Streaming Media East: Netflix Making the Move to HEVC, but Efficiency Gains Lag
“We’re not seeing efficiency gains being claimed by HEVC encoding vendors,” said David Ronca, manager of encoding technology at Netflix.
Ronca made his comments during a session at the 2014 Streaming Media East show, being held today and tomorrow at the New York Hilton. The session, titled “Netflix's Video Workflow: Transcoding, Codec's and 4K Streaming,” focused on those three key areas.
Ronca noted that the Netflix approach of measuring "quality versus time" shows that high-effiency video coding (HEVC) yields about the same quality as x264 with much longer encode times—up to ten times longer.
Yet it appears that the limitations of HEVC today aren't deterring Netflix from making the move to HEVC. “In two years, we expect about a 20-30% efficiency versus x264,” he said.
“We are very excited about HEVC,” he said. “All of our innovation is going to be shifting to HEVC. We’re not pessimistic, but I’m just telling what we see today.”
Ronca also walked through the transcoding process, highlighting the need to be detailed in the reasons an uploaded mezzanine file asset might be rejected by Netflix.
“When we reject an upload, and send it back to a content provider, we have to give a detailed reason,” he said. “We can’t just give an FFMPEG error code.”
Ronca noted parts of the transcoding workflow includes a number of inspections, including both system layer inspections and decode inspections.
“We fully decode the source, with no errors allowed,” said Ronca.
The Netflix team then does a decoded picture analysis, including telecine detection. In addition, tests are performed for commercial black detection—allowing no more than two seconds of commercial black—and crop detection.
Netflix also does audio detection tests.
“Sometimes we’ll be told that a particular set of audio files is a left-right mix down, and yet we’ll find that it’s actually side-channel audio,” said Ronca, noting this would also be reported back to the content provider.
Ronca talked about the difference in the older encoding solution Netflix used, called Matrix, versus the new one called MAPLE, which stands for Massively Parallel Encoding.
MAPLE handles a video job by chunking across multiple Amazon EC2 instances. A progressive model is used to efficiently handle transient errors, eliminating the problem faced in a long encode where it may fail towards the end, requiring the entire asset to be re-transcoded.
To place these parallel encodes in proper sequential order, which are done in small chunks of video on many different machines, scene change detection information is gathered and then used to re-assemble the chunks.
“We have around five nines reliability,” said Ronca. “We’ve never had a false negative, meaning a bad encode has never made it out to our customers.”
Ronca then discussed MAPLE in action, with one example being a worst-case encode: a collection of short films by a famous cinematographer. The asset was 4.95 hours in total length and took 96 hours to encode 1080p in the older “Matrix” system, with over 900 compute hours completed. The high number of compute hours were based in large part on various transient errors.
MAPLE handled it in less than three hours.
Ronca then shifted to talking about UltraHD (4K) and preferred formats.
“One of the more difficult processes for us is working with 2TB files for UltraHD,” said Ronca, noting that a 2TB file being downloaded could take up to 24 hours, which causes issues around transient errors. In addition, most EC2 instances don’t have enough storage space to handle a 2TB file.
“The answer for us was chunked inspections,” said Ronca. “MAPLE allows us to do this.”
“We now have the ability to move 4K assets through our workflow almost as fast as 2K assets,” said Ronca, adding that the “almost” was based on the fact that high-efficiency video coding (HEVC) take a bit longer than advanced video coding (AVC).
Ronca said that UltraHD (4K) is interesting not just because of the resolution, but also because of the high frame rate (HFR) that provides a different viewing experience when content is viewed at 4K60 (sixty frames per second) versus the slower frame rate of 4K30.
“We also like the HDR and wider color gamut of UltraHD,” said Ronca.
Ronca also noted formats that Netflix supports, staring with MPEG-2 Transport Streams (M2TS). He said that M2TS is well supported but doesn’t allow for parallel inspections like those being done for the 4K assets, due to the fact that M2TS is a series of sequential packets that must be inspected in sequence.
ProRes is a very high-quality format also supported by Netflix, but Ronca notes that it’s a proprietary format that’s not heavily documented as a specification. A third format, DPX, is another that Ronca says is high quality, allowing parallel and progressive processing but that it is at almost 2TB per hour and that Amazon S3 has a 5TB object limit size.
“We don’t feel that DPX is practical for scale,” said Ronca.
IMF/MXF (based on JPEG 2000) is an emerging standard with the ability to deliver incremental revisions.
“We are very excited about IMF/MXF,” said Ronca, noting that February 2014 was the first time that Netflix deployed encodes based on IMF/MXF.
“IMF/MXF is the future,” said Ronca. “We’re investing heavily and have messaged down to our content providers that it is the preferential format, and there may eventually be a point where it is the required format.”
Watch the address below and download a PDF file of the presentation.
It's important to think ahead when preparing content for different screens in different environments. Read on for best practices from experts.