Fully understandable my friend! AI rendering use the gpu in a way standard benchmarks and stability tests don’t.
Had to do a lot of trial and error until I found voltage and clocks speed which cause power to hover around 95% +/- 4%. Too much voltage and it punishes clock speed heavily. Slightly too high clock or not enough voltage will cause a crash.
After hours of testing different input formats, resolutions and combination of models as well as a lot of crashes I’ve found the sweet spot.
My 3y old off brand HP rtx 3080 beat all 4090 scores I’ve seen(except for one, which also had considerably higher score than other 4090s which means that owner has also been min maxing his card) when it comes to all of the interpolation models.
For the enhancement/upscaling AIs the fabric 4090 render about 50% faster than my 3080. However, if you’d compare a fabric 3080 to a fabric 4090, the 4090 should be 88% faster if you only compare relative specifications which impact ai enhancement models, such as tensor cores.
At 100% load and 95-98% power usage my temps stay well within the range for minimal degradation. Overnight render won’t have core, hot spot or memory conjunction temps come within 10-15C of its limit, and that’s with a very conservative, but optimized fan curve to lower noise. They average about 50-55% to hold those temps for +12h.
Speaking of temperatures… What do you guys have on a 4090 under “full load”? Like processing a video with Rhea, which seems to be the most demanding model. (Rhea XL is still in its “baby phase” IMO, since I see worse results compared to the original Rhea). I have a liquid cooled machine and the temperatures get has high as 68-70 C? Is this normal/safe?
Another thing that I noticed about models, Rhea in particular because I didn’t really test thoroughly the other ones recently. If I use Rhea I get ~4.8fps processing speed. If I activate Focus Fix (even though I know that is not meant for 1080p or lower inputs), the speed is 3x faster (~15.5fps) and it also produces in a lot of cases better results.
So, how did your 4090 become faster than mine, then? I set it to us max 108% power, which is 485W (iirc, you could go up to 550W). And yours, at 95-98% power usage, is faster?
I also have mine set to ‘silent’ bios option. (I figured 1 or 2 exta percent in a game really doesn’t offset the rather stern increase in noise).
I have my RTX 4090 locked at max 85C (+1C). It rarely gets to that, though, as I have it power-limited (to 486W max); so mine will hit ‘Pwr’ (in GPU-Z) before it can get too hot.
Rhea is the model which has the highest VRAM usage by far, correct?
In general a +1000-2000mhz memory overclock is fine on a rtx 4090.
Overclocking memory is a lot less prone to crashes than core clock. It’s also not always better after a certain point. Can’t remember the exact reson, but say a 3000MHz memory overclock might throttle other gpu loads, generating frames or something along those lines, while 1500MHz can increase performance a lot.
I’m flirting with the ASUS TUF RTX 4070 Ti SUPER OC with 16GB Vram to replace my Asus 470 none TI (NVIDIA AD104) with 12GB Vram.
According to benchmarks, the TUF RTX 4070 Ti is at least 30% faster. What I also like is that the frame and cover is made of solid metal and it should be very quiet even under full load
Do these also perform better with TVAI? What do you think is worth or just a small step upwards?
Personally, I would never do this. For starters, +30% added to an already suboptimal card, is still not worth it. Get an RTX 4090 instead. And if you can afford it, get an RTX 5090 (any time soon now). The latter sits in another price bracket than a 4070, of course; but point being, if you want to buy a new card for TVAI, make sure it’s a substantial step up from what you have now.
Mind you, this is not about money per se; I just want a card that’s substantially better than the one I have, or I wouldn’t buy a new one. Like if they brought out a new RTX 4090 Ti, I would not buy it (difference being too minimal); the RTX 5090, however, I will buy.
Actually, When this came out in beta I actually asked about this, and are they going to allow users to host their GPUs for parallel processing, and Topaz devs said that the reason they do not allow this is security reasons. If they send us portions to transcode, we could theoretically view the material.
They did confirm they are indeed breaking the video into small parts and running it across their farm in parallel.
I thought it would have been a much better idea to make a cloud only version, FREE, so you pay for credits to render, or the two packages, Home user and Pro use. Home use is the app for $40/yr and $100/lifetime without the multi-card farm and license not allowing professional work done on that license, and professional is $299 with the ent portions included. Then you have three full different versions that would make more sense than throwing one large app that does it all and charge everyone ent prices.
I, again, cannot stress how much their pricing model is off and wrong for non-pro users. Will some cheap users actually use the home license and use it for pro work? Absolutely, but small loss with how many home users you pick up, who then get used to it and get work to buy it so they can use it there. THAT is how companies build up users bases.
Thanks.
