Netflix Finds VP9 Offers Strong Compression At 1080p, Approaches HEVC Performance

Netflix, being one of the biggest video streaming services in the world, tested how efficient various video codecs are for a given level of quality. The company discovered that the royalty-free VP9 codec developed primarily by Google is almost as efficient as HEVC, and can sometimes be even better at resolutions of 1080p and higher.

Netflix has over 83 million subscribers who watch billions of hours of TV shows and movies, according to the company. The service was made available in over 130 countries this year, which means it could potentially grow much larger in the future, once more people discover it and have the requisite internet access. More people are using it over cellular networks as well, which makes delivering the video streams as efficiently as possible a high priority for Netflix.

Testing h.264, HEVC, VP9

The company evaluated the three (currently) most popular video codecs: h.264/AVC (still the most popular one by far); the next-generation h.265/HEVC; and Google’s royalty-free competitor, VP9. The company used open source encoders for all three, including x264 for h.264 videos, x265 for HEVC videos, and libvpx for VP9 videos.

Both HEVC and VP9 promise about 50% bitrate savings for the same quality compared to h.264, but Netflix wanted to test for itself to see if this is true. Netflix sampled 5,000 12-second clips from its catalog, which includes a wide range of genres and signal characteristics.

With three codecs, two configurations, three resolutions (480p, 720p and 1080p), and eight quality levels per configuration-resolution pair, the company had more than 200 million encoded frames. Netflix applied six quality metrics: PSNR, PSNRMSE, SSIM, MS-SSIM, VIF and VMAF. This resulted in more than half a million bitrate-quality curves. Netflix’s unused cloud-based encoding infrastructure allowed the company to complete this large test in only a few weeks.

Test Results

The company learned that previous research showing up to 50% bitrate savings for both HEVC and VP9 compared to h.264 turned out to be true. HEVC’s x265 implementation outperformed VP9’s libvpx for most resolutions and quality metrics. However, at the 1080p resolution, the difference was either much smaller (in HEVC’s favor), or, in some cases, VP9 even beat HEVC in bitrate savings.

The fact that VP9 performs better at 1080p or higher is not a major surprise, considering VP9 was optimized for resolutions beyond HD. Google is currently using it for YouTube, where all videos are encoded in VP9.

Codec Adoption

VP9’s predecessor, the VP8 codec, failed to gain traction because it arrived much later than h.264, and because chip makers weren’t too interested in utilizing hardware accelerators for it. This has changed somewhat with VP9, as more and more chips come out with accelerated decoding, and even encoding, for VP9.

However, by the time either VP9 or HEVC become too ubiquitous, a new generation of royalty-free video codecs that are being standardized by the IETF may arrive to replace both. Google, as well as Mozilla, Cisco, Intel, Microsoft, and even Netflix itself are all contributing designs, code, and patents to create a new standard royalty-free codec that everyone can use.

Update, 8/30/16, 2:17pm PT: It's come to our attention that the original headline in this article was poorly worded to the point of being misleading. We have replaced it with a different headline.

This thread is closed for comments
14 comments
    Your comment
  • DocBones
    Wonder why they didnt do 4k testing.
  • I'm guessing 4k testing wasn't done because of lack of samples, time and resources (spare cloud processing). I really wouldn't be surprised if it was in the works. I'm more and more impressed by Netflix, seems like they are really improving.
  • bit_user
    We'll see how long it stays royalty-free. Patent trolls might find something it infringes. Compression is such a patent minefield that it's hard to do anything (good) with no risk of infringement.

    BTW, I don't know why they even bothered with PSNR or PSNRMSE. I hope those didn't skew the results.