GPU advice on 5080

Since it’s a really expensive purchase, I think it’s better to be cautious… It’s better than forgetting to ask after you’ve already bought it…

other reports say the blackwell architecture is weak for AI image>video generation

That report was accurate in the first half of last year, but it’s now just a misunderstanding. That’s because, at the time, ComfyUI and one of its core components (PyTorch) weren’t fully compatible yet.

Blackwell is fundamentally a better architecture for AI than previous generations.

While there are some poorly made Blackwell cards out there, that’s not unique to Blackwell… seriously.

Also, the reason you’re confused is that “workstation and data center-grade cards”—which weren’t part of previous comparisons—have now been added as an option. All the GPUs we’ve considered so far have been designed for gaming. Gaming GPUs prioritize peak performance, consumer-friendly features, and flashy lighting effects. Workstation-grade cards, on the other hand, prioritize power efficiency, stability, computational accuracy, and durability.

“If I could buy a card with the same price, performance, and VRAM—and if it fits in my PC—I’d choose a workstation-grade card without hesitation.”

Since they’re more energy-efficient, they put less strain on your PC’s power supply than gaming cards. Normally, workstation-grade cards are extremely expensive, but… even if the processing speed drops slightly—say, by a few percentage points—they’re still a very attractive option.

What makes this confusing is that people often mix up three different things : the architecture , the specific card , and the software stack. Blackwell itself is not the real problem now: PyTorch 2.7 added Blackwell support, so the old “Blackwell is bad for AI” line is mostly launch-era baggage. The RTX PRO 4000 Blackwell looks attractive because it has 24GB GDDR7 ECC , but it is also a workstation card built around lower power, compact form factor, and pro reliability rather than maximum consumer throughput. At the same time, image-to-video is a heavier workload than plain image generation because memory demand grows with frames, resolution, and model size. Wan 2.2’s official TI2V 5B model card explicitly frames a high-end consumer GPU as the natural reference point for 720p/24fps generation, which is why 24GB-class cards keep coming up in these conversations. (pytorch.org) (nvidia.com) (huggingface.co)

The RTX PRO 4000 Blackwell is therefore not a bad card. It is just a different kind of card. It gives you workstation strengths: ECC memory, pro drivers, lower power, compact size, and video engines. Those things are useful in pro environments, multi-app reliability, or space-constrained systems. But for a normal desktop running ComfyUI image-to-video , they do not automatically make it the smartest buy. What matters more in that use case is usually raw throughput per pound plus how smooth the Windows software path is. (nvidia.com) (pytorch.org)

That is why my answer changes depending on what you want:

If you want the simplest, strongest Windows choice , pick the RTX 5080. It is the cleanest “buy it, install current drivers, use current ComfyUI and PyTorch, and get on with it” answer. The downside is simply that it is 16GB , not 24GB. (pytorch.org)

If you want the cheapest sensible GeForce option , pick the RTX 5070 Ti. It is the easiest card to justify when the current market feels insane, because it fixes the real pain of the 4060 8GB without pretending to be a forever card. (nvidia.com)

If you specifically want 24GB but do not care about workstation extras, the RTX PRO 4000 Blackwell becomes harder to justify. It is very appealing on paper because “24GB GDDR7 ECC” sounds perfect, but the more honest reading is: great reliability card, less obvious value card. Unless you have a reason to want workstation traits, I would not make it the default recommendation for hobbyist or enthusiast ComfyUI image-to-video. (nvidia.com)

If you want 24GB at the best raw value , the RX 7900 XTX is the obvious alternative. The reason people keep bringing it up is simple: 24GB and a 384-bit bus are genuinely useful for heavier image and video workloads. The trade-off is that the Windows path is still less relaxed than GeForce. (amd.com) (docs.comfy.org)

So the clean summary for your purpose is:

• Best overall: RTX 5080
• Best value: RTX 5070 Ti
• Best workstation-only case: RTX PRO 4000 Blackwell
• Best 24GB value alternative: RX 7900 XTX

And the plain-English answer to your original confusion is:

Blackwell is not weak here. What is confusing people is the difference between a workstation Blackwell card and a consumer Blackwell card. The PRO 4000 is built to be professional, stable, compact, and efficient. The 5080 is built to be faster for consumer-style GPU workloads. For ComfyUI image-to-video, that distinction matters more than the architecture name. (nvidia.com) (pytorch.org)