RoncaTech · Video, engineering & leadership consulting
David Ronca built the Netflix encoding platform and led the Meta team that delivered high-quality video to 3.5 billion people. RoncaTech brings that experience to your codecs, pipelines, and teams.
Career
From DVD authoring to adaptive streaming, VMAF, convex hull encoding, custom video silicon and AV1 — a few of the stories behind the work.
David was a key engineering contributor at InterVideo from 2000 to 2007, working on broadcast technology (ATSC) and DVD editing and authoring.
DRM work that David did on the InterVideo DVR system was the inspiration for the PS3 Blu-ray implementation at Netflix.
Shortly after David joined Netflix, work started on the Silverlight client, the first widely deployed and used adaptive-bitrate (ABR) streaming client.
The streams needed for ABR streaming required special processing, and David wrote the back-end tools to prepare these streams for client delivery.
This was the birth of the Netflix Encoding Technology team.
We were challenged to develop an adaptive streaming client for the PS3. At the time, Sony was not working with Netflix, so the only way to build such a client was as a Blu-ray Java app. The problem is that Blu-ray plays muxed streams, so there was no obvious means of handling ABR. David proposed a model for Blu-ray-compliant ABR with client-side muxing, and built the back-end tools. Plot spoiler: we shipped the working app in nine months.
At one point, Sony messaged us saying, “The system was not intended to be used like this. Please stop!”
In January 2010, Apple announced the iPad. Reed emailed Steve Jobs and told him that the iPad was a cool product and Netflix would be a great app. Then we all waited…
In early February, Steve replied that Netflix “would be a great app for the iPad and we want you to launch with us!”
The only problem was that launch was 57 days out, and the iPad used HLS 1.0, a streaming model foreign to the Netflix system: specifically, (1) M2TS vs. fragmented MP4, (2) muxed A/V vs. separate streams, and (3) per-segment files vs. addressable segments in one file.
In the early years of Netflix, video encoding was very slow, often taking more than three days to encode a 1080p stream for a given title. Netflix also used a static ABR ladder for all titles.
In 2010, David proposed MAPLE (MAssively ParalleL Encoding), using a complexity-analyzer model to make encoding decisions. David invited Dr. Ioannis Katsavounidis to join him on the project for the summer. Working with David, Ioannis developed the complexity analyzer that altered the trajectory of streaming video.
US Patent 20120147958 (Ronca, Katsavounidis, et al.) covers the fundamentals of map-reduce encoding, complexity analysis, and complexity-based encoding — the basis for the Netflix complexity-based encoding model.
While working on the complexity analyzer in mid-2010, Dr. Ioannis Katsavounidis was perplexed by an apparent bug in his tool: it was recommending a ~1 Mbps H.264 bitrate for a fairly complex 1080p video. After much analysis, Ioannis realized that the test video was not actual 1080p, but 480p video that had been upsampled.
It was then we realized that our fixed bitrate ladders were likely over-allocating bits for many of the videos we delivered. Ioannis and David proposed a complexity-based model for encoding, where bitrate would be tuned to the underlying video. This was part of US Patent 20120147958, filed in mid-2010.
In late 2014, Netflix decided to launch the service in Japan in September 2015. Japanese subtitles require complex typography, and none of the Netflix devices had the technology to support it properly. To meet the tight launch window, David proposed an image-subtitle delivery model, where the Japanese subtitle assets were rendered to images for client-side display. This still required back-end tools to render Japanese subtitles properly; David designed the workflow model and led development of those tools. When Netflix Japan launched in 2015, its subtitles were recognized as the best Japanese implementation to date.
While working on the project, David saw that the global subtitle standards were incomplete and did not support the major Japanese features. He led the multi-company effort to complete the specifications, resulting in the publication of IMSC 1.1 and TTML 2.0. Netflix was awarded a technical Emmy for the global subtitle standardization effort.
In 2015, David worked closely with Rich Gerber and Dolby to build the Netflix HDR ingest and processing workflow. This included helping Dolby build an SDK that could support Netflix’s distributed processing model.
The Dolby Vision work, along with the IMF workflow built by Andy Schuler and Rohit Puri (leaders on David’s team), enabled ingest and processing of multi-trim-pass IMF HDR masters.
“We are video experts, and we are telling you that this is good video quality!”
Those were the exact words a codec vendor said to us when we complained about the quality of their video technology. David had long been telling anyone who would listen that the industry needed a new video quality metric. One day he was having lunch with Raul Diaz, a former colleague, explaining the metric problem. Raul suggested he talk to USC Prof. Jay Kuo (a former colleague of both David and Raul). After some conversations with Prof. Kuo, David wrote a research grant with provisions that the work would be royalty-free and open source.
Dr. Joe Yuchieh Lin (a PhD student at the time) led the project, and USC built a functional VMAF prototype in MATLAB. At Netflix, David’s team productized VMAF and brought Prof. Kuo’s and Dr. Lin’s great work to the industry.
In 2017, Dr. Ioannis Katsavounidis proposed the Dynamic Optimizer model for encoding video. The model made sample encodes at various resolutions and CRF values, then plotted the convex hull of the rate–distortion curves to predict the optimal ABR ladder. While computationally expensive, Dynamic Optimizer (convex hull) compression produced significant bitrate reductions, allowing good streaming of complex videos at bitrates as low as ~100 kbps. Ioannis’s work on convex hull encoding (US Patent 20190028529) has had a significant impact on the quality and egress of streaming video.
At Netflix, David’s role in convex hull encoding was to immediately see its value and ensure the project had the full support of Netflix leadership and top priority. When David joined Meta, he led the team that scaled convex hull encoding to 3.5 billion users around the world.
David has been working with AOM since 2015. As an engineering leader at Netflix and Meta, he sponsored open-source AV1 projects such as the SVT-AV1 encoder and the dav1d decoder. David has also been a tireless AV1 evangelist, discussing the benefits of royalty-free codecs and AV1 publicly and with network operators, handset makers, and SoC vendors.
As an engineering leader at Meta, David led the build-out of the AV1 ecosystem. As an IC, he focused on expanding AV1 reach through device testing, and worked on a level-based decoding model for software AV1 decoding. Most recently, David has worked on device benchmarking for software decoding with the development of the VCAT open-source benchmarking tool.
In 2019, Dr. Ioannis Katsavounidis shared his work on a cost/benefit model for encoder comparison. It laid the foundation for apples-to-apples comparison of generational video codecs, and finally provided an objective way to measure the value of an encoder implementation. Ioannis’s model provided the feedback loop that enabled SVT-AV1 to achieve impressive cost/benefit wins.
As an engineering leader at Meta, David sponsored multiple codec evaluations based on Ioannis’s work, and brought this model into conversations with network operators, handset makers, and SoC vendors. This work was also fundamental to Meta bringing high-quality video to 3.5 billion users.
As engineering leader of the Video Processing team, David led the deployment of two generations of custom video encoding ASICs, including the Emmy-winning MSVP processor, which enabled Meta to deliver the quality of convex hull encoding to 3.5 billion people around the world.
While working at Meta on the problem of high-scale video encoding, David proposed a variation of the well-known CAP Postulate applied to video encoding:
In any well-optimized video encoding system, you cannot simultaneously minimize Compute, Distortion, and Egress. Improving any one of these metrics necessarily degrades one or both of the others.
Through Meta, David partnered with Ittiam to build VCAT (Video Codec Acid Test), a tool for benchmarking hardware and software video decoding.
Today, David continues to contribute code and direction to VCAT, with a vision to build a full suite of video benchmarking tools.
For ten years, David organized the annual Netflix Waffle Day, an event where Netflix employees pitched in to make a yummy waffle breakfast for everyone.
David is always up for challenges of scale, and successfully scaled Waffle Day from ~350 employees to ~3,000 employees and family.
Transcoding at scale, codec deployment, quality tuning, or building the team to do it.
Contact RoncaTech