheaper per clip at list price. Otherwise it buys speed, not savings. Measure render time per clip on two GPUs, divide, and compare with this table.
Worked cost per clip
Cost per clip is GPU-seconds × the per-second rate. At list price, one minute of GPU time costs:
| GPU | Per GPU-minute (list) | Per GPU-minute ("as low as") |
|---|---|---|
| RTX PRO 6000 | $0.067 | $0.033 |
| H100 | $0.075 | $0.042 |
| H200 | $0.100 | $0.050 |
| B200 | $0.133 | $0.092 |
| B300 | $0.217 | $0.100 |
Some examples at list price:
- A clip that holds one H100 for 90 seconds costs about $0.11. 1,000 clips a day for 30 days is about $3,380 a month.
- The same model at 720p instead of 480p has about 5 times the attention work. If that makes it a 5-minute render, each clip costs $0.38 on the H100. A B200 that does it in 2 minutes costs $0.27 and returns it more than twice as fast.
- Splitting one clip across four H100s with sequence parallelism bills 4 GPU-seconds for every second of wall time.5 If four GPUs cut the time by 3.5 times, not 4, each clip costs 14% more, in exchange for a 3.5 times faster result.
Step 3: account for encoding and the rest of the pipeline
The H100, H200 and B200 have no hardware video encoder, so the final MP4 is encoded on the CPU.6 The RTX PRO 6000 has four NVENC engines with AV1 support, which output frames without using GPU compute time.67 For short clips, encoding is a small share of the request. For long clips or high frame rates, it can become a visible part of latency on GPUs without an encoder. fal's machine-type guidance recommends the RTX PRO 6000 for video generation for this reason.6
Larger pipelines change the answer again. A runner that also holds an upscaler, an interpolation model or a safety checker needs memory for all of them, which pushes toward an H200, B200 or B300.
How to choose, in order
- Find the smallest GPU where the full pipeline fits at your target resolution and length without offloading.
- Render the same 20 prompts on that GPU and on the next one up. Record GPU-seconds per clip.
- Divide the two times and compare with the break-even table above.
- If latency matters more than cost, add GPUs per runner and accept a higher cost per clip.
For the deployment side, covering queues, weight caching, cold starts and multi-GPU code, see how to deploy a video generation model. For hourly rates and billing, see the price guides for the RTX PRO 6000, H100, H200, B200 and B300.
FAQ
What is the best GPU for video generation?
For open models up to about 10B parameters at 480p or 720p, an RTX PRO 6000 is usually the cheapest per clip: it has 96 GB, hardware encoders and the lowest rate on fal's pricing page. Mixture-of-experts models such as Wan 2.2 A14B fit better on an H200 or larger. For 1080p, long clips or the shortest render times, use a B200.
How much VRAM do I need for video generation?
Start at about 2 GB per billion parameters at BF16, then add the text encoder, the VAE and the activations, which grow with resolution and clip length. A 5B model runs on a single consumer graphics card. A 27B mixture-of-experts model needs at least 80 GB with offloading, and more to run without it.
Is the H100 good for video generation?
It works well for models that fit in 80 GB, and it lists at $4.50 per GPU-hour on fal. It has no hardware video encoder, so encoding runs on the CPU. It must render about 1.13 times faster than an RTX PRO 6000 to cost less per clip at list price.
How much does it cost to generate a video on my own GPU?
GPU-seconds per clip × the per-second rate. On fal at list price, a clip that holds one H100 for 90 seconds costs about $0.11. The same 90 seconds costs $0.10 on an RTX PRO 6000 and $0.20 on a B200.



