Anyone try starlight mini with PRO 6000?How does it perform compared to the RTX 5090?
We know that performance will decrease when using system memory due to insufficient VRAM. Will it consume more than 32 GB of VRAM, or is it just a RTX 5090 with most of its VRAM idle?
In some unofficial testing we did see a report of a .1 to .3 frame per second improvement with the Pro 6000 over a RTX 5090 with the same video sample and upscale settings running Starlight Mini and Sharp.
This could cut about 5-7 minutes off an export when doing larger resolution files with a 10 second clip, so instead of 37 minutes the export came out in 31 minutes.
My question is, does SLM assign more than 32GB Vram to a Pro 6000? What I observe is, with my 5090 the 32GB Vram gets not fully used.
That’s also my concern. Is the process currently cropping the image into multiple blocks, processing them, and then stitching them back together? Using a large VRAM could eliminate this step.
I have an RTX 6000 Pro Blackwell card and I have not seen it go above ~30gb of VRAM no matter what I’ve thrown at it, including Starlight Mini 2x upscale.
Starlight absolutely hammers the card more than anything I’ve run thus far. Temps easily hit 80-85c with the fans running at full speed. It’s such a concern I’ve actually nerfed the TDP to 85% to hopefully avoid a melted power adapter.
Because of this, ironically, rather than sell my old RTX 4090, I’m getting an eGPU enclosure for it in order to run overnight / multi-day Starlight upscales.
For those curious, I’m running a 272x480p → 1280x2238p upscale (i.e. the minimum) with Starlight Mini now. 27gb of VRAM used and around 0.4fps with TDP set to 100%. ETA is about 20 hours for a 16 minute video…
In that case, for Starlight Mini, there is not much difference between Pro 6000 and 5090.
That wasn’t the case with the introduction of local Starlight! Now I’m sure they’ve nerfed the local version, because I can’t explain it any other waysomething like that doesn’t happen by accident. Take a look here, the 6000pro as it could still use the VRAM of the card Video AI 7.0 - NEW Starlight Mini (Local) AI Model - #772 by olivier.martin
What you can do is using the old Version (V7, V7.0.2.b, V7.0.0.4.b) So instead “improving project starlight” they made local Starlight worser with highend cards.
I found the performance of Starlight on the RTX Pro 6000 a bit underwhelming for sure. Interestingly, it’s only Starlight Sharp that absolutely cooks the CPU - GPU utilization is 100% non-stop. But Starlight Mini drops to 10-20% utilization every 4-5 seconds, which keeps the temps in check. But I’ve got no idea why it does that, when there’s oodles of VRAM and system RAM not being used.
+1 to the nerf theory?
To my opinion, Starlight (Mini) alternates between a very intensive GPU AI phase (saturated VRAM) and a CPU/RAM phase of reconstruction and writing the result. The ~0% GPU usage drops are normal: the GPU waits while the CPU assembles and writes the frames.
Being owner of Pro license, I tried every possible GPU combination. The results are more than disapointing, especially on the latest models, both in multi-filesand split processing.
The pipeline is “so well” designed that you absolutely must have the display connected to the GPU being used, otherwise there will be a conflict between CUDA and the pipeline management.
Do not expect more explanations from the team, they never deliver such.
Does anyone know the performance difference for a 4090 vs 5090 with Gigapixel Wonder / Redefine subtle and Standard Max?
Synthetic benchmark the 5090 is 20-30% faster than 4090. But I’m not sure how good Gigapixel scales, maybe benefit is lower. What I really love about the 5090 is that 850mV GPU undervolting (1) this turns the 5090 into an efficiency monster. I use Topaz Video and I really have (measured) only about 550W (600W max) total PC power consumption while Starlight renderings (undervolted 5090 takes about 350W).
So if this is important to you and you want to buy new card, and if you have the choice between the 4090 and the 5090, then go for the 5090. If you already have a 4090, it may not be worth switching, or the question is what you can still get for the 4090 and how much the switch to 5090 will cost.
(1) MSI Afterburner, RTX5090 @850mV limited; gpu clock of 2200Mhz & ram clock of 2100Mhz, power consumption drop about 200W, speed loss about 8%
The powerload of the 4090 drops when i use Photo and Gigapixel at the same time with Standard Max and Wonder.
To me it means the gpu is fully saturated and isn’t able to deliver power to the full board.
With afterburner i’m able to turn the powertarget to 115%.
Maybe the 5090 would do better.
5090 has higher memory bandwith could be this helps.
I use 7.0.0.2b for all my Starlight Mini jobs…. it’s the only one that uses up to 45GB of my 48GB modded 4090. Every other version stops at aroung 22GB
yes I know, I use 7.0.0.4.b does also use full vram (and goes over it by adding shared), same as 7.0.0.2b does, I think this two versions are the only one without gpu ram limits. At the moment I do 720p SLM 3x (to 4k) upscale and I have gpu ram usage between 48GB-49GB (31.5GB Vram + rest shared)..speed? don’t ask! (0.1 fps or less 
)
Are you doing this for speed or is the output better?
for quality, some said there is no difference, but i can see the better quality. it will be resolved more finely. With low gpu ram, when you zoom in, you can see blocking, and pixel cluster gets smoother with more gpu ram. There’s a reason why the diffusion model wants to draw more RAM, and if you don’t give it to it, there are consequences. I also hardly ever see tiling when I use the beta, or at least the transitions are so good that you can’t see it.
As higher upscale resolution, as more Vram the model would take. But in live version, they have simply limited Vram usage of the cards. I only see about 22GB usage with my 5090.
Regarding the eGPU idea, apparently they’re not supported by Topaz Video:
I’ve no idea if that means eGPUs definitely won’t work, or it might work (but isn’t supported).
Andy
below is a sample while doing 720p to 4k with SLM and TVAI Beta with actually 43,6GB GPU Ram usage 
so yes it’s horrible slow with that ammount of shared ram, but compare source with upscale, everything is smooth and this leads to fine details (check hair) can be displayed. It is heavy zoomed in to show the effect better.
When I do the same, 720 to 4k with my 4070Ti and only 12GB Vram, it’s like night and day, when you have a close look it’s rougher. Sure when you compare a 480 source upscaled to only 960, then the difference is much smaller, beause less Vram is needed, then the gap is smaller
A = Source, B = SLM @ 44GB gpu RAM
https://imgsli.com/NDQ0NDc4
