Vouk's Render Performance Test

  • Thanks

    I'm using XL to do the computation from the log file, but I need to wait for the encoding to be completed which is time consuming.


  • This is the result I got with Premiere Pro CS6 (v. 6.0.5) on my main rig - 244 fps @ 1920x1080:

    https://i.imgur.com/HnGQAtM.png

    ____________________________________________


    I also tested a few different encodes of the same files to better examine Premiere's decode speed with VRPT. For that, I created 5 identical timelines each containing 5 identical video clips without audio, effects, or transitions, with identical resolutions and framerates. The only difference between these timelines was video codecs used to encode said clips.


    Original ProRes 444 LOG source footage (avg. 215 Mb/s-) - 48 fps:

    https://i.imgur.com/3sKeU9n.png


    The same footage encoded with HandBrake's Production Max preset (avg. 127 Mb/s) - 78 fps:

    https://i.imgur.com/kko8gDl.png


    The same footage encoded with HandBrake's Production Standard preset (avg. 55 Mb/s) - 140 fps:

    https://i.imgur.com/HzGRJNA.png


    The same footage encoded with HandBrake's Production Proxy 1080p preset (avg. 19 Mb/s) - 248 fps:

    https://i.imgur.com/1EMD11a.png


    The same footage encoded with a custom NVENC CQP18 HQ preset in HandBrake (avg. 5 Mb/s) - 356 fps:

    https://i.imgur.com/Zm1UqUQ.png


    Interestingly enough, despite the fact that ProRes and Production Proxy timelines were the fastest in terms of forward and backward live playback and playhead scrubbing, due to their I-frame-only nature, it was actually much faster to render an NVENC timeline, which was the only one among these timelines to actually contain all 3 frame types (I, P, B). You might think that this was due to disk throughput (the smaller the videos the faster the render), but all these videos were stored on a fast modern SSD with DRAM cache, and the disk load during rendering was fairly low.

    ____________________________________________


    All the above tests were run with Mercury Playback Engine GPU Accelerated (CUDA).

    I did try to disable CUDA, but, I didn't notice any significant speed change when rendering these effectless/transitionless timelines with CUDA disabled.


    Edited 7 times, last by dr0 ().

  • I wanted to see how different source formats compare in 4K. So, for that, I encoded a 1m40s 4K drone video into several different video formats and compared the decode speed of each video with VRPT in PrPro CS6.


    The results almost linearly scale if you compare the difference in fps between formats used in my previous comparison and the number of pixels per frame. The only discrepancy is that unlike with 1856x776, HandBrake Production Max turned out to be slower than ProRes in 3840x2160.