GPU slower than CPU. Both slower than in last version

[udkim1]

Hi.
I'm trying to figure out if there is something wrong with my setup, or if it is normal that my GPU is over 2x slower than my CPU when processing a video face swap. GPU is a GTX1080, and CPU ryzen 1700. When forcing CPU, task manager shows CPU at 100%, as i imagine it should, and when running the GPU, task manager shows CUDA at 100% with some drops. based on this i imagine it was installed correctly, since it does use CUDA.
Also, in the last version (with the old UI), Processing was way faster, is there some setting that was changed that might affect it?

[ShrekLovesAll]

Are you rendering using CUDA or CPU? This is different from checking "force cpu for face analysis". The GPU is taxed far more when you're enhancing the face as well.

[udkim1]

Yes, i checked using task manager, that i indeed am using CPU when testing CPU and CUDA when testing CUDA. I need to try without face enhancment next.

[udkim1]

With no face enhancment, GPU was fairly quick, any way to speed up the face enhancement part?

[wesleyliwewa]

Hi, I have a similar situation too. My card is 3060 with 12GB, CPU i7-6700K with 48GB RAM, but when swapping with DMDNET Enhancing, for a 2.5 minutes, it took around 17000 seconds (yup, almost 5 hours), it seems like not normal, right?

I tried to update and restart but no difference, using 2.8.2 version by the way.

I checked, the CPU and GPU only using around 10% to 20%, adjusted many time, but still the same.

Can any one please tell me how to set it up? Thank you.

P.S. If I set None on "Select post-processing", so no enhancer, it will much faster but my CPU is 100%, and GPU only about 1 to 5 %. Then if I select DMDNET again, it will much slow down to like 3 seconds for 1 frame, and CPU and GPU are both about 10% to 20%. And in any situation, my GPU never reach 100%.

[C0untFloyd]

Enhancers are really slow and currently codeformer & dmdnet can't be run in parallel, so there is a bottleneck.
For the plain processing, it's necessary to have a look at vram usage. The default of 8 threads like in older versions is just too much for most people as this will fill up your gpu memory in no time. When this is happening the memory will be shuffled around between cpu and gpu all the time and that's why it's so slow and your cpu is doing a lot of work. You could drastically start with just 1-2 threads and a small buffersize to check it out and look after your dedicated gpu ram in taskmanager. Changing the number of threads/buffersize doesn't need a restart. Also be aware that the allocated memory isn't completely emptied if you're re-running from the same terminal, this seems to be a bug in the onnxruntime on windows.

[wesleyliwewa]

Hi C0untFloyd.

Thank you very much for your reply. I have adjusted to 4 thread and 4 frame_buffer_size, it is much faster now. A 8 seconds of 1920x1080 video in 29.97 frames/s can be done in 29 seconds with no Enhancers, and about 60 seconds for Enhancers.

However, it still take up the CPU first (close to 100%), and the GPU just around 5% to 20%, is this normal?

Thank you again.

[Kurtista]

Hello @wesleyliwewa

Just to make sure you're getting the right reading, assuming Windows, be sure in Task Manager under GPU to switch a graph to CUDA. Default view under performance tab, then GPU will have a top right graph labeled 'Copy'. Click on that label then select 'Cuda'.

[lysxelapsed]

Hi C0untFloyd.

Thank you very much for your reply. I have adjusted to 4 thread and 4 frame_buffer_size, it is much faster now. A 8 seconds of 1920x1080 video in 29.97 frames/s can be done in 29 seconds with no Enhancers, and about 60 seconds for Enhancers.

However, it still take up the CPU first (close to 100%), and the GPU just around 5% to 20%, is this normal?

Thank you again.

3060 here with 12GB too. with enhancers that's too many threads in my opinion. I'm running 4 threads with 4 frames buffer size with no enhancers. reducing to 2 threads / 2 frames buffer size with enhancer. check your VRAM usage with "gpu-z" (free and lightweight tool, windows task manager often doesn't get it right), if VRAM usage hits 100%, start again and reduce the threads. or even better: start with 1 thread and increase one by one. as soon as you hit 100% VRAM usage, performance will drop drastically.