PPro, AME, and CUDA: What's Actually Happening Here?

Question

I bit the bullet and picked up a new Quadro 4K for my Mac Pro.  The videos I'll be making with Premiere Pro will include a scaled-down multi-cam PiP, which the CUDA processors can clearly help with.  However, I'm a bit puzzled by something: what's actually going on here behind the scenes, so-to-speak?  I'm coming at this question from a software/computer geek perspective, not from a video editing one.  I'm sincerely interested in what's going on between PPro, AME, and the nVidia card.
First off, I asked this question back in July.  At the end of the thread, Todd says that if I queue up the export with AME, then the rendering would be done by the CPU.  Fair enough, but something else is going on with AME and the nVidia card.  If I enable the GPU in PPro and then export it through AME, the AME log looks like this:
- Source File: /Volumes/Users/jvp/Library/Caches/TemporaryItems/Multi-Cam Test.prproj
- Output File: /Volumes/Users/jvp/Desktop/multi-cam_1.mp4
- Preset Used: Custom
- Video: NTSC, 1920x1080, 29.97 fps, De-interlaced
- Audio: AAC, 160 kbps, 48 kHz, Stereo
- Bitrate: VBR, 1 Pass, Target 6.00, Max 10.00 Mbps
- The source was deinterlaced
- Encoding Time: 00:20:09
If, however, I disable the GPU in PPro and export the same sequence through AME, the log reads:
- Output File: /Volumes/Users/jvp/Desktop/multi-cam-no-cuda.mp4
- Preset Used: Custom
- Video: NTSC, 1920x1080, 29.97 fps, De-interlaced
- Audio: AAC, 160 kbps, 48 kHz, Stereo
- Bitrate: VBR, 1 Pass, Target 6.00, Max 10.00 Mbps
- The source was deinterlaced
- Encoding Time: 01:18:32
A noteworthy difference in time.  Further, during the first export in AME, the GPU is definitely being used at about 33% or so.  So clearly it's doing something.  What I'm curious to know is: what?  Is it actually doing the rendering of the PiP while AME is exporting the video?  Or... ?
jas

Answer 1

Harm Millaard wrote:
Apart from other effects, the PIP means scaling and scaling is GPU assisted. That could well be a serious determinant for the differences in encoding times.
Right Harm, I'm certain that the PiP scaling is what's causing the immense difference in exporting (encoding) times.  But what I'm curious to know is: why is it mattering when I punt to AME?  To me, it implies that either:
PPro is continuing to do some of the work during the export, in the background, even though AME is open and working, or
AME is actually making use of the video card.
That's why I'm curious about this and wondering what's going on from a technical perspective.
jas