very, very interesting/important, thanks! @Topaz-Employees did you note this?
I prefer the 5090 sure it’s too expensive, if you refer to the performance, but I still don’t regret the purchase. Of the four NVidia cards I’ve run with TVAI in the last few months, the 5090 runs most reliably super stable undervolted at 850mV the card is very efficient; a TDP drop of over 30% is remarkable and power consumption on the 12V connector keeps always below 400W.
1 Like
Once SLM or SLS is running, open a terminal with cmd and enter the following code and share the results:
nvidia-smi
I can try this 
but later I’m at work. I’ve posted something about 5090 undervolting months ago here
Nice!
I use OCCT for temps and power monitoring.
I purchased both Video AI and Photo AI; the apps live on my Mac and now so does the studio versions; can I delete the others to free up the space or do i need both? Example: I have Topaz Photo and Topaz Photo AI; Photo is 16.11GB and Photo AI is 22GB. do i need both?
Any update to this as I have the same issue.
I’m certain you can delete Video AI and Photo AI to free up space. The Studio versions are designed to be independent from the previous AI versions.
Andy
This is EXTRMELY useful. I’m going to look into all of what you said.
Thank you very much!
Last night, I also looked back through many user forum posts to gather and organize what people were anecdotally experiencing in SLm regarding VRAM/DRAM with various hardware and 0-100% mem slider settings.
For those with a single NVIDIA GPU (which also drives the display) that has >= 16GB of VRAM, keeping the slider at 70-74%, with CUDA sysmem fallback policy = driver (in NVIDIA Control Panel) worked best most of the time.
If the GPU VRAM is <16GB, I think you have no choice but to fallback to DRAM. A 12GB is borderline if the job is minimal. So, in the case of <16GB, you may as well set it to 85-90% (depending on what you are running/doing while SLm is running on the card, i.e. set to 85% (or less) if, for example, you are going to have Davinci, Photoshop in use or a high res video playing.
It was also pretty clear that there are 2 quality steps that occur near these VRAM usages:
11-12 GB and 23-24 GB. It seems you must give SLm a minimum of 12GB (from all sources) to get good quality and over >=24GB is excellent quality. Less than 12GB is noticeably much worse.
user @Mayday previously wrote:
seems to be, but not for sure
8GB - 11.6GB ->quality 1
11.7GB (not sure) → 16GB ->quality 2
24GB → quality 3
I agree.
I recommend looking at @Mayday’s posts along with @jojje’s, @robertsgonsalves1 among others…
In all cases, it seems that setting the slider = 100% can/will cause some issues at some point. The issues that come with 100% are quite varied, but common enough that it seems 98-99% would almost always be better than 100% (if trying to push the limits).
In my case, where there is one 32GB GPU dedicated to TVAI and a second GPU driving the display, my testing last night revealed a much better setup for this SLm job that was estimated at 1 day 4 hours.
As I mentioned, I thought the job would be about 8 hours (1 fps), but, once launched, TVAI said 1 day 4 hours (0.3 fps).
It turned that by doing the following I was able to get ~1.3fps (I say 1.3 because, at first it was 4fps, then dropped to 2fps, then it became mostly 1.2fps with some 1.4fps readings).
-
I set memory slider to 98% (100% had minor glitches in getting the run kicked off expeditiously)
-
In NVIDIA control panel I set the program C:\Users\All Users\Topaz Labs LLC\Topaz Video\models\runner.exe CUDA sysmem fallback policy: prefer no sysmem fallback, assigned it to RTX 5090, no vsync
2a) I assigned the same settings (no fallback) to the TVAI ffmeg file: C:\Program Files\Topaz Labs LLC\Topaz Video\ffmpeg.exe
-
In BIOS I disabled all features related to Virtualization/Hypervisor including IOMMU. In Windows, all Virtualization/Hypervisor settings were disabled. I left Resizable BAR enabled.
-
I re-rendered the file to be horizontally landscape oriented. It was a “tall” resolution (480x640)
-
I cropped the video slightly to become 640x400
-
I set the upscale to 2x
When I ran this, the VRAM usage was still well below the 32GB, but the card was much more consistently utilized with a result that was 4x faster than the “default” settings.
With that much performance uplift available - why wouldn’t the Topaz team be laser focused on providing much more sophisticated controls and guidance in this area??
For starters, how about some some or all of the following (which should be very easy to implement):
- Give the user a button to provide a pre-run estimate that shows element such as:
Estimated time & fps
Estimated memory usage profile for current slider & driver settings and installed hardware:
Peak VRAM usage: xGB
Peak DRAM usage: xGB
Minimum memory required
Peak memory usage if unlimited memory is available; output quality level = aaa; estimated time & fps
Peak memory usage if xGB is available; output quality level = a; estimated time & fps
Peak memory usage if yGB is available; output quality level = b; estimated time & fps
Peak memory usage if zGB is available; output quality level = c; estimated time & fps
Optionally, you could also offer estimates for the various (primary) upscale choices whether they be whole integer factors like 2x, 3x, 4x or scaling to a standard size like HD, 1280x960, FHD, UHD etc.
Perhaps, you could even indicate what the limiting factors are in each scenario (in order of the bottlenecking vs. a high-end standard benchmark) (GPU cores/generation, VRAM, swapping, PCIe Gen/lanes, CPU, DRAM speed, BIOS feature etc.)
-
The next step would be to allow the user to select one of these choices to implement. Further refinements per @naxci1 recommendations could be embedded behind some of these quality/speed choices.
-
While the job is running, display more statistics such as:
VRAM usage
DRAM usage
GPU utilization
CPU utilization
PCIe throughput
with a dynamic highlight showing which element is the momentary bottleneck
I’m sure there are plenty of experienced users at this point to get some excellent, crowdsource ideas into the product. The ones I put here are just on-the-fly…
Let’s get to it!
7 Likes
Any tips for adjusting VRAM if I have a RTX-3090 (not Ti)? I plan on use Remote Desktop to interface with the machine… do you think that will use my CPU’s iGPU or will the 3090 be wasting some resources by rendering the remote session?
Bug : Not sure if it is the same for others, but Topaz Video 1.0 doesn’t work offline. it ask me to login (while i already did that and it work online), while Photo and Gigapixel run ok offline.
I see in folder
C:\Users\All Users\Topaz Labs LLC\Topaz Video\models
are json files associated with all the models (and their major sub-configurations)
These also seem to have the settings you’re talking about.
And, they also seem to be associated with specific binary/encoded model files that are the large files that get downloaded when a run begins (if not already present).
I wonder if changing any setting in the json file would “break” the run in a bad way causing it to fail (because bin/enc file expects json to have certain settings). Or, I wonder, if changing the json would do what we want and the binary model file is agnostic to the json changes…
It seems like maybe the json could be modified and still work. This is obviously worth testing.
The SL/Astra json files all have the “slm” prefix.
Let’s see how it goes with ChatGPT recommendations!
If I understand… your setup is:
You have a dedicated pc with a 3090 for TVAI usage (only).
You will remote into that pc to setup Topaz jobs and kick them off.
Periodically you will remote into that pc to check progress and other “light” usage.
It is unclear if (1) you will be driving a local display screen at the remote pc and/or (2) you would be able to assign the CPU’s iGPU (or other, non-3090 GPU) to drive that remote pc display
The answer to 1 & 2 will matter with regard to settings.
If the remote 3090 pc has no display to drive that helps a lot and you can mimic my settings I just described (98% memory slider, CUDA no fallback; note this could still fail if your job is very heavy resolution-wise)
But, if the 3090 has to drive a display… then set the remote resolution to super low 640x480, color bit depth as low as possible and refresh rate super low - like 30Hz. Avoid having anything use the remote GPU. Note that when you remote into that machine, the remote GPU will have to do some “light” work, but your local GPU will also do some work to make the remote session function. All of it is “light” work but since your remote resolution, color depth is low it will be very “light” VRAM usage on the remote. I suggest remoting in and shutting down all unneeded apps and looking at VRAM usage in HWiNFO to get a baseline usage need for the system… that amount will be off-limits for TVAI. I would guess you could get that usage down to around 1GB)
Then you could set slider to 88%-90% for jobs that are not super “heavy” SLM jobs (meaning the source resolution/upscale resolution VRAM requirements are below 17GB-18GB) as well as set NVIDIA driver sysmem fallback = prefer no fallback. You might have to kick off the run to see what’s going to happen and then modify your settings based on that peek.
But, if you know you have a heavy job, you may need to keep sysmem falback on but set slider to 50% = 12GB, which is where quality level 1 begins and see what happens.
Ideally you would want a job to use 16-23GB of VRAM and never have to swap to DRAM. That way you get speed and quality level 2.
I would also consider turning off all virtualization/hypervisor features and IOMMU in BIOS and Windows. Also, make sure you have full PCIe Gen#/lanes set in case you are going to be swapping to DRAM (this may not matter though… the biggest Tx/Rx burst I have seen appear to easily fit in Gen4 x8…)
I am trying to learn all this on-the-fly, like everyone else.
So, seek more advice and test… and beg Topaz for some better guidance, mem control, job estimation stats, and real-time stats with bottleneck indicators… etc.
The application no longer closes properly after processing a video and attempting to exit, as if there is still something running in the background. Screenshot attached.
3 Likes
I could run any local display via the iGPU output, if needed. I wasn’t sure if the iGPU or dGPU is what renders the remote session.
But here’s a more pressing question ha… How much of an upgrade would a 5090 be, purely for local Starlight rendering? I keep trying to stop myself from pulling the trigger on the $1999 Walmart PNY OC 5090.
I don’t do any gaming. The GPU would strictly be for Video AI/Studio. If I can use the iGPU and dedicate the full 24GB of VRAM to rendering, is that going to get me at least 50% of the way to a 5090? Is 16GB of system RAM going to be good enough, regardless?
I did some poking around online, the 5090 might be 2X as fast as 3090? But we’re talking 1fps vs. 0.5fps. I guess if something takes 30hrs vs. 15hrs, it’s maybe not worth the $2K because we’re talking a ‘set and forget’ type project, regardless.
So my studio transition is complete yet I’m still unable to use the starlight models. Can anyone from Topaz team please help!
My PC has 8GB VRAM, so I know that starlight sharp won’t run locally, but now even the original starlight isn’t running. This wasn’t the case on the legacy Video AI software. And everytime I relaunch the app and select starlight, it does the download again and again. This is a crippling bug.
And on top of that if send my video for cloud render, it returns the exact same degraded video with no enhancements. This is the case for both Starlight and Starlight Sharp!! I’ve wasted so many video credits because of this.
The only reason I even considered the Studio upgrade was because of starlight sharp and if that’s not gonna work, then there’s no point.
Same problem here. Did you find a fix?
Just that you don’t want the precision reduced as this will give better speed at the cost of quality..
Thank you for clarifying, I had the same question. Intel Macs in recent years came with AMD GPUs, some of them very powerful (ie 2019 Mac Pro).
“Starlight… now runs on Mac and Windows systems with AMD GPUs” definitely sounds like you are talking about those Intel Macs with powerful AMD GPUs. You at least need a comma between “Mac and Windows”.
“Starlight now runs on Apple Silicon Macs, and on Windows systems with AMD or NVIDIA GPUs” would be a clearer way to communicate about this in the future.
Glad to see Apple Silicon GPU support so I can make use of my investment in M1 Ultra hardware, thanks!
3 Likes
My advice would be to keep the (perpetually licensed) older apps installed, because those purchased apps will continue to work forever (as long as the OS is compatible).
The new Studio versions of the apps will stop working the moment your subscription ends, if you ever stop paying for it.
That was not the case for all prior versions of Topaz software until now. You could always keep using the versions you had previously paid for. Now with this Studio subscription model, if your subscription ever becomes inactive you will lose access to any Studio versions of apps.
