How to choose GPU for AI video upscaling?

Question

I’m trying to pick a GPU mainly for AI video upscaling (Topaz Video AI / similar tools) and I’m getting lost in the VRAM vs CUDA/Tensor core debate. Most of my source clips are 1080p/4K and I’d like to upscale to 4K (sometimes 8K) with reasonable render times, but I also don’t want to overspend. I’m on Windows and currently using an older 6GB card that keeps running out of VRAM on heavier models. How should I choose between something like 12GB vs 16GB+ VRAM, and how much do GPU generation and memory bandwidth actually matter for AI upscaling performance?

uhpyggezqo · Accepted Answer

+1 to what was said earlier — VRAM is totally the “will it even run?” gate. Been thinking about your question cuz I went through this with my old 6GB card too… Topaz would just randomly bail mid-export, super annoying lol.

What I learned the hard way:
- **VRAM spikes aren’t linear**: heavier models + higher output + longer clips can suddenly jump memory usage, even if 4K sources felt “fine.”
- **Gen/bandwidth matters once you fit**: if you’re not VRAM-bound, newer Tensor cores + higher memory bandwidth usually means noticeably better fps. When you are VRAM-bound, speed doesnt matter cuz it’s paging/crashing.
- **8K is the trap**: it’s not just 2x work vs 4K, it’s a lot more memory pressure.

What’s your typical clip length and do you batch render overnight or want near-real-time-ish?

ouygvwlqyh · Answer

For your situation, I’d treat VRAM as the “can I even run this model?” gate, and generation/bandwidth as the “how fast is it?” knob.

- **Option A: 12GB (budget)**: NVIDIA GeForce RTX 4070 12GB-class. Pros: great speed per $, modern Tensor cores. Cons: 12GB can still choke on heavier Topaz models or 8K, esp with big tile sizes. I had to babysit settings a lot.
- **Option B: 16GB (sweet spot)**: NVIDIA GeForce RTX 4070 Ti SUPER 16GB / NVIDIA GeForce RTX 4080 SUPER 16GB. Pros: way fewer VRAM errors, smoother 4K->4K and some 8K. Cons: costs more.
- **Option C: 24GB+ (no drama)**: NVIDIA GeForce RTX 4090 24GB. Pros: basically brute-forces everything. Cons: $$$

Also, memory bandwidth matters more than ppl admit… when VRAM isn’t the bottleneck. What CPU/RAM you on?

gdqgyszomr · Answer

Seconding the above — VRAM is the “will it run without dying?” gate, 100%. For Topaz, 12GB is ok for lots of 1080p→4K, but if you’re doing 4K sources + heavier models or wanna do 8K sometimes, I’d honestly target 16GB+ so you’re not constantly babysitting exports. Value picks: NVIDIA GeForce RTX 4070 Ti SUPER 16GB is a really nice sweet spot; if you wanna go cheaper, NVIDIA GeForce RTX 4060 Ti 16GB works but it’s slower (memory bandwidth matters a ton). And yeah gen matters too—newer Tensor cores + better drivers = faster per watt. Also quick tip: render in smaller chunks / lower preview quality when testing models… saves so much pain lol. gl!