Rent a 48 GB GPU, load the official Wan 2.2 template, and generate a video. It will work. No error, no warning, no out-of-memory crash. It will also take three to four times longer than it should, because the card never actually loaded the model, and ComfyUI doesn’t flag it.
To find where it starts, we benchmarked 23 GPU configurations on Wan 2.2 and LTX-2.3 through ComfyUI, measuring wall clock, peak VRAM and cost for every card. A 5-second Wan 2.2 clip ranges from 8 minutes to 69 minutes depending on the GPU. Most of that gap isn’t about how fast the card is.
So the best GPU for AI video generation is less about picking the quickest chip than about three decisions in order: which model tier you render at, whether the card holds what you load, and only then how fast it runs.
Why a 48 GB GPU runs Wan 2.2 three to four times slower
Wan 2.2 at fp16 needs about 64 GB. Cards that hold it show a peak VRAM figure around 47 to 61 GB. Cards that do not show 10 to 16 GB, because they are holding a fraction of the model and fetching the rest from system memory as they go.
The line falls between 48 GB and 80 GB, and it falls in the same place on the fp8 build. That’s the surprising part: the fp8 weights are about 36 GB, which fits inside a 48 GB card with room to spare, and the 48 GB cards stream it anyway. Every card measured at 80 GB or more holds fp8 at a peak of about 34 GB. Every card at 48 GB or below peaks between 2 and 14 GB.
So the practical rule, measured at ComfyUI’s default template settings, is simpler than the model size suggests: cards below 80 GB stream the weights, at either precision. Quantizing to fp8 buys you speed on the cards that already held the model. It doesn’t move the line.
ComfyUI’s memory settings likely play a part too. A 48 GB card has 14 GB of headroom over the 34 GB working set and offloads anyway, and two cards with the same 48 GB land in different places, the L40S at 5.2 GB and the A40 at 12.7.
Wan 2.2 fp16: speed and cost by GPU
1280x704, 81 frames at 16 fps, a 5.06-second clip, 20 steps, ComfyUI’s official Wan 2.2 template. Fifteen cards, sorted by wall clock.
Cost spans 3.1x while wall clock spans 5.5x, and the two don’t line up. The lowest per-clip costs come from cards that stream the model, and they finish mid-pack or slower. The most expensive, $1.1163 on a B300, buys the fastest render in the set.
The useful comparison is among cards that actually hold the model. There the RTX PRO 6000** at $0.4261 and 734 seconds is the best value of those**. The next cheapest holder is the A100 SXM at $0.5696, which costs 34% more and takes 1.76 times longer. The B300 is 1.44 times faster and costs 2.6 times as much.
Wan 2.2 fp8: speed and cost by GPU
Same template, same clip, fp8_scaled experts. Twenty-two cards.
fp8 helps on cards that hold the model and barely moves the ones that don’t. An RTX PRO 6000 goes from 734 to 656 seconds, an 11% saving. An H100 SXM goes from 652 to 616. But an A40 goes from 2,810 to 2,749, which is 2%, and an RTX A6000 from 2,616 to 2,654, which is slower.
An RTX A5000 renders the clip for $0.2585, the cheapest Wan clip in this dataset, and takes 57 minutes to do it. An L4 takes 69 minutes and costs more than an RTX PRO 6000 clip that finishes in 11.
Don’t quantize in order to fit Wan 2.2 onto a smaller card. It doesn’t work, and the cost figure that tempts you there is measuring a card that is idle most of the time. Quantize to make a card you already chose faster. If your budget stops below 80 GB, the productive move is to change model tier rather than precision.
LTX-2.3 distilled: speed and cost by GPU
768x512, 97 frames at 25 fps, a 3.88-second clip, 8 steps, ComfyUI’s official LTX template. Twenty-three GPU configurations, including two MIG slices. This is a different quality tier from Wan and the numbers aren’t comparable to the two tables above.
The RTX PRO 6000 is the fastest card in this table, ahead of the H200 and the B300, at under a quarter of the B300’s cost per clip. The RTX PRO 4500 at $0.72 an hour is fifth overall, faster than an H100 NVL costing 4.4 times more, and the second cheapest clip in the set.
Note the A100 PCIe at the bottom. It holds the model, it’s an 80 GB datacenter card, and it renders this clip 2.7 times slower than the A100 SXM. Short renders like this are sensitive to the machine around the GPU, which the next section covers, so don’t read this row as a verdict on the card.
Model choice moves the number more than GPU choice does
Wall seconds per second of finished video is the one figure comparable across clip specs. On a single RTX PRO 6000:
On the same card, Wan 2.2 fp16 takes about 75 times longer per second of video than LTX-2.3 distilled. Within Wan 2.2 at fp16 the spread from the best GPU to the worst is 5.5 times. The decision that moves your render time by two orders of magnitude is which model and precision you load.
These are different quality tiers at different resolutions and step counts, and this dataset measures no quality at all, so this isn’t an argument that LTX is better. It’s an argument that the tier decision comes first and the GPU decision second.
Decide the tier before you shop for a card. If a distilled model meets your quality bar, almost any GPU in this article is fast enough and the smaller ones are inexpensive. If it doesn’t, you’re in Wan territory and the 80 GB floor applies. Switching cards later is easy; switching tiers can mean renting a different card.
Which GPU should you pick for AI video generation?
Best overall: RTX PRO 6000
It’s the fastest card measured on LTX-2.3 at 7.5 seconds, and on Wan 2.2 it’s the only card that is both near the top on speed and near the bottom on cost: 734 seconds at $0.4261 on fp16, 656 at $0.3807 on fp8. Every card that beats it on time costs at least 1.5 times as much per clip.
Fastest for Wan 2.2: B300, then B200
509 seconds at fp16 and 482 at fp8 on the B300, about 5% ahead of the B200 at both precisions. A B300 clip costs $1.12 at fp16 and $1.06 at fp8, and a B200 clip about $1.01 and $0.96. These are the most expensive figures in the set. Worth it when turnaround matters more than budget.
Cheapest per clip that still holds the model: RTX PRO 6000
Among cards that hold Wan 2.2 rather than streaming it, nothing undercuts the PRO 6000. Cheaper figures exist in the table, but every one of them belongs to a card on the wrong side of the memory cliff.
Budget pick for LTX: RTX PRO 4500
$0.002642 a clip at 13.2 seconds, fifth fastest of the 23 configurations, at $0.72 an hour. The RTX A5000 is marginally cheaper per clip at $0.002625 but takes 35 seconds.
What latency looks like when a card is shared
Ten cards ran an offered-load ladder on LTX-2.3: requests arriving at random intervals against a warm ComfyUI server at a fraction of the card’s own calibrated ceiling. ComfyUI renders one video at a time, so throughput is bounded by single-clip time. What the ladder shows is the wait.
At half load the p95 wait runs two to four and a half times the single-clip time on every card. On an RTX 6000 Ada , the slowest 5% of requests waited about 27 seconds for a clip that renders in 11.
The A100 SXM and RTX PRO 4500 ladders are excluded from this table. Both ran on machines that render the single clip roughly twice as slowly as the cards’ published medians, so their ladder figures describe those machines rather than those cards.
Why the same GPU can be twice as fast on another machine
On short renders the host CPU is a first-order term. An A100 SXM rendered the LTX clip in about 15 seconds at one datacenter and 34 at another, same graph, same card. An RTX PRO 4500 ran 13 seconds at one site and 33 at another.
The effect disappears as the render gets longer. The B300’s Wan 2.2 fp16 row spans machines with two different CPU generations and they agree within 3% at a 509-second render.
Treat the LTX-scale numbers as approximate and the Wan-scale numbers as firm. If short-clip latency is what you’re optimizing, benchmark on the pod you actually got rather than trusting any table, including this one, and redeploy if the first one is slow. This is a property of short renders rather than of any provider, which is also why benchmarking two providers against each other on a 10-second clip tells you very little about their hardware.
Common questions about GPUs for AI video generation
What is the best GPU for AI video generation?
The RTX PRO 6000. It renders an LTX-2.3 clip in 7.5 seconds, faster than any other card measured, and a 5-second Wan 2.2 clip in 734 seconds at fp16 for $0.4261, which no card that holds the model beats on cost. A B300 is about 44% faster on Wan at fp16 and costs 2.6 times more per clip.
How much VRAM do you need for Wan 2.2?
80 GB, at either precision, at ComfyUI’s default template settings. Wan 2.2 at fp16 is about 64 GB and at fp8 about 36 GB, but every card at 48 GB or below streamed the weights from system RAM rather than holding them, including on the fp8 build that fits inside 48 GB on paper. Cards that hold it show peak VRAM of 34 GB on fp8 and 47 to 61 GB on fp16.
The fp8 result is worth treating as a setting rather than a hardware limit until tested. A 48 GB card has ample headroom over a 34 GB working set, so if you’re on one, try ComfyUI’s VRAM flags before moving to a larger card.
Can I run Wan 2.2 on a 24 GB card?
Yes. An RTX PRO 4000 renders the fp8 clip in 2,375 seconds, about 40 minutes, compared with 656 seconds on an RTX PRO 6000. It finishes without errors, but much of the extra time goes to streaming the model from system memory. At $0.38 a clip, that can work for iterating on prompts, but it’s slow for volume.
How long does it take to generate a 5-second video with Wan 2.2?
Between 8 and 69 minutes depending on the card, at 1280x704 and 20 steps. At fp16, a B300 takes 8.5 minutes, an RTX PRO 6000 12, an A100 SXM 21 and an A40 47. At fp8, a B300 takes 8 minutes and an L4 69.
Is fp8 worth it for Wan 2.2?
On cards that hold the model, yes: about 11% faster on an RTX PRO 6000, and 5 to 6% on the H100 SXM, B200 and B300. On cards below 80 GB, no. An A40 improves 2% and an RTX A6000 is slightly slower at fp8 than at fp16, because both are streaming the weights either way and the bottleneck isn’t arithmetic.
Which is cheaper per clip, a smaller card or a datacenter card?
On LTX-2.3 the cheapest clips come from small cards, an RTX A5000 at $0.0026 and an RTX PRO 4500 at $0.0026, and the most expensive from a B300 at $0.0190. On Wan 2.2 the picture inverts once you account for the memory cliff: the cheap figures belong to cards that are streaming the model, and among cards that hold it the RTX PRO 6000 is both the cheapest and nearly the fastest.
Do I need a B200 or B300 for video generation?
Only if wall clock is the binding constraint. They are the two fastest cards on Wan 2.2, at 509 and 538 seconds, and the two most expensive per clip at $1.12 and $1.01. An RTX PRO 6000 takes 734 seconds for $0.4261, which is 44% more time for 62% less money.
Compare these figures interactively, alongside LLM, image and speech workloads, on the GPU Compare tool.
