I haven’t tested the 5.x.x series as daily usage seems to be about the same as 4.x.x.
The best performance to watt is the AMD 7900XTX. If you’re working 4X upscaling then agree the performance difference of 4090 with 7900XTX at 4-5 fps isn’t going to make much of a difference, but the 4090 is the fastest right now.
But since the 3080Ti isn’t much slower than a 3090, buying a 7900XTX isn’t going to make much of a difference.
So not much difference form Video AI 4.x to 5.x for the 4090.
I’d recommend waiting for AMD RDNA 4 (8800XTX using DDR6) GPU due out Dec 2024 or Jan 2025. It’s about 25% increase over the AMD 7900XTX (which should make it slightly better than a 4090) and operates at 410 Watt max (about 200 watts less than the 4090 as same level of performance).
With that said, nVidia 5000 series due out Q1 2025 should use a little less power at 500-550 watts but 50% performance increase (using DDR7).
Anyway, back on topic, please hire some good programmers Topaz. Maybe you feel sympathy for whomever is currently coding Video AI, but if you want to keep me interested, the development team will need to do better, A LOT better … need to see some progress here and that’s just not happening.
H100 $28,000 each or H200 $40,000 each. But I think you’ll be shocked that the single GPU H100 is not even 2X faster than a 4090.
Where H100 benefits is that you can run 8 or more together but again with very specialized other components to make that work.
Alternative render out to a farm at $2.30/hr for 1 GPU shared allocation. Add multiple GPUs along with any “correction” and you’ll be paying A LOT daily. I think the H100 and H200 are used for large companies where AI models are used for a multitude of things other than just upscaling or improving video.
I had a 3080Ti / 5950X rig and read through a lot in the Topaz benchmark threads. At the time, both components were “balanced” in that upgrading either one would not gain me any meaningful uplift because the old component would hold back the new component’s performance.
