Since I have a Linux VM with a big GPU lying around I thought I’d take a quick stab at hardware accelerated video transcoding using ffmpeg. Quick results:
- software decode, h264 software encode: 0.355x real time
- software decode, h264_nvenc hardware encode: 2.7x real time
- cuvid hardware decode, h264_nvenc hardware encode: 2.8x real time
So the hardware h264 encoder is about 7x faster than software. Hardware decoding is nice but not a huge improvement.
This is all on a 1 vCPU, K80 Google Compute server. Crappy little CPU and big beefy GPU. Normal desktop machines probably have less of a spread in results.
I really wanted to test this with H265 encoding, but the hevc_nvenc encoder does not work with the Kepler-class K80 hardware. I’m regularly doing H264 -> H265 transcoding at home now and on my i7-2600 it transcodes 720p video at just about real time.
Between the video transcoding and the machine learning I’m itching to own my own Linux box with a GPU in it. I could add a GPU to the existing server but it seems to be $200 minimum. The old server is 6 years old, maybe it’s time for a full hardware upgrade. Go for SSDs. Ubuntu 17.10 while I’m at it; these LTS releases are reliable but limiting.
I wonder if there’s a way to use the on-chip Intel GPU more? Probably not worth the hassle.