I am desperately trying to use my gpu for 2 things at the same time, upscale using starlight precise 2.5 while i also play some video games, i’ve tested nearly everything but i cant get it to work. It seems if i set max memory to something like 10% it still ends up using 70% of vmem on the gpu, and the game i am playing cant fit on the remaining 30%. Is there any way to actually cap the vmem usage of topaz?
ive tinkered around in the game too (poe2) and i cannot make it consume less than 5-6gb of vmem, and it seems this is enough to put it over the edge with starlight running in the background ending up nearing the peak of 24gb, once i hit the limit you can tell, the games framerate goes from stable 100fps to 20fps and stuttering (unplayable) then i have to reset the game and reboot it, now the memory is cleared and well be able to play a little bit again, rinse repeat, it would be nice if i could just cap topaz to not consume so much, is it possible?
Edit: maybe i should add gpu info, i am using a nvidia 4090
Running a process with Starlight Precise is going to put a significant strain on your system, even with a max memory cap. It is not recommended to try to run other graphic-heavy processes or tasks when using Topaz Video, even if you are running the lighter models.
i understand, but maybe i should pivot to request a pause button for topaz video then, it would be really nice to pause the upscale process then use the gpu for something else, then be able to resume.
right now it seems like topaz video already has a resume function because if i kill the process and open it again it will essentially start from where it left off but its not a nice way of doing it, a pause/resume button would be a lot better.
The Starlight models are unable to be paused at this time with the current feature in the app, due to how the models run. This is on the list for the team to try and work out a way in the future to make this a reality.
Saying you can’t interrupt a video diffusion model is just a YouTube myth at this point.
Temporal inpainting is perfectly doable by freezing previous and future frames to guide the context. It’s obviously a massive KV cache and VRAM optimization challenge for any engine, but it’s definitely not a theoretical limitation.
Assuming that your CPU is capable of handling a Topaz app and a game at the same time, your best solution is probably a second GPU. Topaz apps allow you to specify a processing GPU. Not a gamer, so no idea if games also allow this, but you want the two programs processing on different GPUs.
