NETINT Technologies about Streaming Media Mastery: Edge Computing, Low-Latency Delivery, and More

welcome to nen's voices and video and thanks for joining us today we host will law a Chief Architect at aami who has over 23 years of experience in streaming media and expertise in media over quick Dash hls and low latency Ott delivery will is a rambling wreck from Georgia Tech and a heck of an engineer we won't ask how he drinks his whiskey who holds a master's degree in aerospace engineering and an MBA and has a notable industry presence having co-chaired the w3c web transport work group in the CTA wave task force that will simplify streaming for the entire ecosystem from publisher to player vendor personally will was always a patient and knowledgeable source for many articles that are WR for streaming media on topics like CF and low latency he's taught me a lot and today he he's here to teach you a lot specifically in this podcast we'll discuss a range of topics including Edge Computing for media and its practical applications per session watermarking and encryption techniques manifest manipulation low latency video where will was a true Pioneer Google Chunky Monkey and will law if you haven't seen that presentation media over quii and Trends in high density transcoding welcome will thanks for joining us thank you very much that was a very kind introduction so I appreciate it and it's been a pleasure working with you over these years and it's nice to have this time to to have a talk so thank you for our pleasure so we're going to go about 45 minutes today and we'll try and answer questions as they come in um if not we'll get to them after that well let's let's Jump Right In um tell us you know I've kind of covered a little bit about yourself but why don't to add anything you feels essential and then jump into aami and talk about its architecture because I think its architecture really lends itself to a lot of things we're going to be talking about today I've been at aami 16 years now which was probably not the plan when I joined it at the time but it's it's been a great place to work and it's a place that's evolving and changing so I think a lot of people consider acami as a CDN and they put ACMI purely in the CDN bucket but CDN today is is about a third of our business and the other two legs of the aami stool that that are now grow the fastest growing legs are security security side of the business and then the newest one we have is compute so between delivery security and compute it's it's a really thriving business we had excellent results two weeks AO go and we've hung in there I think there's there's there's some stat I don't want to misquote it about startups in 1999 that are now worth more than $15 billion do and are still around I think there's only five of them or so and it's some like Google that you know well but one of them is aamine so we've actually uh been a player for a long time and there's there's still a startup mentality in many people's minds but it's been two decades now so we really can't claim startup I think we've we've passed the the the threshold for uh getting into as a mature company impressive yeah I I came in aami always on the media side so I came in running player engineering came from flash background flash one day solved all my interrupt problems I had with Windows media and real media and I think it was a great tool I think it's its history was skewered a bit by the failure of of Apple to support it but we certainly had demos media capabilities in the mids that we weren't going to have for another 10 years with with raw HTML so there was a bit of a setback there but that's when I joined aami and then I worked a lot on the player side um wrote libraries and wrote software then I became more of a manager of our player engineering team and then I shifted more into an architect role and I think for the last five or six years I've been in the Chief Architect role and for those who want to complain about our product map I'm not the product architect that's Kevin Jones he's an excellent product architect he has a team of product Architects under him uh so I work a lot in the industry I am I spend time in the standards world and I try to work on things that lift the entire industry it lifts the aami boat but it also lifts other boats but at the end of the day it's that's good for business so that's why we invest a lot of time and money at eomi inter interop interoperable standards one of the first s we connected was with cmf and I just didn't get it it's like what do we need another format for and um cmf has really come a long way I mean that's the whole focus of the of the wave project which really where where's the wave project today I mean how far along is it on the on the Jour wave is very active so wave is a funny name it had an even funnier name in the beginning it was called the give project global internet video ecosystem was terrible people thought it was a charity for video Engineers which is not a bad idea but it's not a for video Engineers it's it's a the CTA is is the group that runs the CES Trad show so they represent all all the vendors of of uh electronic items but also of televisions and they were quite a stute in realizing that the the the entertainment landscape was switching uh and that Ott was a new and Powerful entr in in the media and entertainment market and there were still a lot of gaps the the broadcast world is heavily defined by standards and it's mature but in Ott there's still interrupt gaps so the the goal of wave very simply is to remove interrupt problems for rot streaming and there's interrupt problems about the specification of content there's interrupt about how a TV device renders video how how it plays it back inside a browser so web's done a bunch of tooling it's I think it's 8 years old at this point or something like that it's been around and it's quite productive and then we've branched out recently into observability so cmcd is one of the most popular specs we've made in wave that's a group that I ched uh for V1 it's been out for two years we're getting a a lot of pickup right now and we just started work on cmcd V2 so Public Service Announcement if if people care about what V2 looks like and there's certainly a good diversity of opinion and what it should look like um then I think that people can uh get involved in wave and if you're an sbta member you get automatically the ability to participate in the wave meetings which is a nice on train we get a good good group of people coming in through SBT it's kind of remarkable you can criticize ntsc but at least you knew when you when you turned on your TV it would play video from ABC and there's there's no consistent specification that does the same thing for streaming so Netflix and Amazon Prime and Hulu spend a fortune on compatibility of players and it's it's silly and and really that's what you're that's one of the things that your project you've got to trade it off right because if you if you also look at television it's monolithic there was only one way to send TV to TV sets for a long time right three three evolutions of of broadcast from analog switching to digital that was a big one and now we have next gen TV right but these are these gaps are are decades apart and you can't go off and innovate on a slightly better version of TV you have to lift the whole country up at the same time so you get the reliability and security but it's from a very common Baseline whereas an Ott we can have someone who says I want to do HDR or I want to do 3D I want to do something different and you can go do it it's a much more flexible distribution mechanism but at the same time the the problem you just raised which is diversity of players diversity of formats and interrup between them so you want to strike a blend you want standards like hls and dash which have lots of permutations within them but still still provide a general field and then you can provide constraints within those around which you can run businesses right but you can always make something new and that's the exciting part of it why don't you know from a CDN perspective could you describe aami's architecture and and kind of what makes it special um or at least yeah so Eis differentiated itself from the beginning from other C n just on the diversity it's highly distributed so a lot of CDN or content providers will say we have 10 or 20 or 50 or 100 pops around the world points of presence places from which you egress content aami today is about 4,200 so our goal is to be inside every as or significant as delivering media to customers in the world so we're in in a city we're in multiple as's with and peering points Point within those cities were highly distributed and there's a cost for that it's more expensive to go Mount servers in 4,000 locations than it is to mount them in 40 locations but we reap a huge benefit we get a huge surface from which to deliver content we get a huge surface from which to deflect dos attacks and we can run an overlay Network we can from each of our points of presence we can map connectivity to our other points of presence and if we don't have to follow bgp routing we can route directly at least that's what we did for many decades today we we've starting to connect fiber between our major points it's just we have so much traffic that it's worth our while to lay our own network down and we've laid what's called AC my fabric uh and it's our next level inter interconnect between our centers but we're we're a highly distributed Edge and we used to only use the internet to move between the edges but today we have a private Network that we can fully utilize and it gives us very good cost efficiency for that so when we talk about Edge Computing you guys you guys have a lot of edges we have a lot of edges but we didn't have a lot of compute right we just had delivery we had our code running on our edges and nobody else's code so the big change was last year when we acquired Lode as a as a as a great step off point into the world of compute and now we're starting to marry Len no's core compute with the notion of edge compute so the node we still don't we have we had we bought 11 core sites I think we added 10 more of this here but the the core sites or the the really centralized high capacity compute sites will number in the tens right but we still now want even deeper Edge compute than that so we're having next to your CS deploy about a 100 much smaller satellite computes that are deep within the networks and married very closely with our onet distribution capabilities so that's a great for many media use cases you want to be doing your compute and this is for higher level compute we also have Edge workers which is our serverless layer our JavaScript serverless compute that is executing in all 4,000 uh regions so we have serverless then we have this Edge based heavier compute and then we have our core compute in the center and I think that that hierarchy suits itself very well to what many applications want you want an outside layer that's got a high surface and we'll talk about all the things you can do right at the edge especially with servol list and then you come down to a heavier layer where you might want to transcode or prepare content and then deeper down is where you're running your databases your synchronization and your heavier business logic let's talk about that our focus is is primarily media but what what Edge applications are you seeing make sense for your customers these days there's many but the some of the obvious ones are ones that require personalization for every user who's connected so for example uh AB watermarking you want to synthesize a unique stream of ab segments for each user that's something that's an edge based function you you don't want to run every request for every user down to some Central Authority you'll just swamp it so Edge based border marking also edge-based encryption we can do hls envelope encryption today a unique encryption per user as the contents going out that's another action you want to perform there um then anything involving manipulation of playlists so the the beautiful thing about adaptive HTTP adapted segmented media HS and dash is they're powered by playlists right they're powered by text files and these text files are beautiful opportunities to manipulate content at the edge they're very lightweight but you can make change to make changes to them very easily and those changes are quite powerful so if you want to localize your content you can have a fat playlist that's listing 16 languages and right at the edge as the user connects in their as you can modify that manifest or playlist and strip out all the other languages and just leave the languages that that client needs for hls for example if you want to re hls takes the first variant playlist as a signal for the one to start loading first so you can manipulate that your network might have knowledge about what the throughput of the connectivity is for that client and we can prepare an appropriate playlist for them so that they start at the right bit rate you can filter content you can put 4K in you can take it out um you can do a lot of manipulation so that's geolocalization and personalization it's really popular talk about the you know who is who is doing this it sounds like the major Publishers and you know getting over the technical you know what are they doing from a user perspective how do I see this if I'm if I'm watching in a in one of these regions it's very hard for you to see it because you're going to get something and you have no idea how that was syn you might get a playlist that has four bit rates and two language options but actually it was a master playlist with 16 bit rates and eight language options so as a user you can't tell that it's it's it's opaque to the end user and so it should be you should just get content that's appropriate for you for your device it should be localized for you uh and should work um that's the whole point of personalization and you can also personalize actual pseudo live playlists at the edge which is another area a lot of VOD Distributors are getting into today if if you're a news organization and you've built up 10 years of VOD material the fishing fishing in in Iowa or hangliding in Atlanta right these are very Niche subjects there'll never be a live broadcast channel for them but you can construct a pseudo live channel from VES and put in Dynamic Advertising and stream make a channel that's just for Yan or just for will that we like and it's got advertis that that advertising is a much higher CPM because it's highly targeted and it's not broken out across a larger cohort so Dynamic content synthesis especially of of linear streams is also very popular at the moment you know what about cloud gaming you know we're hearing a lot about that is that an edge Computing thing or is that going to be done you know at the or yeah so there's some cloud gaming that's done at the origin you might want to authenticate against An Origin but there's two parts of cloud gaming that really require Edge participation so the first is we have multiplayer games massive multiplayer games or just multiplayer games right the the users downloaded a large game download and the game is executing and running on their device but the game playay instructions need to go to a game playay engine that is synchronizing playback and all allowing the users to play the game and to be fair to the users you have to be very careful where you place that gameplay engine imagine you just had one on the US East Coast and West Coast and one in Europe then somebody from Africa joins or Australia joins right they they are going to be so far behind everybody else because they their rtt is higher so and you can you know the IP addresses of your connecting clients and you there's there's a lot of advanced algorithms for picking the centroid the best set the fairest centroid to to run your game engine but what if you don't have a data center there if you've only got 10 in the world you're stuck so that's why having thousands of places you can run a game engine lets you move it closer to the optimum Point um and the second one for gaming is actually game synthesis which is another level up in terms of edge compute but this is where you have a Very Thin Client your game is actually being played and rendered in a virtual machine at the edge it's making unique stream for every player and all the players doing is watching a video essentially and then feeding back user interface instructions back down to the engine so that's also popular because it lets you play sophisticated games that you don't have the hardware or the software for on the client the trouble comes in monetization it's very expensive to encode a 4K or an 1080p stream for every single user but with vpus and I think you're more well versed in that world than I am that cost of encoding is coming down and I think that's going to mean the intersection of um profitable game streaming from the cap the the capability to do it versus the ability to monetize it that that might overlap a little bit more in the future you know one of the things that's kind of confused me a little bit is the whole VOD to live and we talked about that a little bit before and you said that was wasn't an edge application that was more of a central application what are the use cases you are seeing for VOD to live these days what what are they look like who wants it VOD T typical VOD to live is a sports game right the sports gamees played out live and there's there's two use cases people either watch or three people are watching at the Live Edge or people are late to the game so they want to start watching one hour back from the Live Edge but then they they want to watch it play out in live and then the third use case is people come back tomorrow and want to watch the game so live toad is about allowing or DVR window live toad I get it's it's VOD to live that's a little bit confusing oh sorry sorry I I confused it around so VOD to live so if you think about a live broadcast on a television station even that one we're all familiar with it's a sequence of programs right those programs are all VOD assets that were shot last year or last month or 10 years ago if you're watching rerun somewhere and then you interpose them with customer advertising which is also VOD assets right and maybe news which is shot live in a studio and that's that's your typical linear TV so it's a sequence of a bunch of va assets interspersed with some live ones um and most channels even don't have live news so they are literally just a sequence of odd assets that are played out as if as if they were live and we can do the same same thing on the edge yeah why wouldn't you just store it as a as a an En coding ladder that's already been transcoded seems like the transcoding cost is a major cost of that why would you store why would you go back to the mezzanine file for the broadcast and you're not going back to Mays at any point you're taking these assets were encoded and you got a bitrate ladder and and you you played your live you played your live event right and now it's a VA asset or you've got a vet and you want to play it out as live you don't need to re-encode them okay you can you can you can adjust the time stamp if you want to or you can play tricks with your manifests to offset the starting time be it Dash or or hls but really you're just it's a sequence it's a playlist of v files that the user is getting and then interface on their device makes it appear like it's a live stream so VOD to live isn't re-encoding it's just making the not at all it's the avoidance of re-encoding that's you're just playing with text files for VOD to live you are not touching the audio video end codes what is content steering so content steering is new spec uh hls came out with it last year Dash DH mirrored this year so normally players get instructions from a playlist or a manifest right and in those depending on the format there's alternate sources for their content and they only move to these sources if there's delivery problems uh that they experience but different players will react in different ways so some if the player gets a four or four on a segment some players will retry twice some will retry three times some will wait 10 seconds and try again others will switch bit rates others will give up completely and go try an alternate source so the failover logic in both h less and dash is not specified tightly so it's very hard to get a cohort a large cohort a million players who move consistently if you want to move them from one CDN to another at any one time there's no before content steering there's no way to do that so what content steering does it adds in a URL that the player connect to to something called a Content steering server and that server sends a little Json package with some information for the player and what it does is it steers the player between sources of content that already baked into the playlist and it can actually clone or synthesize a new source of content that doesn't even exist yet but is derived from looking at one of your existing sources of content in your playlist so the whole point the original point was to move uh players in a cohort players in a very deterministic fashion um for qoe purposes to improve if you knew a CDN was having performance problems you can move them over it turns out in the last year that there's other reasons to want to move cohort to players you might want to do it for economic reasons one CDN is more expensive at prime time but gives you better performance but you want to switch away from it outside of prime time because then the other cdns are okay or you want to switch because you're switching content so there's a use case now where some people have 4K on one CDN and 1080p and Below one another and they use content steering to move their premium subscribers and steer them at the the original CDN and I think we're going to find more opportunities for Content steering as we go ahead it's a very interesting way to control a set of players um after having given them um some initial play playback instructions and I think people are going to leverage it in different ways what level of customer are we talking about for all of these techniques that you're describing is it you know we've got you know the typical pyramid of Publishers on the internet we've got the huge companies on top and that's the you know 05% perhaps I mean how far down the pyramid have to go so it's going to be mostly like the .5% cuz they're the ones with a million players that they want to steer and they get the most benefit they're the ones most likely to have multi CDN contracts to begin with and to want to invest in a steering service sufficiently to get the gains to to move their players so I really it's it's new like I don't know I know brov are building a Content steering service that can operate at scale uh but it it's not as if there's a big Market or established market for this year this is quite new and it introduces a risk too your steering server can now become a single point of failure if your steering server goes down or gets bad information or stale information now you've made your your QE problem worse than it was before at least your players were autonomous and they went off and did their things now they're they're tied to a service that's misdirecting them okay be the bigger people first but that doesn't mean if you're a small company you can't use it or you might want to steer uh in an Innovative way so you don't even need different cdns you could steer within different delivery properties within the same CDM you could have a delivery property optimized for apj and one optimized for emia and you could steer people between it for examp and what is one of the other kind of edge Computing topics we talked about was program replacement what is that who needs it who's using it program replacement is doing something like inserting a bumper into a VOD file and it might be for a sports team or it might be for a VA asset you can think of advertising as as program replacement but usually it's advertising is a separate insertion mechanism and then you simply want the ability for g a given playlist to inject uh a subset of files and it's nearly always something to do with with bumpers or or pre-rolls something like that so that's another thing you can quite easily do with text file manipulation at a serverless layer as you're delivering the playlist or the Manifest to the end mus so let's let's switch topics to load latency and you know one of the things that happens in the industry is is you and I tend to live really at the Forefront of standards as they come out and we write about them and we talk about them we argue which is best but the market adopts them very slowly so where are we with the the adoption of Technologies like low latency hls and low latency Dash and C we're we're in the slow market adoption phase I would say yeah I was all excited dashow latency came out probably four years ago H llh came out two years ago but we've we've we've and we saw early deployments on the dash side because their their requirement on the origin is much lower right the same same origin an HTTP server works with low latency Dash and standard latency Dash is just using chunk chunk transfer for the delivery or an aggregating response I should say because it works with htb3 and that doesn't have junk transfer um hls on the other hand places a much higher bar on the intelligence of the origin server it has to follow hls delivery directives it has to follow Skip instructions it has to hold back so low latency hls takes your timing problem that Dash solves with in basically by requiring the client to have a notion of time hls the client doesn't need a clock but it does need to know when to request content and since that can get sloppy and your sloppiness is a function of your segment duration um you need you need to find some component that knows the right time to release content and that's the origin so hs's timing model is basically shifted down to the origin which needs you need a bigger origin so we've just seen AWS Elemental release low latency hls support so that's that's that's the part right and now we put a horse in front of it but there was nothing for the horse to pull before that um so I think you're going to see more llh deployments and there's also competitive pressure on the industry once once the major sporting provider lowers their latency from 20 seconds to 8 seconds or 6 seconds which I don't really call low but that's honestly where a lot of people are targeting right now broadcast equivalent latency once they do that then it's looking bad for everybody else and then that will pull everyone else down to 6 seconds and then below that we get into what I call Interactive levels of latency this is the latency level where you want to tweet or throw hearts up on the screen or shout or cheer or make comments with some broadcast uh and that's typically around the one to 3 second Mark so we're seeing low latency hls low latency Dash Target in production Target three seconds and up they both have demos below 3 seconds but when you when you look at scalability it's higher hesp from theop player does operate at the 1 second or 800 millisecond Mark um but there's less players for it and it's not as broadly supported today but it is an Adaptive segmented solution that took want one of the problems with both hls and dash which is you can only switch it at at your segment boundaries and it solved it by giving you you know much more diversity in segment boundaries for example or the equivalent of it so you can switch quicker and switching quicker is the secret to low latency in stable playback um if you can switch out of trouble if you can detect trouble and switch out of it quickly then you you can keep your stream going with a small buffer but otherwise you have to buffer up to cover essentially the blindness you have about when you can switch and also waiting until the next switch opportunity so you're looking back at the you know the 0.5% we talked about are they embracing low latency now and what's what's their typical Target we've got the the interactive applications for gambling or for auctions or for any type of um interactive but you know the 05% the broadcasters what are they doing and why are they doing it be be careful of the language you use by saying 05% it's it sounds minimal right it sounds like a trivial amount that I don't have to worry about those 0.5% are probably delivering 90% of all video on the internet so let's rather talk about 90% of all video on the internet and not worry about who would comes from so yeah the first the first instances we're seeing of latency are are people looking for broadcast equivalency for sports and that's happening at a big scale that's happening at millions of of concurrence what is the latency of that that's 5 to 10 seconds or it's in there 5 to 10 depends where you are in the world if you're DB T2 in Europe you're going to get 3 seconds if you're in the United States on a cable connection you can get 10 seconds so it's not constant you need to tune it if you actually want to be precisely aligned with your OT your overthe a or your cable broadcast you need some way to tune the latency that the player is going for so that they are align uh but in general it's it's five to 8 seconds or I'd say six six to 8 seconds somewhere in there and for people have been Opera operating at 30 seconds they call that low latency for people who been operating at 8 seconds they call 3 seconds low latency and the web RTC guys are going that's really not low latency at all and they're all right they just have different frames of reference well and different applications and different goals for those applications so what is what is quick what is uh quick is RFC 9,000 it's ietf protocol for uh basically a bin transmission a quick quick is a protocol that gives you concurrent streams of reliable data and it also gives you datagrams so you can use this to move HTTP but you can also use it to move media and it turns out to be very useful and you might say why do I want unreliability in media well the answer is because reliability comes at a cost it comes at the cost of time in retransmitting data so that it does arrive reliably so having the option not to waste time in retransmitting and to look at other schemes for how you might recover your data that's nice flexibility so quick is is the evolution of HTTP it's replacing TCP within the HTTP stack um but I'm also working on media over quii which is looking at how can we how can we change media to start leveraging uh quick and and leverage some of the properties are quick so that we do actually get better performance at the end of the day than we do by simply using hls or Dash with HTTP 3 we have a question from Monica Gonzalez mono um telefonica Madrid why doesn't hld admit bitmap subtitles something you can address no I don't what is HL I I don't know what HL I'm sorry hls it was was a miss it was a oh why do they do bit map subtitles I don't know um I know a lot of Dash so does dash do bit map yes there's a big argument you know a lot of directors like bitmap subtit image-based subtitles essentially but they're much heavier weight uh and they can't be translated or for example as can imsc or um web VT subtitles so uh Monica unfortun I don't know the answer is to why hls didn't support that I would have said they did support it so that just shows you my level of knowledge about uh hls and supporting uh image based captions okay what about U media over quii from an overall architectural standpoint is that is that going to give Publishers a single architecture as opposed to having to support web RTC and low latency variation yeah so couple of things media of is very interesting um it has a number of goals it's trying to solve problems um and one of them is tunable latency so medov Qui allows you to do realtime conversations which is the conversation we're having right now sub 400 MCS from cam from glass to glass 400 milliseconds or less and then it like you can also do interactive one second to 3 seconds and you can do higher latency 5 seconds and up and you can even do Bo which is infinite Laten so you can you now have a protocol that runs across the Spectrum so you can tie together use cases you can tie the web conferencing use case or the webinar use case you can combine that with sports streaming and if you're bold you can even combine that with your V delivery so instead of having three separate networks now you can have a single Network and when a single Network can monetize itself across more of the market it's cheaper for everybody cheaper for the people running it and cheaper for the people buying it because if that Network's not delivering a live sports event it can be delivering your Web Conference and therefore uh it's it's cost per bit is lower so Medi over Quake is a tempting to be a protocol that can that can be used across a tunable latency spectrum and also through networks of thirdparty relays AKA cdns and that's one of that's if you look at the early the early and and the other attribute is it's a pub sub protocol right we're not if you want to watch a media of a quick stream you say hey give me the sport stream at 1080p you make one subscription request for it and for the next hour you will get data you don't have to ask for anything else whereas hlso Dash every two seconds 4 seconds 6 seconds you're asking for a text file and a video segment and an audio segment you're making lots of requests it's very chat um so pubsub is interesting pubsub suits itself to Media because media is inherently temporal the encoder knows what the next segment is the client wants the next segment and it also wants it in the same order that the encoder is producing it so the smoother we can make the pipe between the two the lower the latency and the more efficient the transmission but the challenge becomes how do you cach that and we need to cach that to get scale we need to cach it to get the ability to play behind Live and Not only at the live play here um so that the challenge with Mark is cach ability through thirdparty relays with tunable latency and with a common protocol that's used for both contribution media workflow and distribution when do we see Technologies like AI start to influence any of this stuff so AI is already being used I mean the the examples that show up in most of the media conferences are object extraction so if we ran AI against my video would say oh there's there's a fern in the corner and there's flowers there and there's a male caucasian of dubious age with a blue shirt and it would extract all this metadata and information out of the video and that's very useful because you can search it then it it makes opaque video searchable um I just saw output from bite dance research paper where they're doing synthesis of video so you can give a prompt just like you give a prompt to an image it takes an in video you can prompt it and it can change it in real time and produce a different video and today it's it's still a little awkward but the image synthesis if we go back two years was equally awkward with geni and today it's so it's scarily good so the video synthesis is is the worst it's ever going to be right now and it will only get better and it's going to get better to the point where you can say I want to synthesize a video engineer from AC Ami talking for an hour and it will go it will basically give me back so I don't think we're far from that and that's that's interesting I'm out of a job but you got to train it on some data so maybe I can license my data yeah don't stop publishing what a what about from a deliverability perspective or is this problem just so simple it doesn't lend itself to AI in delivery no so there's a whole lot of M well I don't like to use the term AI for engineering applications it's machine learning right and and machine learning is really statistics we're we're got a bunch of data points we're drawing a line between them it's more complex than that but it's also not more complex than that we shouldn't we're not going after general intelligence to do useful things so anomaly detection is a core part of machine learning and that's very useful for log files you know aami gets over a 100 million requests per second coming in right so a human can't possibly say is this normal for right now um does this look like what it did last Thursday at this hour in this location in this city is this a is this traffic request pattern coming in is that something that's typical so for security analysis and for finding anomalies and and detecting when a a request might be malicious versus non-malicious is a whole Suite of applications for ML so you can see ML on the the data analysis side it's very good at at doing that it's very good at monitoring it's very good at finding problems it's very good at detecting when things are not quite right you talked at at the start about aami adding compute capabilities and security capabilities where is is is transcoding coming in as part of those compute or is that coming in as part of the CDN so we're not so we're not offering compute as a service today we offer VMS on which people can then host FFM big or bring your your favorite uh encoding solution and run it run it at the edge so people are doing that today and you know Edge based transcode makes sense especially in a user generated world because you've got video coming up from people's mobile devices it's relatively heavy you don't want to move it too far you want to be able to process it close to where it's being produced and then once you've processed it you've shrunk its size potentially unless you're just relaying it and then you want to Route it to people who are consuming it so we're seeing applications on our Network um the nice thing with aami is we have this Edge compute married with a very large delivery surface we just hit a new traffic record last week 384 terabits per second which is a lot of traffic if you go back to aami 2001 we crossed one gigabit per second that was the entire network traffic and it was big deal and there was celebration one gabit so 384,000 times more traffic crossed our Network last week than it did in 2021 that's ridiculous growth and I think the next 20 years will go up there'll be PAB bits per second PAB bits per second within years um it the the rate of increase is growing the rate of consumption is growing um so so where is trans coding you know high capacity transcoding fitting into your product line back to transcoding so it makes sense to do some transcoding at the edge it it makes sense to do other transcoding in the center of the internet where it's cheaper and more scalable I think of edge compute as Miami Miami real estate right Edge compute is your Miami Beach real estate you've got to want to be there because you're going to pay through the nose it's expensive it's a limited resource so you need an app application which is I want to have a vacation and I want to walk out onto the beach that's your equivalent I'm going to pay to be in the beach hotel but if you just want to work there you can move five miles Inland it'll be much cheaper and you don't have the need to be at the beach so a lot of use case a lot of media processing use applications are power intensive CPU intensive and if you're simply just transcoding a library video that's not something to do with the Ed right that's something to do in your cheapest data center and cheapest is not just where it's running it's I have to ingest these mes files and the further I have to move content if my data center it might be cheaper to run it next to a hydroelectric plant in Buenos arees but I've got to move all this data down to Buenos arees and then move it all back again so your optimal location is going to be a blend of your transmission costs both in and out and your actual cost of trans coding but it's almost certainly not going to be in a highly distributed Edge which is a lot of moving and into a lot of more expensive environments uh for content and coding so how has transcoding I guess morphed from CPU to GPU to Asic based how are you seeing that working out so I think you're closer to that world than than I am I certainly I've seen a whole lot of people doing CPU up till now and I think IBC really opened my eyes to just the power the the the huge jump or the in in capability or the decrease in cost per gigabyte out that that vpus are starting to bring to the table now and I think economically if you're not using some type of vpu you're going to be challenged like yes it costs money and yes it's Hardware but it's giving you such a bump in your capability that I think you're going to see it adopted in the cloud platforms I I can't we're a public company I can't speak about AC my road map but it would make sense to me that if you're offering transcoding as a service in the cloud and Amazon r that you should be looking at at some type of vpu to to basically give you the the cost benefit if you're not other entrance are going to pop up and do that so I think it's it's a shift but at the same time codecs are getting obscenely compute intensive and they you know av1 is that much harder to encode and so is VVC so it's a little bit of a game machines get more powerful the vpus get more performant but then the codecs get ever more demanding yeah we we talked about the adoption cycle for load latency Technologies isn't there kind of an analog with codec adoption every you we were talking I guess av1 shipped in 2018 what percentage of video on your network would you and I know you don't have a a real no but you did a very interesting table actually didn't you just do that chart where you showed that like 60% of all media on the internet today is ABC which is a 20 year old codc yeah it wasn't I think that was the uh Tommy Flanigan and the and uh Mr Alex Davies that rethink had a had a chart like that I think I reprinted it the numbers are some there the reality is the V the majority more than 50% and I would guess in the 60 70% of all video on the internet today is ABC h264 and that is a 20-y old coding and it I was for years I've been saying look at the efficiencies vp9 hvc is now well supported on mobile devices name a device that doesn't support hvc yet it's still not being used and the reasons and hvvc is 10 years old at this point as well so I asked I asked a vendor who's well aware of the performance efficiency they get and and the ability to re reduce their delivery bill in the long term and I asked a vendor why don't you do this and he said because in the short term it's more expensive I'm producing ABC and I got to produce my next gen codic at the same time and I have to hire people who understand that I have to redo all my encoding profiles I've got it I've got this cost the short-term cost that comes in and the trouble is media and entertainment is a competitive business and a lot of people don't have road map to say I'm willing I'm willing to suffer for the next year in cost so that the next five years I can benefit it's just too hard to get over that hump and people do hbc's coming up av1 will come with the you know with the Apple giving it the Boost in Hardware decode support and I think VVC will come as well but they come far slower than you would expect uh and a lot of it is to do with this this cat this cat the friction in an exothermic reaction action you need you supply some energy to get it going you need the same to convert your uh encoding your Cod do you remember your cost per gigabyte for high volume transfers in in 2013 because I think that's a big piece of it too I'm remembering break curve coding like 50 cents a gigabyte um not that it was in this it was in the sense it is scary how how much it has fallen I I I don't know if I can quote current ACMI prices but I it's safe to say they well well well under one cent per gig like well under one cent per gig and you can go to Amazon and look at their prices we're not going to be that different but the point is to deliver a movie encoded at 1080p for two hours might be two gigs right encoded in or less encoded in a good codec so it's costing the vendor like sub one cent to send that that video to you and you're paying them $18 a month right so the costs of distribution are really small really in the scheme of things there's other big costs content acquisition costs way more than content distribution does so and the sad thing about the CDN business and one of the reasons aamai is really attracted by compute and security is that if you look at your CDN Bill the price per gigabyte declines by about 15% per year every year for 20 years like would you like to run a business where the cost of your core output goes down every single year by 15% it's hard to do that it has to be offset by an increase in volume right otherwise we go out of business and yes there are increases of volume but it's a Perpetual battle for a CDN to offset the expectation that my bandwidth is cheaper and cheaper and cheaper every single year so from a technical perspective what's what's keeping you and other engineers at aami up at night what are the what are the big problems you see coming up that that keep you keep you at your desk my problems eiz we have people dealing with very very different problems our problems are highly personized like these people worrying about scale right these people worrying about security these people about building out for sufficient capacity uh and schedu you can't just get terabits of data in the Middle East next week you have to plan months ahead for deployments like that you have to deal with outages and these systems are getting so large and complex that you you can't rely on traditional QA to find bugs QA is about here's a list of problems that might happen I'm going to write unit tests for all of them and run my unit tests and I know I won't have any problems but in complex systems you you have a combinatorial explosion of things that might happen that you can't anticipate all the conditions under which things go wrong so you end up having to build systems that compartmentalize they detect when problems are happening and they stop them spreading and that's that's more important than trying to figure out up ahead of time all the problems that might happen so in the networking world it's it's complex my what what keeps me up at night in in particular right now so I'm I'm focusing on Medi over quick um it's in a Nason stage but I think it's it's a very interesting opportunity you don't get many opportun unties to look at a Next Level media distribution system come up I think it's a very interesting one and the challeng is to make it better than HS or Dash and actually solve the problems it's easy to produce architecture slides that say we're going to do this and this and this it's very different to make working systems at scale that do these things so that's my current Focus it's so funny cmf came out when it was cmf was about 2014 2015 somewhere in there and it really started paying dividends in the in the 21 22 23 yeah I remember you were moderating a a session at streaming media West once and I was like the seaf fan and everyone was like yeah don't want another format but yeah these things take time you know CF is just a it's a common file format and it's like hey if we use the same file format then we get less less different stuff to debug better playback and less diversity of content and we're slow to getting there we still have the problem with cbcs and counter mode but it's eventually it'll all be C cbcs and then we won't worry so much and we'll look back on it but yeah I'm I'm always saddened a bit by the rate of I shouldn't be right but I'm these things take time to roll out even Medi over quick it's it's going to take years so let's not do a we well media quii is not successful because no one's using it in 2024 that's fine let's look at 2028 and and see where we are Terry Fier added a comment I don't know if you saw it that Comcast and Warner Brothers claim they stream HBC to 60% 60 to 70% of their devices which um they're pretty aggressive we know that you know Netflix is pretty aggressive in in what they do I'm sure Amazon is as well Hulu so how you slice I I believe Terry's number Terry Terry should know if if you look at new devices right if you look at devices that came out of the last three years that number is closer to 90% it's old set toop boxes very weaker Android phones people who haven't updated stuff that's eight years old so it depends how broadly you want to support your market and no one wants to give up 5% of their addressable market so it forces you to carry ABC and at that point if ABC is good enough I'm not going to incur the expense to go to HBC however it's a simple C they look at their CDN Bill and they go ah H if I do atbc for the popular content I I only encode 20% of it but I reduce by half my my cdmv and then they do that they they reach the point at which the economics makes sense it it's probably worth setting acemiz codic agnostic right you don't care you just want to car CDN it's binary objects we move them we are the we are FedEx for your media right we don't open the boxes we can't open the boxes it's encrypted much of the time uh and we don't want to open the box because anytime you're opening boxes you're slowing stuff down so we're just we're just moving them but and and you know we care like yes if everyone used the latest codec the volume in theory would go down but that doesn't happen a codec swings both ways you can use it for the same quality parity reduce the size you can also for the same bit rate parity increase the quality and actually your end users Jud judge you not on the bit rate you send them they judge you on the quality of what you send them so there's Market pressure to use a codec to raise the quality not so much to lower the bit rate and and of course there's 8K and HDR and and uh meta um so there's you know whenever compression gets more efficient we want to bigger experiment we we expand like 8K is not so unreason you can do 8K with 40 or 50 megabits per second quite easily and I'm talking with hvc let alone VVC and the Codex get more efficient as your pixel count gets larger and it's um if I just this data is like two years old right but I did a study of looking with with cmcd we get the players reporting what's the throughput that they're seeing and I looked for a large US distributor of vaud content they were reporting year year and a half to two years ago 34 megabits per second what the player was estimating as the throughput right so that's well above what I need for 4K that's average yes there's people who are sucking and there's people who have fiber but the average was 34 so we're not that far from sending 8K but 8K is ostentatious wealth in terms of distribution it's it's the Ferrari of driving and most people don't need it they just want to get from it they just want to watch a movie and so I think it's it's going to be very hard to make a market where you can warrant moving that much content all the time we started before we went on air talking about our common time in Atlanta and I remember the house we bought uh off north side and struggling to get a 288 modem working and looking at real video at you know 19 kilobits per second and it's you know in in in the the 25 years since then it's just shocking how uh how much this has progressed yeah do you know what what's interesting I was actually looking at a curve of um WiFi throughput capability for for a phone it was like a Samsung a something a phone that's been produced for a decade and they plotted with time they plotted per Mo's lore which is doubling every two years and they plotted Wi-Fi it was exceeding Mo's law so the ability to stream data in your house has crept up hugely and we used to like be laying rj45s everywhere in our houses because that was cool and it was a tech world you just don't do that anymore it's way fast your Wi-Fi exceeds your Your Capacity to consume and we get to a point as you mentioned with meta that eventually if if I deliver like 16k either I there's diminishing returns on delivering more data as a human just can't perceive it but at at that point I have to saturate my vision I have to send an entire world but there is this it gets harder and harder to come up with use cases that increase but for now I I think uh metaverse type applications will be the next bump up in media content because you have to send a ton of data you have to send a ton of data as long as you can't do something like fiated rendering but you actually have to send a tiny amount of high quality data and then and then lower quality data will suffice for the rest and I think our head tracking and our ey detection might get good enough that the direction we go is in sending a tiny amount of very high quality data by being able to direct it just to the center of our eyes okay we're we're out of questions and we're out of time will it's a we could probably talk all day it's a pleasure to see you again it's been a while and um I really appreciate the time you spent and the uh expertise and the knowledge that you shared so thanks for coming well it's my pleasure to thank you to you and Anita for today and I I look forward to chatting with you and others in the industry at the next conference cheers take take care good to see you

2024-01-09 22:42

Other news