从两位部署者那里学到的关于AV1部署的经验
What I Learned About Deploying AV1 from Two Deployers

原始链接: https://streaminglearningcenter.com/articles/what-i-learned-about-deploying-av1-from-two-deployers.html

## AV1部署:Meta与Evolution Gaming的实际经验 最近的Streaming Media Connect面板讨论,Meta、Evolution Gaming和Rethink TV分析师Alex Davies分享了关于AV1编解码器采用的实用见解。讨论的重点较少关注未来潜力,而更多地关注当前的实施挑战和优势。 两家公司都使用基于软件的SVT-AV1编码,通过FFmpeg实现。Evolution Gaming运营着数千个直播赌场流,选择软件是为了可扩展性和在全球工作室中轻松复制,认为硬件编码速度太慢。Meta利用AV1用于Reels、Stories和Feed,优先考虑计算效率,并利用设备能力过滤器,仅在适当的时候提供AV1。 解码策略各不相同:Evolution严格向*具备*硬件解码的设备提供AV1,而Meta则采用基于设备规格的预测模型,自动为能够支持的用户启用AV1。 虽然显著的比特率节省目前还不是主要目标,但两者都看到了价值。Evolution长期目标是节省10-15%,优先保证稳定的播放。Meta报告AV1现在驱动了大约70%的视频观看时间,收益因地区而异——在带宽受限的地区实现观看,并在可能的情况下提高质量。分析师Alex Davies指出,Netflix报告称80%的*新*设备已通过认证支持硬件AV1解码。 结论?AV1部署并不像预期的那么令人望而却步,特别是与FFmpeg集成后,但需要仔细测试和制定战略性的解码方法。

Hacker News 新闻 | 过去 | 评论 | 提问 | 展示 | 招聘 | 提交 登录 从两位部署者那里了解到的 AV1 部署经验 (streaminglearningcenter.com) 6 分,by breve 1 小时前 | 隐藏 | 过去 | 收藏 | 1 条评论 cmbernard333 0 分钟前 [–] 我一直在将我的 H265 (HVEC) 收藏转换为 AV1。我仍在尝试找到 FFMPEG 的确切参数,以针对我使用 plex 的设备生成最佳输出。到目前为止,有点像打地鼠,但如果设置正确,它似乎在文件大小方面是理想的。回复 指南 | 常见问题 | 列表 | API | 安全 | 法律 | 申请 YC | 联系 搜索:
相关文章

原文

I recently hosted a Streaming Media Connect panel titled “Benefits and Trade-Offs of Adopting and Implementing Codecs.” The contributors were Hassene Tmar from Meta, Behnam Kakavand from Evolution Gaming, and analyst Alex Davies from Rethink TV. The most striking information was not about future codecs. It was about what AV1 deployment looks like when real companies actually do it. You can watch the video on YouTube, here.

By way of background, Evolution runs thousands of real-time, low-latency live casino channels where even small timing errors affect gameplay. Their entire workflow is software-based, built on FFmpeg, and they recently completed their first AV1 deployment using SVT AV1 for live encoding.

Meta uses AV1, encoded with SVT-AV1, across its major VOD surfaces, including Reels, Stories, and Feed on Facebook and Instagram. Their work focuses on massive-scale distribution, per-region tuning, and device-capability filters that determine when AV1 can be safely delivered.

Here are the key takeaways, in their own words, on encoding, decoding, and what they have seen or expect to see from their AV1 integration.

Encoding Strategy

Both Meta and Evolution use software SVT-AV1 encoding. Their reasons are different and revealing.

Behnam Kakavand, Video R&D lead engineer at Evolution Gaming

Evolution runs more than 2,000 live channels across distributed studios. Hardware did not fit their operational model. Behnam said, “Imagine for 2,000 live channels, if you want to buy hardware encoders, it’s not very feasible. Software is the easiest way to replicate the same setup everywhere in the world.” He added that some competitors use hardware-based AV1, but their early testing showed “the per-frame encoding time was too long for us.”

Evolution’s AV1 pipeline is built on FFmpeg. According to Behnam, integration was trivial. “It was just a matter of adding it to a test machine first to see which machine can keep up with it, then rolling it out on our live production machines.” When asked directly how long that took, he said, “a couple of hours of work and then you’re ready to go.”

Behnam further elaborated that Evolution added AV1 directly into the same FFmpeg command that produced their H.264 ladders. “Yes, they are in the same command line. We want these streams to be in sync.” Older machines could not handle the combined load, so Evolution now uses newer Xeon-based servers that can run up to four simultaneous AV1 and H.264 ladders on the same box.

Meta’s Encoding Story

Hassene Tmar, Meta Technical Program Manager

Meta’s encoding story is not about plumbing. It is about cycle budgets, tuning, and scale. Hassene said, “At a scale like Meta, compute does matter.” They compared AVC, VP9, HEVC, and AV1 in terms of both compression efficiency and compute. His conclusion was simple. “AV1 can cover the whole range” while delivering a compression efficiency boost at the same complexity. He also pointed to SVT AV1’s threading model. The encoder can “quickly turn from a VOD state into a live state using multi-threading without any loss in compression efficiency.”

One feature both referenced was the SVT AV1 fast decode mode. It won the IBC Innovation Award, given to the SVT AV1 team this year. Hassene described it as “a mode that makes the encoder optimization decoder aware,” and said it allowed Meta to target “a less than 100 dollar Android device” and still decode a 720p AV1 stream.

Decoding Strategy and Targeting

This was the largest divide in the panel. Evolution takes a simple rule. Deliver AV1 only to devices with hardware decode. As Behnam put it, “We are delivering AV1 just to the users with mobile hardware decoding capabilities.”

Evolution rolled out AV1 on one of its simpler channels, which he described as “talking head kind of streams.” They watched for stalls or complaints. “Nothing bad happened” was his description. After that, they gradually expanded. They turned it on, waited a few days, turned it off, turned it on again, and looked for regressions.

Meta uses a predictive gating model. They do not test each viewer. They inspect the device specification. “We do not benchmark every viewer. We look up the device specification. If it has enough cores, enough memory, enough things like that, we believe it is ready.” It is a static decision tree. When a device meets the threshold, it gets AV1. When a user upgrades their phone, “they automatically get AV1. We do not have to rebenchmark that device.”

Battery and thermal behavior matter. Hassene said, “When the battery level is very low, we try not to use software decoding whatsoever.” They fall back to hardware decode or lower resolutions. QoE signals drive enablement and disablement. If a device shows increased freezes, stalls, or scroll performance problems, Meta stops serving AV1 to that device class.

A small but important note from Evolution concerned hybrid ladders. They initially tried serving AV1 only at the top resolutions and H.264 at the lower ones. Switching between codecs caused black frames. “We saw black frames when switching,” Behnam said. The workaround was simple but expensive. Encode AV1 and H.264 at all rungs. This doubled their encoding work.

Actual and Expected Benefits

Neither company made dramatic claims about bit rate savings. Evolution is not tuning for savings yet. They want clean playability first. “Right now we are getting comparable quality with similar and in some cases even slightly higher bit rates compared to the current H.264 streams.” Their first goal is operational confidence.

Their second goal is to prepare for more complex content. They produce game shows with explosions, bright lights, mixed animation, and shifting sets. “We think AV1 will help us for sure” on these high-motion sequences. Their long-term target is modest. “Let’s say 10 or 15 percent of savings would be a big amount annually.”

Behnam also noted that playback length did not decline. “We have not seen any drawback. We have not seen any drop in the length of playback for the users who are actually consuming AV1.” That was an important signal for their business model.

Meta’s Observations on Bitrate Savings and QoE

Meta’s benefits are broad but not uniform. Hassene said, “We do not really have a flat number on bit rate savings because it is tuned differently for different apps, different surfaces, different regions.” In some regions, savings matter more. “Lower bit rates in some regions allowed people to watch videos they could not load.” In other regions, quality is the main driver. “Users really like higher quality. The more you give them, the more they like it.” The result is scale-dependent. AV1 now covers “about 70 percent of all of Meta’s watch.”

Meta’s benefits also extend into HDR. They noted their work with Dolby Vision on Instagram and said it looks “really good” on iOS devices.

A Note from Jan on VVC

Before turning to Alex, one observation I didn’t discuss during the event. I made a significant effort to recruit a VVC deployer for this panel. I posted on LinkedIn several times, contacted a range of colleagues, and had one near match that fell through right after I sent the question list. This may simply mean it is early for VVC. It may also mean there isn’t much deployer activity to point to right now. With that context, here is what Alex said.

Alex Davies on the Hardware Trends

Alex Davies, Senior Analyst at Rethink.

By way of introduction, Alex is with Rethink Technology Research and recently published their updated Media and Entertainment Video Codecs report, a device-based forecast covering 2025 and 2030.

Alex was careful not to overstate market adoption, but he was direct about where AV1 has momentum. AV1 hardware support “has been proliferating through the market,” particularly on mobile and PC platforms. For services that choose the best available hardware decoder, “there is a very good chance AV1 is that option.” He also pointed to a notable data point. Netflix recently announced that “upwards of 80 percent of new devices they certified had AV1 in hardware.” In his forecast, AV1 climbs because VVC adoption stalled, not because the industry made a coordinated shift.

Closing Signals

Evolution and Meta both gave clear forward-looking statements about the benefits of deploying AV1. Behnam said, “I would highly advise starting to deal with it. Deployment of a new codec is not an easy task. The sooner you start, the better.”

Hassene said, “For the users, the quality of experience has increased significantly. The metrics just keep trending up.”

Finishing Up

My thanks to the participants. These panels are time-consuming, and we appreciate the people and their companies for sharing their knowledge and experiences.

If you want the full details, the transcript is worth reading because nothing in AV1 deployment is hypothetical for these teams. It is their day job. What they described is a codec that is already trending upward in real production environments, not a lab curiosity waiting for its moment.

A few simple takeaways. Integration is not the heavy lift many assumed, particularly if you have an FFmpeg-based encoding stack. It is simple, but you still need to test, watch for playback issues, and expand cautiously.

You also get to choose your own decoding strategy. You can follow Evolution and rely strictly on hardware decode. You can follow Meta and build a software-and-hardware decision tree that adapts to each device. Or you can land somewhere in between.

AV1 is clearly on the upswing – I will write more about this in the weeks to come.

联系我们 contact @ memedata.com