Upscaling SD video to FHD with Topaz Video AI: how long could it take?

[Mounter81]

Hello!

So, I have a briefly entry-mid level computer, with 16GB 3200MHz RAM and an i3-10105 @ 3.70GHz processor with UHD630 IGP (so no dedicated VGA jet).

I'm planning in the future to upscale some SD videos the to at least FHD(-ish) resolution, with also enhancing the videos picture quality as much as possible. And therefore I'm up to buy a "relatively" good performing (but still somewhat budget) VGA, like an Nv 1660 Super.

So my question would be, that: "estimatedly" how long would it take to convert an SD (lets say 1 hour long, with 1414 kb/s bitrate and 680x430p resolution) video to FHD with the best possible picture quality enhancer options with an 1660 Super VGA and Topaz Video AI?

Hope that someone here has some kind of relevant experience regarding and could answer somewhat helpfully to this question...

Thank you!

[RetsimLegin]

Using an entirely different AI upscaler on a box with
i7-1165G7 (2.8GHz, 4.7GHz Turbo)
64GB Corsair 2666MHz SODIMM DDR4
NVIDIA® GeForce® GTX 1650 Ti - 4.0GB GDDR5 Video RAM
a 45 minute PAL SD title takes around 9-10 hours to upscale 2x.

[Mounter81]

Using an entirely different AI upscaler on a box with
i7-1165G7 (2.8GHz, 4.7GHz Turbo)
64GB Corsair 2666MHz SODIMM DDR4
NVIDIA® GeForce® GTX 1650 Ti - 4.0GB GDDR5 Video RAM
a 45 minute PAL SD title takes around 9-10 hours to upscale 2x.

And during the upscaling (checking in Task Manager) how busy is the CPU and how much is the GPU?

Thanks for the reply!

[RetsimLegin]

I haven't ever bothered to watch the resource screen, but I can tell you the cooling fan ramps up to somewhere like max and stays there the full duration. So it's obviously very "busy".

And by way of comparision, a simple re-encode or simple Lancoz upscale using avidemux takes ~20 minutes for a 2 hour movie.

[Mounter81]

I haven't ever bothered to watch the resource screen, but I can tell you the cooling fan ramps up to somewhere like max and stays there the full duration. So it's obviously very "busy".

And by way of comparision, a simple re-encode or simple Lancoz upscale using avidemux takes ~20 minutes for a 2 hour movie.

I'm not very familiar with Avidemux, so (more than) just for curiosity: does Avidemux make any visible picture quality improvement with this upscaling, or it's just a simple "resolution" upscale?

[RetsimLegin]

It's nothing like AI. Very similar (to my eyes) to the upscaling done by the display if I feed it PAL SD "raw". Maybe a tiny bit of an improvement, but not so I'd notice any "wow"!

But for many other purposes, I love avidemux. It has some good deinterlacers, and can be used to, say, fix PAL speedup (by retiming both video and audio to the defacto standard 23.976fps). I have adopted a practice of re-encoding any SD content to 1280x720p (note, not 1080p, for reasons of space/bitrate and that it doesn't offer much if any advantage with SD sources) at 23.976, regardless of PAL/NTSC in the source. And many more video filters.

[Emulgator]

Topaz (I am on 2.6.4) wants to see some more elevated GPUs.

Running Topaz on Intel UHD750 is a waiting game:
Proteus-upscaling any 240p..SD-ish source to 4k runs 0,3fps, using a 11900K 17..24% plus Intel UHD750 90-100%-ish.
Proteus-upscaling any 240p..SD-ish source to FHD runs 0,7fps, using a 11900K 17..24% plus Intel UHD750 90-100%-ish.

Running Topaz on the same system, but pointing it to a GeForce 3080 comes out 8..12 times faster here:
Proteus-upscaling any 240p..SD-ish source to 4k runs 3fps, using a 11900K 40..100% plus a GeForce 3080 30-40%-ish.
Proteus-upscaling any 240p..SD-ish source to FHD runs 10fps, using a 11900K 40..100% plus a GeForce 3080 30-40%-ish.

If Topaz, I wouldn't spend any time on 9xx,10xx nVidia cards.
With 20xx I have no experience, only hearsay of being less capable then their (initial) price might have suggested.
From 3060 on Topaz and other AI apps start to play ball.
Any returns from going -> 3070 -> 3080: may still be worth an investment, was to me at last.
From 3090 on into 4090 diminishing returns are reported.
Worth reading:
https://www.pugetsystems.com/labs/articles/topaz-ai-cpu-gpu-performance-analysis/