Webinar Compares Codec Comparisons
In this article we’re going to dig a little into codecs. id3as (pronounced ‘ideas’) ran a webinar that looked into the topic of codecs, with Jan Ozer, a codec expert. He expressed the view recently that
“there is no way to compare codecs, you can only compare codec implimentations”.
He gave an outline of the point in an article here.
Better Codecs Essential for 8K
Codecs are really important to the development of 8K and beyond video as there is at least four times as much data in an 8K video stream, compared to 4K and 16 times as much compared to FullHD. Storing that data, processing and delivering it adds increased demand which can add to cost. So, the 8K Association is always interested in codecs, such as VVC, which can deliver the data in a more compressed form without compromising quality.
The first speaker in the webinar was Adrian Roe, CEO of Norsk, a company in the id3as group, that tries to reduce the complexity in video production workflows to enhance the possibilities of creatives. Roe said that his firm’s tools have interacted with eight different H.264 encoders. five different HEVC (H.265) versions, three AV1 implementations and “a whole bunch… of less well-known codecs”. That’s slightly surprising as, he said, “H.264 is H.264. surely?”. Ozer has published the comment on LinkedIn here. His key point was that:
“To identify the benefit of any codec, you have to choose an implementation of that codec, and analyze it using your own encoding/transcoding and distribution practices.”
Roe said that the choice of a codec and its implementation are both business decisions that can make a ‘BIG’ (his capitals) difference. Attendees at the seminar were using H.264 the most, H.265 next, AV1 and also VVC. Roe said that he was pleased and surprised to see VVC in the list. Resolutions used by participants are also increasing, with 4K at two thirds the level of FullHD and above 720P (although it was a small and unrepresentative sample, he added). Higher and higher bitrates was also a trend with more than a quarter of respondents using more than 15 Mbps.
Ozer Explains his Point
Ozer then started by saying that his presentation is available at http://bit.ly/JONorsk25 or here. He said that when he gave a talk last year, he noticed that there wasn’t a single number to compare codecs and there is a particular difference between live and VOD encoding. Although studies can support particular numbers, they might not be accurate or represent the true picture for a particular use case.
Ozer said that one reason that AV1 is not so widely used is that for live use, its performance is not much better than HEVC, yet. Using software to encode you may be able to get 50% better efficiency. However, when you look at hardware compression, the numbers are much closer, reducing the advantage of AV1. To make sense of this, he wanted to come up with some numbers, but found a big variation according to the codec implentation. Although it is often assumed that codecs deliver the same quality, they often don’t. When he talks about quality, effectively, he said, you mean bandwidth. If you have to boost the bandwidth (e.g. to meet a quality threshold such as VMF 95 or another metric), your costs rise. Saving money on the codec doesn’t help if you then have to spend more money on bandwidth.
Ozer did the cost analysis based on both cloud or on premises use. He used data from three studies.
He explained that it’s really important to select the right parameters (“the gruesomly boring stuff” in his words). The studies looked at implementations from Main Concept and Tencent and Ozer emphasized that these can be better than FFMPEG. The commercial encoders could end up cheaper even if you have to pay royalties, if the compression is more efficient. The tests that he looked at in the studies were done with average bitrates although, he pointed out, ‘everybody uses variable bit rates (VBR) or constant bit rates (CBR)’. Tests can also use constant quality encoding, but again ‘nobody uses that in production’. Ozer detailed the differences between the different test methods in the studies.
Ozer identified four ‘red flags’ in bitrate choice, presets and different command strings. He also said that ‘tuning is challenging, there’s no right answer’.
Looking at the results, he explained that the performance was normalised to H.265 at 100% with the other numbers showing the relative amount of power needed to match to H.265. So, the Tencent software numbers showed that the Best AV1 was just 5% more efficient than the best HEVC. Looking at hardware encoding, the best HEVC was 15% more efficient than the best AV1 tested. That explains that although Nvidia, Intel and AMD can support it, AV1 is not being used for live production.
Ozer looked at the cost implications of the different decisions. He calculated the cost for a 100 channel FAST streaming service and he showed how different processor choices in hardware would add to the bitrate costs (from a 23% reducton for an Intel Arc implementation to a +78% $578K boost for the AMD-powered option).
In Ozer’s presentation, he went through in some detail the financial trade-offs between encoding cost and bandwidth cost. Fast encoding, which might mean a lower cloud-based cost, may cost more in higher bitrates. Even at low numbers of hours, the slower encoding with higher coding efficiency (quality) is cheaper than ultrafast. Big service providers should spend the maximum to get the highest possible quality of encoding, he believes.
There was an analysis of the right choice of instance to minimize cost in the cloud. The right choice makes a big difference to the overall cost. He did a similar analysis for on premises solutions and showed the big differences in solutions depending on the physical size and power consumption of different solutions.
More informations will be in an article to be published in April in Streaming Media.
Multiple Points to Consider
At this point, Ozer passed over to Eric Schumacher-Rasmusson, CMO of id3as who said that his firm knows that every case has seven or eight points to consider. Roe said that availability of server instances can be a limit on the choices available. He also pointed out that consumers and viewers tend to be bandwidth limited, rather than quality limited. That means that if you cut the bitrate needed for a particular quality, consumers may switch to a higher quality, so you don’t make the bitrate cost savings that you might have hoped for.
Roe explained that energy consumption is very important and he expects more interest in this as the industry matures. Compatibility can be a challenge if you are trying to support older devices with particular codecs. Latency is also a factor to think about. Ozer agreed that cost may not be improved but offering a better experience can make a difference to the user and Schumacher-Rasmusson emphasized that this can be important at the lowest levels of bitrate.
The on-premises market has changed a lot with the ability to run a lot of channels in one single 1U server. This is less significant in cloud-based processing, Roe said.
Ozer described how he had worked with a provider that had planned to use a fixed 8Mbps stream, but when he did quality measures (VMF 95), he found that nearly all their content could be streamed at just 5Mbps without significant degradation, so they switched, with no response from viewers.
Audience Size Matters a Lot
Audience size can be a big factor. The audience size increases bandwidth, but the proportion of encoding cost gets smaller. Low number of viewer gambling streams can be high quality in terms of bitrate, but not major sports events for mass viewing, where you need the lowest possible bitrate.
Roe pointed out that the costs of CDNs don’t always scale with the way the services are priced. He also said that sometimes it may be better to have fewer quality ‘rungs’ on the ladder if the user can perceive when the rung position changes. It may be better to be at a fixed quality level for the viewer experience than to chase better VMF numbers that interrupt the viewer.
In the almost universally switched off analogue television, each point on the screen has many signal levels, which is used with great effect, for example, in multi-bit SD cells and can be used not only for storing but also for transmitting information, for example in DTT for high resolutions and better use of the limited number of existing TV channels.