Open jmunson opened 6 months ago
Closing the window (X) will leave fancontrol working, it doesn't close the process. There is still a tray icon active at the bottom right to bring the window back up.
i was just about to make an issue about this. it would be a really useful feature
Same problem for me as I'm on a very tight VRAM budget for gaming (two 4K screens plugged into a 8GB GPU) so any VRAM saving is welcome
Yes, the VRAM consumption of it is to high. It's the only big red flag for me with this tool. Always have to fully close FanControl when I use AI to be able to squeeze the models into VRAM. I'd rather have it eat 4x times the RAM than VRAM if I could choose.
Have seen it use up to 250 MB when in foreground and in background (I guess after some time when C# does garbage collection?) it usually drops to 70-100 MB.
Isn't there a way to make .NET render the UI without using GPU memory? Like many browsers or Discord have a switch to disable hardware accelerated rendering -> which usually drops VRAM consumption to or near zero.
Wish I could look a the source code now to get an idea of what's going on. Sometimes it's simple stuff like the UI loading some unnecessary large unoptimized pictures. Unfortunately FanControl is closed source. So I have to look for an alternative.
It's a WPF app so it uses direct X to render the graphics. Ui library used is https://github.com/MaterialDesignInXAML/MaterialDesignInXamlToolkit
If you start the app minimized and never bring the window up, it shouldn't take additional memory.
This is with the app starting minimized.
I noticed that if fan control is open it constantly reserves 100MB of dedicated GPU memory, even when the window is minimized. The only way to free up the memory is to close it entirely.
It would be nice if there were a way to entirely unload the UI while still retaining fan control, or at least a way to disable GPU acceleration of the UI so that this becomes normal memory usage and not gpu memory usage
It's a pretty minor thing but for users that want to work with LLMs especially on smaller devices, this can be the difference between fitting a model in vram or not.