NAB 17 Codec Roundup
I only spent about 23 hours at NAB, but even in that short amount of time I was able to learn plenty about the latest developments in codecs and related delivery protocols. Here are some of the highlights.
I started my day with Haivision, which came out as major investor in HEVC codec x265, particularly for low-latency live encoding. Chief marketing officer Peter Maag explained that working with the open-source community allowed Haivision to offer its customers a flexible array of GPU/CPU-based encoding to meet their particular density, quality, and latency-related goals. In other words, an encoding one-stop shop.
Haivision also open sourced its SRT (secure reliable transport) technology, and formed the SRT Alliance with co-founder Wowza to foster the adoption of SRT. Briefly, SRT is a "video transport protocol video transport protocol that enables the delivery of high-quality and secure, low-latency video across the public internet." Basically, it's a competitor to Zixi, which is often used for first mile transmissions between live streaming producers and their cloud transcoders.
I asked Maag about the economics of open-sourcing a technology the company had developed internally. Beyond the expected flowery benefits to humanity and industry (and all that), Maag pointed out that anyone using Wowza for live event production would want an encoder with SRT, and Haivision just happened to have "an industry leading selection." All of a sudden open-sourcing made a whole lot more sense.
Next up was Harmonic, which along with VIVE Lifestyle Network and Globecast was demonstrating "the first ultra-high-definition (UHD) high dynamic range (HDR) OTT managed platform as a service (PaaS) for delivery of linear content over the internet to connected TVs." Of the trio, VIVE supplied content, Harmonic the live multi-channel HEVC/HDR encoding, and Globecast the distribution.
I had just finished a story on HDR video for the June issue of Streaming Media magazine and hadn't actually seen any until I entered the Harmonic booth. I have to say that it certainly lived up to expectations. Unlike 3D or even VR, in ten years it's certain that all TV will be HDR. It will be interesting to see how quickly these three companies can build a business case for video that only plays on a relatively small, but rapidly growing number of smart TVs, and requires 25Mbps to deliver the top quality 4K stream.
Video codec and optimization vendor Beamr had two announcements. The first related to real-time encoding support for the Beamr 5 HEVC software encoder on the Intel Xeon processor E3 family. According to Beamr, running on an Intel Xeon E3-1585 CPU v5 with integrated Intel Iris Pro graphics P580, the encoder can perform real-time software-based 4K encoding of 10-bit HEVC HDR video.
The second announcement related to the long expected marriage of Beamr's video optimization technology and Vanguard Video's HEVC codec technology into a content-adaptive HEVC encoder, Beamr 5x. Briefly, Beamr optimization is a post-process for H.264 and HEVC streams that iteratively reduces the data rate of the stream until the quality measure analyzer that drives the system signals that the target quality had been met. This schema slows operation and means that compression was applied twice to each file.
In Beamr 5x, the quality measure analyzer serves as the rate control mechanism for the HEVC encoder, so only a single layer of compression is applied. This schema lets the encoder target a specific quality level, rather than a data rate, making the process content-adaptive, so talking head videos will encode more efficiently than soccer matches. I look forward to testing 5x, though unfortunately it won't be available for the upcoming HEVC comparison at Streaming Media East.
Walking the upper level of the South Hall, I ran into Capella Systems, developer of the Cambria FTC Encoder, the first commercial encoder with a per-title encoding feature that adjusts the data rate of the encoding ladder up or down based upon the complexity of the source file. The latest release of the software includes Dolby Loudness Management, and Dolby E and Dolby Digital Plus encoding, and the ability to convert text on slates and in end credits to text for adding to the file metadata.
A few booths down was cloud encoding vendor Hybrik, which debuted segmented encoding, multiple flavors of HDR encoding, and, closer to my heart, post-encoding video quality measurements including SSIM, PSNR and VMAF. This, plus their rock bottom pricing convinced the Streaming Media editors to grant Hybrik a Streaming Media Best of NAB Award.
The key event for me was Netflix's "From Pitch to Play: Building the Netflix Global Studio," held Wednesday night at the House of Blues. As the name suggests, the presentations focused on how Netflix original content moved from pitch to production to play in what's called the Netflix Digital Supply Chain. As part of the event, five internal groups integral to the supply chain, which included Media Engineering & Partnerships, Quality Control, Studio Technologies, Content Platform Engineering, and of course Encoding Technologies, huddled around small tables and discussed pretty much anything you wanted to discuss. For me, it was a understanding how Netflix used and viewed the various codecs available to them.
I first spoke with Anne Aaron, director of video algorithms at Netflix. I had missed her excellent session with Google's Matt Frost the previous day, but she graciously covered a lot of the same material. (The session is available for viewing here.)
We first discussed VP9, which Anne indicated was at the heart of Netflix's low-bitrate encoding mobile push. Anne recounted extensive Netflix testing that found VP9 approximately 40 - 50% better than AVC in the same uses, and shared that even at the low bitrates targeted for mobile downloads, VP9 was 36% more efficient than H.264.
We then transitioned to the Alliance for Open Media's AV1 codec. In their tests, Netflix found AV1 about 20% more efficient than VP9, which was compelling at this stage of the development cycle. I then asked Anne how soon Netflix will start to utilize AV1 once the bitstream is frozen. She responded, "we had to integrate AV1 into our encoding pipeline to run our initial tests. We'll keep updating the code as it develops, and once playback is available with Chrome or other browsers, we'll be able to incorporate AV1 into our production workflows."
As an aside, in the above referenced video, Google's Frost claimed that AV1 was currently 30-35% more efficient than VP9 in their tests, and that he expects the AV1 bitstream to freeze by the end of 2017. He also shared that the Chrome team, and other browser developers, were currently experimenting with AV1 decode, and indicated that browser-based AV1 playback should be available when the bitstream freezes, if not before.
Then I asked where HEVC fit in. Aaron explained that Netflix currently deployed HEVC primarily on Smart TVs, and saw the codec as integral to its HDR strategy. For computers and mobile, however, H.264 and VP9 are Netflix's primary codecs, and the focus of most of its current research, which will soon include AV1.
I then chatted with David Ronca, director of encoding technologies. Discussing Netflix's current workflows, he mentioned that they weren't using the stock open-source VP9 encoder available from Google. Rather, they were encoding with technology from a company called Two Orioles, which was founded by Ronald Bultje, a lead VP9 developer. Ronca mentioned that in particular, Netflix had found that the open-source VP9's two-pass rate control mechanism was suboptimal, and that Bultje's encoder is currently about 8% more efficient than Google's.
Ronca also revealed that Netflix was using the Vanguard HEVC encoder (now part of Beamr), which it adopted after extensive competitive testing. I asked about the potential for Beamr 5x, which merged Beamr optimization with the HEVC codec Netflix was already using. He responded that though it looked compelling, Netflix had built its own codec-independent, per-title encoding workflow powered by its own rate control mechanism (Dynamic Optimization), and using VMAF to drive quality decisions. Though Netflix's approach is compute intensive, Ronca explained that the scale of its operation, and the number of times each video is played, means that it makes sense to invest whatever compute resources are necessary to produce the most efficient encode possible.
I asked about the future of HEVC given the imminent release of AV1, and the fact that Netflix was a founding member of the Alliance for Open Media. Ronca laughed, explaining that Netflix was still encoding in VC1 format for some older platforms, and that HEVC is important for their UHD/HDR experience, though he noted that the HEVC license uncertainty is an ongoing concern. However, as Aaron had mentioned, he shared that a major focus of Netflix was low bitrate transmissions to mobile, since this would be an enabler in many markets, and that HEVC didn't factor into those efforts. Transitioning over to HDR development, Ronca explained that all of Netflix's HDR Masters are Dolby Vision, which allows for a single workflow that can output Dolby Vision, HDR10, and rec. 709 video.
I ended my night at the Encore Tower, where V-Nova company demonstrated PERSEUS 2-encoded video at bitrates as low as 100Kbps for mobile consumption and 1080p HD video at 1Mbps. With partnerships with infrastructure players like Microsoft, Eutelsat, and Imagine Communications, V-Nova has made significant strides in making its codec efficient and easy to deploy; it will be interesting to see the pace of commercial adoption over the next twelve months.
Companies and Suppliers Mentioned