bobobo1618
commented
7 years ago For fun and a ballpark idea of where Turing fits, I ran some basic tests with ffmpeg (and the patch currently on the mailing list and libturing build from Git master). I used a 7.3Mb/s, 59.96 fps, 10-bit (yuv422p10le), 1080p HEVC file I had and looked at how much time it'd take to get to approximately the same PSNR (~49.5) in a reasonable (less than 10 min and or as fast as possible) and what bitrate the resulting file would have. The machine has a relatively weak i3-4150 CPU (which also lacks a hardware HEVC decoder, which is a problem).
time ffmpeg -t 20 -i input.mov -c:v libx264 -psnr -tune psnr -preset medium -crf 22 -an -y /tmp/out.mkvconsumed 133s of CPU time and resulted in a 671kb/s file and a global PSNR of 49.573time ffmpeg -t 20 -i input.mov -c:v libx265 -psnr -tune psnr -preset medium -crf 20 -an -y /tmp/out_hevc.mkvconsumed 422s of CPU time and resulted in a 513kb/s file with a global PSNR of 49.295time ffmpeg -t 20 -i input.mov -c:v libturing -turing-params 'speed=fast:qp=20' -an -y /tmp/out_turing.mkvconsumed 4428s of CPU time and resulted in a 868kb/s file with a global PSNR of 49.582
Given that Turing Codec is speed focused this is pretty surprising to me. I also tried with a raw yuv file and the turing executable but it was even slower. I suspect that might be because disk.
There are claims made about "high perceptual qualities" and "good compression performance".
Has there been any effort to quantify Turing Codec's compression performance? Could the results be made available? Is it likely that Turing Codec could be evaluated by MSU in their regular encoder comparisons or added to the AreWeCompressedYet benchmarks?