I did, 172 Episodes, duration 45min/episode median (at least, there are some double lenght episodes) costs about 696.600 credits. When 3000Cr costs 200$ total ammount for Season 1 -7 is → 46.440$ 
if it takes 15min to render 10s film, total rendering time (all episodes film lenght = 7.740min.) is 696.600 min. or 11610 hours = 484 days
I have houndreds of hours of TVAI with 800mv on my 3080 without instability. The card automatically uses even much lower voltages if there is little or no load. There is no inherent instability just because of low voltage. Only too high frequency for any given voltage will cause instability.
Here is what I used and it runs absolutely stable and very efficient:
Even 737mv runs absolutely fine and is ideal for even lower wattage:
@Mellow
Your voltages doesn’t seem to be from the 3000 series of cards (too high clock speeds for that), rather than the 4000 which are from a different chip manufactoring process. Also, using almost 1000mv will not result in a big efficency gain which is the main point of undervolting imho.
see post above
Has anyone compared other models against Starlight to see how much difference it makes?
It’s easy seeing the original with Starlink, but not Iris LQ at manual settings against the same footage but with Starlight.
In some of the comparisons I’ve seen, I feel IrisLQ could pull off similar results if done on manual with recover detail and improve detail set high, maybe not one attempt but two passes.
I did in my earlier post above vs Artemis LQ (only model that really was the best before at reducing aliasing/moire with my footage). I used Starlight with a Proteus second enhance pass with Dehalo at 60.
I’ll check it out, thank you. I find Artemis LQ-MQ gives a waxy image.
Try this: Make two encodes, one with Iris and second encode with Proteus OR Iris + Rhea (XL when aretfacts are not to strong). Do not use “Revert Compression” or “Dehalo” they are invasive, try first using Deblur only which keeps naturalness, maybe use additional Rec. Orig. Details and Sharpen but nothing more.
Then overlay frame exactly both videos in your video editing program and set opacitiy for example to 50%.
Import the merged video into TVAI: 1st model “Artemis Medium Halo” (does denoise+sharpen a little bit and merges overlay better together) + choose second model “Artemis HQ V12” (<- makes pixels tructure more homogenus and minimizes plastic look from Artemis Halo)
It is a 5080, like I wrote in the post. My previous card was a 3080 which i started to run TVAi using my previously (stable in everything, but far from stable in TVAi due to the nature of loads it put on your gpu which is not like synthetic benchmarks, stability tests or gaming) stable undervolt. After 10 or so crashes where I changed values up/down/left/right using the same approach as you. A flat down clock - > select voltage - > increase only that voltage - > apply was not stable.
I finally settled with standard curve and flattened the rest like in the example and that was stable.
Only because it’s stable for you doesn’t mean it’s stable for everyone.
The first example, which I removed was with the approach you use.
Yes, every chip is different. My recommendation of +100Mhz over default at those low voltages is conservative but ofc it could be that somebody with a below average chip experiences instability. It’s just not a problem of the voltage by itself (otherwise the default curve would be instable at that voltage as well) but a too high clock at that voltage.
Limiting to the card to a voltage from the standard curve should always be stable at any available voltage.
What type footage do you usually upscale? I usually use Artemis HQ or Proteus on already high quality 1080P to give me faux 4K or 4K to faux 8K.
But Artemis LQ has done the second best for me on low resolution lineskipped video (Starlight has done the best). But if the plastic and smoothness can be addressed, that would be great.
That was what I tried immediately. No better than any of the other combinations and/or models. Low res faces = no go. 
It’s a bummer if the faces are a no go. The type of footage I shoot mainly is telephoto wildlife slowmo so maybe that’s why it’s done the best for me personally on lineskipped aliased/moire video. I need to do more testing on faces tho.
See here for a comparison Starlight vs. IrisLQ V1 with tweaked settings (spoiler:Iris won by quite some margin)
I agree. The Iris LQ worked better here. Iris LQ has actually been a go to of mine for Sony A7S III footage. It’s strange how we’re getting different results in terms of color. Starlight is adding more contrast and saturating the color more in my shots, but you’ve been getting the opposite effect. What format file are your input videos? MP4?
False. Due to several factors.
A considerable undervolt, eg. 800mV even at fabric clock can 100% be less stable than the same thing but at a higher voltage, eg. 950mV. You’re also reducing stability by not maintaining the default curve up to the voltage you have set.
What you are claiming is simply not true.
Has anybody tried to rerun the same clip with Starlight? I redid a clip today that I previously did last week, and it seems to have alleviated some of the issues with oversharpening/overestoring/banding that I had initially. Looks like a good sign that the team is continuing trying to improve it.
Not surprising, but holy smokes!
In that case (setting the undervolt at the default curve), the clocks at lower voltages are actually lower than standard and therefore hardly make it more unstable.
Well, I enjoy my completely stable undervolts with nice efficiency gains since years (I also do overclocking in community competitions). You can enjoy what u are doing. Have a nice day, sir 
Frankly, I don’t see a lot of difference between the two. TVAI has evolved into miracolous video enhancement application and I’ve spent literally thousands of dollars on high performance hardware to run it on. - I find it strange that this Starlight announcement does not say anything about any improvements to the “latest” desktop version of TVAI I’m already running, And I’m not certain the Starlight is little more than an off-loading of local processing only to cost money and likely save time. - That’s not an attractive option for me personally. - I hope this new “innovation” doesn’t arrive at the cost of de-emphasis future improvments to the quality of the desktop version local image processing.
Well, for the example given I do.
But that is a not so hard case where you could likely have had good success with some of the old offline models, too.
