Thanks for posting a driver update does jack shit. Now I don’t have to update drivers.
Ok so I went and bought an MSI RTX 4090, installed the latest studio driver, and compared to my old RTX 3080 the results were…not impressive.
Upscaling a 1080p to 4K video my seconds per frame dropped from 0.46 to 0.35 using Gaia HQ, however Artemis HQ was the same and Proteus 2 actually did worse, increasing from 0.46 spf to 0.51
Doubling a 720p video to 1440p again Gaia improved from 0.23 spf to 0.17. However Artemis HQ only got better by 0.01 spf and Proteus 2 got worst by that amount.
Doubling a 480p video to 960p all models performed better, however it was only by 0.01-0.02 spf.
In conclusion, if you want to buy a graphics card just for Topaz VEAI save your money and get a cheaper model.
3 Likes
To me, this is already a very telling test in confirming that still TVAI Beta does not exploit the Tensor Cores of RTX GPUs well!
Welcome to the “upgrading didn’t speed up VEAI that much” club. At least the new 3 beta reduces GPU pwr/consumption by a good bit to bring nice speedups on some mobile RTX GPUs. Now for the question of, does stacking a A770 + 3070/3080 have a speedup greater than a 4090 at the cost of a 3090? 
1 Like
No surprise there, a 3090 runs at 47% average utilisation on my machine for 1080p → 4K. Can barely tell the difference between a 3070. All in all I think RTX cards have a lot of untapped power when it comes to Video Enhance. I have not tried the v3 beta but I heard usage is better there. How many upscales can you run with a 4090 without any noticeable slowdown? With Gaia on a 3090 at that resolution it’s about 2.
You mean how does it perform running parallel jobs? Is that even possible with 2.6.4 as I haven’t tried out the beta yet.
you can do it with 2.6.4 by launching the software several time like 2 or 3 times.
for the beta/EA, it’s done in preference, it’s an option to turn on, and fortunatly it does it in one app.
1 Like
Using my 4090 I tried upscaling two 1080p videos to 4K with Gaia at once. When ran separately the first had an estimated time of 1:15 and the second one 1:42. When I ran them both at the same time the times jumped to 1:45 and 2:20.
I repeated the test upscaling two 720p videos to 1440p. When ran in parallel the estimated times jumped even higher, from 2:50 and 1:40 to 3:42 and 2:13.
Finally one last time upscaling 480p videos to 960p. This time the jumps were far less, from 1:35 and 0:42 to 1:54 and 0:51. Adding a third video caused the times to jump even more to 2:05 and 1:01.
So while running parallel jobs does slow things down it still saves you time compared to running them in sequential order.
ask @Imo, he’s the one i know that used a lot 2.6.4 version separatly at the same time.
i have a s*** card with 4Gb ram, so never did this.
did you tested the 4090 with the 3.0 beta/EA ? do you have better processing time than 2.6.4 ?
anyway, from what i read/know, speed improvement has not be done yet on the 3.x
3.x has a parallel processing included.
anyway, between the drivers and very recent 4090 release and the softwares not optimised at all for this card, it’s looking pretty normal that it’s not running as it could be expected by “owning” the best card on the market actually (for “normal” user).
certainly at some point, the optimisation for such a card will be done but not on 2.6.4 but on the 3.x. ask the dev (suraj) what he think about the 4090 Rtx and the future of VEAI for such a card (as a suggestion 
)!
Can you please check with nvidia-smi.exe in what power state your 4090 while upscaling 720p?
On my 3090 it is always P3, but when I resize frame and make it 5-6% larger, card sometimes comes in P0 state (and sometimes not) and it speeds up process.
As a result in Proteus FT mode pure 1280x720 shows 0.16-0.17s for frame and same video resized to 1344x756 may show 0.14-0.15s
This is what it says when upscaling a 720p video to 1440p using Proteus 2:
When I change it to 206% upscale it still says P3.
so in both cases it in in P3 state and use only 15% of it power…
I doubt they are paying attention to this anymore. All focus is on the 3.x beta and early releases. 2.x won’t be updated.
On this note… if you’re using Windows (or just check it anyways), running Telegram can drastically cap clocks for no good reason. Dunno why, but it manages to do that, my 3070ti normally runs upscales at a peak of 1985-2040Mhz but occasionally I’ve found it to be bugged with Telegram open (minimized / in tray even!), reducing the clocks to 1765~ MHz which is one of the middle power states. This hurts all GPU compute (Hashcat too, lol!) for no good reason. Once I closed Telegram and restarted the render/program, things go up to full speed and I gain a few % more perf.
Could you check out the 3.0 betas https://community.topazlabs.com/t/topaz-video-ai-early-access-v3-0-0-8/35610 and see if that makes any difference to your power consumption? Oddly enough I found my 3070’s power usage to drop considerably (same with a A1000 dgpu in a laptop) to keep it at a lower power state, without any performance differences versus the previous version. Maybe your 4090’s so fast, combined with the existing bottleneck (of something within Topaz AI causing 3070s/3080s/3090s/4090s to perform within similar speeds) that your card is ‘idle’ as there’s no way to push it frames fast enough
This topic was automatically closed after 25 minutes. New replies are no longer allowed.
