Achieved Better Performance When Lowering VRAM Usage in Preferences?

Topaz Video AI General
1

Just now when upscaling a video I noticed I was getting artifacting in a YouTube video i was watching at the same time. Originally I had VEAI set to GAIA High Quality and VRAM usage in Preferences set to max. The video was showing an eta of 1hr and 20 minutes. Since I had the artifacting on YouTube I decided to turn down the VRAM usage in VEAI and use the same GAIA High Quality preset. This time though it is showing a eta of 26mins. Why would a video take longer to process with VRAM at max but a shorter time to process with VRAM usage turned down?

1 Like
2

I have seen artifacting in videos I was watching on VLC while using TVAI - I had not noticed this behaviour until recently.

Interesting to note… I’ll try adjusting my VRAM usage and report back on results.

3

I noticed with Gaia in v3 that it showed 2 very different completion times for the same video and settings. You can test this by selecting Gaia upscale as default template and then closing/opening VAI each time, selecting the test video and pressing Export. If I do this like 5 or 6 times, 2 or 3 times it will take like 5 minutes and 2 or 3 times only 3 minutes. The output size between these 2 different processes will differ slightly. So there seems something wrong with the Gaia initialization

4

For starters, I think it’s almost always a very bad idea to let any program use all of your video memory. And precisely for the reason you mentioned yourself: you want someting left for youtube et al.

Why does it run slower with max VRAM? Hard to say without seeing their code, but I venture to think that the youtube artifacting you mentioned is reciprocal: aka, your VRAM needed for TVAI gets corrupted too, checksums fail, and TVAI needs to redo some chunks? Who knows?

5

When I run Gaia in V3 with only one video it loads GPU up to 100%.
When running Gaia with 2 video with the same settings in parallel, the GPU loads gets only up to 85%.
Same happens when any other program (like VLC) charges GPU.
Nevertheless Gaia processing is about 25% faster than Proteus, loading GPU only less than 40%.
GPU memory load is about 4GB per video (with 12 GB available in RTX3060) I don’t run more than 2 videos in parallel in order to have enough resource for other jobs.

1 Like