Voices of Video
Explore the inner workings of video technology with Voices of Video: Inside the Tech. This podcast gathers industry experts and innovators to examine every facet of video technology, from decoding and encoding processes to the latest advancements in hardware versus software processing and codecs. Alongside these technical insights, we dive into practical techniques, emerging trends, and industry-shaping facts that define the future of video.
Ideal for engineers, developers, and tech enthusiasts, each episode offers hands-on advice and the in-depth knowledge you need to excel in today’s fast-evolving video landscape. Join us to master the tools, technologies, and trends driving the future of digital video.
Voices of Video
Silicon Showdown: CPUs, GPUs, and VPUs Battle for Data Center Dominance
Hybrid cloud infrastructure has finally arrived for video streaming, and it's transforming how organizations balance performance, cost, and sovereignty requirements. In this enlightening conversation with Stefan Ideler from i3D.net, we explore how the gaming industry's early adoption of hybrid approaches now offers valuable lessons for video streaming platforms facing similar pressures.
Stefan reveals how i3D.net's massive global network, exceeding 30 terabits per second across 65 locations worldwide, provides the foundation for both gaming and video delivery services requiring ultra-low latency. Having started in 2004 before AWS even existed, i3D.net has evolved from gaming server provider to infrastructure powerhouse supporting companies like Discord through unprecedented scaling challenges.
What truly differentiates modern infrastructure providers isn't just their network footprint but their approach to customer partnerships. "You talk to humans, to engineers directly," Stefan explains. "We're in a Slack channel together, directly talking to your engineers, giving insights, sharing dashboards. We're really part of your operational team." This collaborative approach, combined with world-class anti-DDoS protection and network quality, creates value beyond pure cost considerations.
The conversation oscillates into the evolving silicon landscape within data centers, where specialized accelerators increasingly complement CPUs and GPUs. Stefan predicts that within five years, at least one-third of processing capacity will shift to AI inferencing chipsets, while VPUs continue growing in importance for video workloads. As power constraints become critical, especially in Europe, these efficiency gains aren't just economically advantageous - they're operationally essential.
For organizations struggling with cloud costs or sovereignty requirements, pragmatic migration paths exist. Through direct connectivity between cloud providers and bare metal infrastructure, companies can gradually shift static workloads while maintaining service continuity. The key is collaboration between engineering teams willing to develop innovative solutions together.
Catch i3D.net at the NETINT VPU ecosystem booth at IBC to learn more about their infrastructure solutions and how they're helping shape the future of video delivery.
Stay tuned for more in-depth insights on video technology, trends, and practical applications. Subscribe to Voices of Video: Inside the Tech for exclusive, hands-on knowledge from the experts. For more resources, visit Voices of Video.
voices of video. Voices of video. The voices of video voices of video well, welcome back to another amazing episode of voices of video. I can call it amazing because I'm the host, mark Donegan. Hey, we really do appreciate all of our listeners, all of our viewers, our audience, coming back week after week and listening to these wonderful interviews. Today I am joined again by Stefan Eidler from i3dnet. Stefan, welcome back to Voices of Video.
Stefan Ideler:Thank you, mark, it's great to be back again.
Mark Donnigan:Yeah, well, we've had some great conversations, you know, both on this show and, of course you know, at various conferences and trade shows, of which I'm going to go ahead and plug the fact that you are going to be in the NetEnt VPU ecosystem booth at IBC. It's going to be amazing.
Stefan Ideler:I look forward to talk and meet with everyone.
Mark Donnigan:Yeah, that's right. That's right. Well, let's start this conversation with. You know, hybrid cloud has finally arrived for video. For a long time, you know, vendors and implementers have been talking about the benefits, or even the dream of being able to flex into the cloud, maybe primarily run on on-prem, you know, in an own data center, and then flex Sometimes the other way. You know, run primarily in the cloud but flex to on-prem. For certain, you know applications or use cases, but it's real and it's here and you guys run a massive network. In fact, I think it was like 26, 28, 30 terabits per second, but now it's bigger, right?
Stefan Ideler:It's already certified plus, and every time we talk it's already more, because the demand side keeps going and going, yeah, yeah.
Mark Donnigan:Well, so let's start our conversation here. Well, so let's start our conversation here. Tell us you know what you guys are doing to enable hybrid cloud because you are an infrastructure provider, so you've got an excellent perspective here. Tell us what you're doing.
Stefan Ideler:Of course.
Stefan Ideler:So originally we come, of course, from the games industry, where we see the movement towards hybrid a little bit earlier than the general enterprise market.
Stefan Ideler:As many know, the past two years, especially the last year, you see that the games industry is under pressure.
Stefan Ideler:There's a lot of focus on cost and the COVID bubble of investments has burst. So we see many, many game companies looking at the bottom line and, like that, thinking what the hell are my teams doing with the cloud? So we see this now spreading to the whole enterprise market in a lot of the tenders that we also work with. Also in the video industry, there's now more and more focus on the bottom line. So hybrid solutions using cloud for the flexibility, which is absolutely perfect, but then getting your costs a little bit more control with that baseline of predictable capacity that you can put in your own data centers or someone else's data center and so on, and also mainly in Europe in particular, we also see a strong push towards the sovereignty aspects, where many organizations are quite worried due to the current political and economic realities. So that also is a major, major push for hybrid and even some on-prem solutions in all of the projects and clients that participate in the last six months actually.
Mark Donnigan:Yeah, yeah, yeah, we're hearing and seeing that as well. You know a lot of government regulation. You know we can talk about the EU, which obviously is where you are, but it's not just in the EU, it's really almost everywhere. You know there's quite stringent, you know requirements now, or at least that people are anticipating, so the sovereignty is an issue. So you know, stefan, I'm going to assume that a lot of people by now have heard of i3D. They know your roots, but I don't think everybody has. So maybe you should give a little one minute, two minute how you know, what are the roots of the company. You did mention you started in gaming, but you do a lot. It's very impressive what you built and who you power. So maybe you should let everyone know, first of all, where you come from.
Stefan Ideler:So it all started around 2004 in primarily a business-to-consumer operation where we provided game server capacity that people could play together. Well, the games industry evolved yeah 2004. That's like early yeah, before there was cloud and people were a little bit talking about virtualization and you had single core processes.
Mark Donnigan:AWS was founded only in what 2006, right, exactly so wow, that's amazing.
Stefan Ideler:And of course, the industry evolved. It turned into the B2B business where the game companies and publishers want to take the games under their own control, including the online lifecycle, and that's how we expanded. So we see ourselves as experts in being the partner of someone who needs to scale. Games at release is always a scaling problem. You never know how the game is going to perform, so you want to work with a partner who can scale with you worldwide and all of these locations. This has now also expanded to the voice and video communications and the general entertainment industry. We also assist with the scaling as parties like Discord, especially during the COVID days. That was some pretty exciting stuff.
Mark Donnigan:Yeah, I can imagine Discord video. Traffic must have just been exploding is an understatement, I would assume.
Stefan Ideler:And now we fix and solve the scaling problem in roughly 65 locations across the world with our own global network, which is, I must say, from my perspective, one of the best networks in the world in terms of connectivity and latency.
Mark Donnigan:Yeah, amazing, amazing. Now. So you've been doing gaming for 20, 21 years or so now more than 20 years but you've been focused on video or not focused, because you haven't left gaming but you've been also powering video traffic for the same period of time and you just didn't really talk about it, or is that kind of a newer capability?
Stefan Ideler:It is something about the last 10 years. I mean, we had some clients of individual industry but we didn't really really actively focus on it. But now we've realized that a lot of the things which matter in gaming good connectivity, no packet loss, stable connection, great insight to what's happening to your users within your network, those performance metrics and so on actually also matters for people in the video and broadcasting industry. Especially if there's also interactiveness involved with the user, it becomes really important to know what's going on.
Mark Donnigan:Yeah, so we're going to get a little more technical and into the infrastructure side and talk about where you see ASICs going and the role of the GPU and CPU and all that. I'm not going to name other vendors, but you know, let's just say I'll use in air quotes the traditional CDN, because it strikes me that obviously you are a CDN, your content distribution network, but there's kind of more to the, to the story, to the package. You know the, the CDN, and then you know also how do you position yourself against. I'll just say the public clouds that are that also. You know you can look at them. And then you know also how do you position yourself against. I'll just say the public clouds that are that also. You know you can look at them and say, well, you know they operate a CDN as well and you know they just happen to have a whole lot of compute hanging around. So how do you position yourself? And you know, where do people? You know, might someone want to choose i3D over one of those other platform types?
Stefan Ideler:That's a good question. To me, it's a combination of different factors. If it was purely a price discussion, then you have no right to exist, right. You need to add something that people also perceive as value, that other parties don't have or don't have as good, and in our opinion, it's a number of things. One is, of course, a resilient focus on that network quality, which was always very important for the gaming and voice and video communications that we do. That, we believe, also is very important for this sector.
Stefan Ideler:The second component, which also originates from the game industry originally, is the world-class NTDDoS services that we provide worldwide on top of our services. And the last one is definitely the support. You talk to the humans, to the engineers, directly. There's no five layers of ticket system. We're in a Slack channel together, directly talking to your engineers, to directly give feedback on what is happening, giving insights, sharing dashboards. We're really part of your operational team and that's so we try to make success together and that, I believe those three things is definitely a value perception which I think is worth going from a traditional provider or cloud to one's own, and, of course, it's a bonus if you also save money right.
Mark Donnigan:Yeah, of course, Of course. Well, it's. It's interesting. We also work with gaming companies. In fact, that's initially how we even got introduced. There's some very large platforms that run on i3D and they're wanting to do things with VPU. But one of the things that I think for those who come from a video background and haven't worked much around streaming of games, so in other words, like cloud gaming, I'm not talking about like object delivery, but is that those platforms are largely bespoke. And, yes, there's some commonality. Certainly there's some architecture commonality, but I just have seen that. You know, when you talk about needing to work with a client, work with a customer, like in the gaming world, that's just required. It's not an option that you could just bring a platform and say here's my documentation, Let us know if you have troubles.
Stefan Ideler:Take off the glove, rub your sleeves and really.
Mark Donnigan:You have to. Yeah, yeah, yeah, yeah, so. So I just make that point, because when you say that you work very closely, you know, with with your customers, with your clients, that comes from your roots of building these. You know very sophisticated gaming platforms, so yeah, well, let's you know very sophisticated gaming platforms. So yeah, well, let's you know. Let's talk about where, as a infrastructure operator, and an infrastructure operator at massive scale, so what is the role, what have you seen the role of CPU, gpu and ASICs been and you know we can keep it, of course, kind of focused more on the video processing, video encoding side, but also I would like you to comment on just even the network operations side. You know where you're seeing, you know some of those technologies, maybe expand.
Stefan Ideler:Oh yes, so actually I see it on a number of layers right. So on the network side of things, on the core routers, I mean, if you look back 20 years ago and now there's such an evolution in chipsets, dedicated custom-baked silicon to to FPGAs, to also normal ASICs again, but it's always focused on dedicated resources in hardware at scale, and now it's funny that we also see it in the server components happening as well. Like, you normally have your traditional network card right, A gigabit, a 10 gigabit, 25 gigabit or even 100 gig card. But in some of our services, especially at Antodidos, for example, we really need to go really deep into the packet layer, into the inspection layer, and for that as well we've moved to A6 and even some FPGA-based cards on the network side of things, Similarly what we see in the ai offering right. So, um, yes, you can go with the famous brand from nvidia and always buy the ferrari for everything that you do exactly and 100, it will work for sure.
Stefan Ideler:but um, if all you, if your model is trained right and all you're going to do is inferencing, then it might make sense to go for again a specialized ASIC chipset who is awesome, for example, at inferencing, workload and model switching and doesn't need all the other stuff that you're not going to use, which in turn, in theory, substantially lowers your cost per token. So I'm seeing, in general, many components and also, of course, with 3BUs, you specialize certain components If it matches the workload that you do operationally the majority of the time, and then it makes a lot of sense to to to go with specialized chipsets.
Mark Donnigan:Yeah, is your. You know I'm curious, um, your choices of, uh, silicon support or actually a large language model support, you know? So we're talking here in the context of accelerators, so dedicated silicon. I know that there's a number of manufacturers out there that are focused on Lama. You know there's others that are optimized around Anthropic and, by definition, you kind of have to pick your foundational model that you're going to accelerate and then you build silicon for it. I'm just curious are you seeing or hearing more requests for any particular foundational model or technology, or do you have to support them all?
Stefan Ideler:Pretty much the latter. We have to support a lot of them and what's really interesting for us is that there are more and more chipsets which align with that idea. We say we are good with most of the models, even the small ones, to some of the big ones, and we're really good at switching them. Let's say, if you have an agentic workload which puts a certain query through multiple different models to get to the end result, I believe we're going to see more and more of these kinds of workloads. There's also another client that we work with who is active in the trust and safety space on voice moderation and as well they have the user voice sample. It goes to different models after each other based on the complexity of the race. So actually, if you can switch really quickly with your existing chipsets, you don't have to keep them running all at the same time on multiple boxes, but you can just process. So that's some of the innovations that I'm now also seeing in this space. That makes sense.
Mark Donnigan:It makes a lot of sense because, it'd be safe to say, I'm dabbling in building my own agents, and mostly to bring efficiency to common marketing tasks. So you know nothing like like what you're building. But even in that very, very, very, very simplistic way I've I've learned I learned very quickly that when you can run one model and even a particular version of that model, you know and really optimize a prompt and optimize it, and then you have a different model for some other function. When you bring it together you get a really powerful like. To me, that's where these workflows really start to become pretty breathtaking. The foundational models are amazing by themselves. The fact that you can just write a fairly simple prompt and say, give me research on X, y and Z and you know, go here, go here, go there, and then what you get back is a 15 page like incredible paper. I mean that's amazing. But when you build these agentic workflows and you optimize the prompts and you apply different models, wow, that's when it's like the X10 acceleration.
Stefan Ideler:Because you're doing this, you're already ahead of 99.9% of the general AI users.
Mark Donnigan:And that is where the difference comes. Yeah, most people are still going to chat GPT and saying write me a blog post. Yeah, yeah, incredible, well, so, okay. So you're doing? You know you're running this large network. It's a CDN, but it's much more than that, because you have compute hanging off of it. Is this built like an edge? Everybody's talking about the edge, right, you know. So is this built like an edge network, or is it still largely consolidated into a smaller number of data centers? Or what does the actual architecture look like? That's part one of the question, and then, after you answer it, I'll let you know part two.
Stefan Ideler:So originally, like 50 years ago when we started building out the network yes, you start out somewhere right, so you start out with a concentration in Europe.
Mark Donnigan:US and one in.
Stefan Ideler:Asia, but nowadays, in order to serve all those very demanding gamers and all those voice and video users as close as possible.
Stefan Ideler:It's definitely an edge network, all the 60 locations there's a few bigger than the others, as always, you cannot avoid that.
Stefan Ideler:But also we want to make sure that all of them have sufficient capacity to serve the needs and also that if one location would for any reason fail natural disaster whatever that, the impact or the blast radius is sufficiently minimized because there's so many other locations in the region. I believe that that's definitely the way forward. But what you also need though not in all use cases, like if it's a one-directional CDN, so if you're just watching video but there's nothing interactiveness, it's fine. But if it's interactivity or you're in a shared room with users together watching the stream and talking and everything, then the connections between all of those edges to each other also starts to matter, because then you're having traffic that goes from one location to another location and if you have no insight over it, it goes. You have a party, you have no control. There can be so many things going on. So that's what the that's what differentiates our network a little bit more compared to a general uh CDN based network only focused on one direction of traffic.
Mark Donnigan:Yeah, yeah, and I think you know that's also another interesting thing about what you bring to market. This stands out, in my mind at least, is that, as if cloud gaming is not hard enough, you're powering this ultra low latency gaming experience, which I guess you could say that's kind of the ultimate test of interactivity, right, because you know, when I move my controller, it better move on screen, you know, and not be laggy.
Stefan Ideler:Exactly. And also, all the gamers know to find social media pitchforks. A lot of them are tech-savvy enough to do a trace route. So it's not just that, the picture your clients know they find us. Message me on LinkedIn. It goes pretty far.
Mark Donnigan:Yeah.
Stefan Ideler:But it also always allows us to get better.
Mark Donnigan:Yeah, yeah.
Stefan Ideler:For demanding users. It always helps to get you to the next level.
Mark Donnigan:But on top of that gameplay, though, you also have a gamer, who then is, you know, posting on social media or live streaming to Twitch, or you know whatever I mean. In other words, not only do you have to keep that game session running, and game is monetization.
Mark Donnigan:Yeah, exactly, exactly. So, wow, that's, there's a lot going on. Well, part two of my question is this how, then, are you guys thinking about video encoding and you know, video processing, now that you have an edge network? Now that you have an edge network, because it opens up a lot of really interesting possibilities. So how have you built, or how are you building, your video solution, given the fact that you have these?
Stefan Ideler:It's more than more compelling right. So being able to offer that close to all of the end users and if you have an application which can benefit from a VPU, it's a no-brainer to do it. Of course, just because of GPUs, not all workloads will fit in a GPU, just a GPU, but if you have one which makes sense, it's a no-brainer. And also looking into with my team of solution architects is both GPU on demand and VPUs on demand, so you can rent flexible capacity and, when needed, you add a GPU over IP, for example, or even a VPU over IP, to have that ultimate flexibility and also control of your custom resource, because the most expensive component of compute nowadays is, of course, that graphical or additional CPU component and if we can enable ways for users to nowadays is, of course, that graphical or additional CPU component and if we can enable ways for users to use it more efficiently, that widens the gap with of course, the PCIe, Absolutely so.
Mark Donnigan:In the data center, are you using some sort of a composable technology that would allow you to have sort of a virtual PCIe connection?
Stefan Ideler:Yeah, I mean there's that we're playing with it. Keep you side of things. There are some more mature solutions nowadays, but we see the same for the GP over IP. It might not be the ideal solution if you're not doing game streaming, I mean that just benefits from having a card right into the box. But for AI inferencing, query loads, or if you just need to pass a check that you have a GPU for your build processes, for your build pipelines, I believe it makes a lot of sense, and similar for VPU workloads. I think there are definitely workloads that would make a lot of sense to use it like this as well, and I think in the next coming years we will see more and more dedicated component cards, as I like to tell them, which are hyper-focused in specific tasks that might benefit your application.
Mark Donnigan:Yeah, amazing. Well, so it's wonderful to be an infrastructure provider. You know, and I, this is greatly simplified, but you know like we need roads, you know, drive the cars on and we want those roads to be smooth. And a couple, a couple of weeks ago, I took everything. Yeah, yeah, a couple of weeks ago I took my, my, my, my son, who's 19. And, yeah, a couple of weeks ago I took my son, who's 19, and you know he's busy doing his thing, and I said, okay, hey, take the weekend off. You know you need to spend some time with your dad.
Mark Donnigan:So we went on a little road trip up into the mountains, get out of the heat. Roads that you know you're cruising along 70, 75 miles an hour, got the radio going, it's like you barely hear anything. And then you know, all of a sudden it just, you know it's like man, I felt like I was like driving over I don't know, rocks. You know it's so rough and so, anyway, so you know, as an infrastructure provider, it's critical to have the smooth roads, right, but at the end of the day, you know, people, you know, need a car to drive on the road, and so this is the analogy is often we describe that the VPU is like an engine, you know, and an engine of course goes in a car and propels the car to drive over the road, you know, which would be like the network, but you need the car.
Mark Donnigan:So I'm curious how you guys are addressing or solving or thinking about as a network provider, infrastructure provider. You know, how are you approaching this when people come and say this is amazing, I want that smooth road, I want the wide lane, you know all the goodness right that you bring, but I need a car. You know I, you know I don't have the engineering team or the money or the time to go build it. How are you handling that? Are you doing partnerships with software ISVs that are then deploying on i3D, or are you building it yourself? Or what are you doing?
Stefan Ideler:Well, I mean, if I had a limited resource, of course I would build it myself.
Mark Donnigan:Yeah, of course I know you're an engineer and you're a builder?
Stefan Ideler:That's a really good question. We went through the same evolution in the games industry as well, where we know what we do best having the solar infrastructure foundation inside and the needles and so on. But let's say you need a matchmaker, leaderboards and everything. We have a number of partners that we work with very closely and we are doing the same also now on the video and broadcasting industry, so we can give you that engine with the nice flat roads, no potholes. But if you don't have the other components yourself, we can recommend you quite a few number of partners that we trust in, we're close to with that. We also put their teams on Slack. To keep the lines short, we believe strongly in partnerships and also, if you're watching this and you'll be interested in also joining our partner program, of course, also reach out to me.
Mark Donnigan:Absolutely yeah, reach out to Stefan for sure. Yeah, yeah, yeah, yeah. And I ask this question because the challenge, I think, in video is that a lot of platforms were built from sort of one or two foundations, one of two foundations. One is just a collection of open source projects of which you know by now are incredibly mature. You know FFmpeg x264, x265, you know, and there's a range of other projects x264, x265, you know, and there's a range of other projects but they're built that way on general purpose compute. So it worked really well on the cloud and you have a small engineering team and over time maybe the team got bigger, but you know you build it, you build it, you build it.
Mark Donnigan:The problem is is that you're never going to have enough engineers. You know, I learned this early on in the industry, where you know I used to assume, you know I won't say names, but you know you'd approach some very large, you know, media and entertainment company or a large cable company or a large. You know somebody and you know and I approached them with a perspective of like you have this amazing engineering team, who am I to even try? And I remember early on learning, you know they look back and went yes, it's true, we have very good engineering team. Yes, it's true, we have hundreds and hundreds or even thousands of engineers working on our platform. However, nobody is sitting there waiting for something to do. Everybody is fully utilized.
Mark Donnigan:And then some, you know, and and so so a lot of the platforms have been built that way. Or you know, there was sort of the jumpstart of like hey, you know, we can utilize maybe some microservices on a public cloud, and over time those microservices became even more capable, more integrated, and eventually you sort of could just buy, you know, maybe not 100% of what you need, but 80%. And then you, you know, you know you buy or you engineer the other 20%. So the challenge then that we run into and this extends to even to like I3D is people say, wake up one day and say oh, wow, I can no longer justify, or I just simply don't have the funding for the big engineering team for my in-house solution. So what do I do? Or, on the other side, you know, it's a similar economics problem. Like my CFO has said, we've got to cut by 30% and there's just no more to cut. We need to change platforms.
Stefan Ideler:So it's very hard when you're 80% relatively locked right.
Mark Donnigan:Yeah, yeah, yeah, yeah, so you know. So I'm curious if you have some insights. Maybe even you know, you can share a little bit of firsthand experiences of walking alongside customers who have faced, you know, this same dilemma. They know they need to move off of whatever their current platform is, whether it's in-house, kind of custom built, or it's somebody else's platform, or it's just a third party commercial platform that's just gotten too expensive, you know, and um, now.
Stefan Ideler:I think we have some great analogies for that from some recent cases. We did so, it course, in the games industry, but exactly the same problem. That was, of course, a problem where people were having on-prem, but it was only themselves as the client right. So games scaling up and down left either a big overcapacity of hardware or under-capacity and players could not join. So we actually helped them migrate towards cloud I know it might sound strange nowadays, but still keeping in mind and helping them to migrate in an agnostic way. So you would just, for example, create containerized workloads and make sure you can place them in any cloud provider so that you're not stuck in. We've also seen the other way around, where we see, actually quite often, where people went to the cloud because it's extremely convenient, extremely accessible and especially when it's small the credits and everything it's small, it's inexpensive, it really is so yeah, I don't blame anyone for that.
Stefan Ideler:However, um, as you say, when the cfo might knock on the door and like, oh, what the hell is going on? Now we're successful, but a large part of my potential profit is directly to the cloud provider. So for that we look pragmatically. What can we do? So in the gaming sphere? With a number of solutions.
Stefan Ideler:We created connectivity between the cloud provider and us, either via tunnels or direct connect, and started identifying the components which are static and we migrated those towards the bare metal infrastructure, while keeping the connections to the cloud provider intact, so that infrastructure at our data center, which is only one millisecond away from the cloud data center, can keep talking to all the internal cloud services that it's still logged into and can work with. And this allows piece-by-piece migration paths to get at least a certain percentage of that 80% toward a base layer. Yes, you still have some traffic costs between the cloud data center and our data centers, but it's a start to start to get that that info up your back essentially. But it does allow. It does require some trust. It allows, it requires our engineers and your engineers in that slack channel talking pragmatically, coming up with brilliant solutions, how to make that connection. That's right. I believe it's always possible yeah, yeah, it really is.
Mark Donnigan:And uh, you know I hate to simplify it, but you know, where there's a will, there is a way. Um, and you, you know, look um five years ago, six, seven years ago. Cfos are maybe dreaming or whispering about doing this, but it wasn't quite as urgent and we were in maybe a little bit of a different financing phase for a lot of startups and money was certainly cheaper. There's a whole lot of reasons why people talked about reducing costs, but it never really drove decisions. Now it is. You know, I mean, you look at it and you know it's very sad for the recent Microsoft employees, for example. Here Microsoft is posting record profits and yet I think they just announced was it 10 000 job cuts or something.
Stefan Ideler:I mean a lot in the game side of things as well yeah, yeah and and and.
Mark Donnigan:Look, you know I, if I were a microsoft employee, and certainly if I were affected, I'd really be scratching my head going guys, you know, like I, I see the stock price. I listened to the, you know to the earnings report. What's going on here? But the reality is is now, you know, companies are being really brutal and I think that's the right word and where, how they're deciding where they're going to allocate capital. And you know, if I put my business hat on and you know, look, you're a I, in fact you're a co-founder right Of i3d. So you know. So, look you.
Mark Donnigan:You know you also have to put your business hat on at times and look at decisions and say what's economically prudent and and optimal for our shareholders, for the employees that own a part of the company, et cetera, et cetera. So, um, I think you know, and I think the industry now has come to terms with us uh, is that cost reduction, becoming more efficient, doing more with less, is not just a phase that we're passing through? And then you know, we're going to be able to go back to spending money like it's, you know, like it's nothing. This is the reality. And for those organizations who get efficient, who get you know fit, you know who go to the gym. They're going to thrive. They're going to really do well.
Stefan Ideler:Especially now you have this layered numbers of issues pushing this.
Mark Donnigan:That's right.
Stefan Ideler:Interest rates. Because there was too much capital invested in the games industry, a lot blew up. We see this now also happening in other industries. And then there's the AI crunch a lot of do-op we see this now also happening in other industries. And then there's the AI crunch. Like you get that push for efficiency to see if you can work more efficiently using AI and, yes, there's some fluff. Right, ais cannot code a complete product. You still need diligent software engineers to double-check if that's really actual good code right, but there are so many mundane tasks that you can already automate away, so it's kind of like a triple crunch. That is that is right, really cutting on on on people's jobs.
Mark Donnigan:Uh, it's well, there's one. There's one that you didn't explicitly name, but I know that you viscerally feel, and that is you can't get power in Europe right To your data centers.
Stefan Ideler:I was hoping to avoid that subject.
Mark Donnigan:And no, but. And by the way, I mean I'm saying Europe, you know you're in the Netherlands, but but, but it's true, I mean you know. You just look at, you know we, we leave politics out of, you know, but look at what our president here in the US is talking about. I mean it's power, power, power, power, power power. We need more power. We need, you know we need more power, and you know all kinds of debates on how we can get that power and whatever, and that's not the purpose. But the fact of the matter is we need power to our data centers, and so that means that any efficiency gain that we can bring operational efficiency now I can bring in those extra GPUs or I can bring in that extra infrastructure to expand my network. Otherwise, even if I have the money, the capital, I can't power it.
Stefan Ideler:I don't have the energy, especially here in Europe, like in the Netherlands. They say it might be fixed. Maybe in 10 years from now you can get some more capacity, so there's a lot of innovative thinking. We have other places in Europe where, fortunately, there was a lot of power capacity left, so there's plenty of room still for us to operate. But it's starting to hurt and that's why, if you can use solutions like specialized AC chipsets, like the VPUs or even the AI chipsets, do the tasks that you need for your business a lot more efficiently, instead of buying 10 Ferraris who'd consume 100x the power. That's right.
Stefan Ideler:That's making sense and it all adds up to, to, to, to the whole equation, yeah.
Mark Donnigan:Yeah, yeah, yeah, it's, it's. It's a great discussion. I really appreciate your perspective because I think, also, you know what I've observed working in the video space for for a long time many, many years, is it and I want to be careful how I say it because I don't want to imply that people didn't care about, like energy consumption or environmental impact. That's not at all what I'm saying, but it's almost like the ball was passed to the network provider or to the public cloud, like it's their problem. You know, and, and I can recall being in conversations and not so much recently, but you know, five, six years ago where you know I was involved in, in, in some technologies where one of the value propositions was, you know, energy reduction, and, and you could pull it. It wasn't um at that time, it was before.
Mark Donnigan:Uh, you know, I was working with NetEnt, but it still had a similar value proposition and people would quite candidly or sort of, you know, like off on the side, say, hey, look, you know I can't really say this publicly, but like you know, like we just assume AWS has taken care of this. You know, and, and the fact of the matter is, is that the buck has to stop somewhere, you know, and and so now no longer can the industry sort of assume that, well, I just leave it on my public cloud to do that. Because guess what, if anything, it does translate into increased cost. Because if that data center provider is having a hard time getting energy, well, they've got to light up their data center somehow. That means that their costs are going to go up, however they do it, which is going to get transferred along, you know.
Stefan Ideler:So, even setting aside the environmental concerns and all of that, you know it's it's very exciting work that we're seeing both in the on-prem data centers, but also in the cloud provider networks and cloud provider offerings and also many SaaS and PaaS offerings. Now, every year it's way inflation plus a lot that prices go up with.
Mark Donnigan:Yeah, yeah, yeah.
Stefan Ideler:More and more pressure.
Mark Donnigan:Yeah, absolutely. Well, this is a great conversation. Let's wrap it up here with this question, and you know it's a little bit of a crystal ball question. You know like forecast the future, but not really In fact I don't want you to forecast the future, I want you to answer the question based on kind of what you know today and what you see today. But I am curious about the mix of architectures or technologies that you think will be in i3D. You know, let's say over the next, you know two to three years, but let's say like three years. So you know there's CPU, obviously for game engines, rendering the GPU is not going anywhere. And now inferencing, you know that's very, I think. I think there's a huge potential to move that onto dedicated Silicon and there's probably other aspects of the AI. You know workflow, tool chain, et cetera, and then for video processing we have BPU, but I'm just curious like tool chain et cetera. And then for video processing we have VPU, but I'm just curious maybe how you see those splitting out in terms of percentage.
Stefan Ideler:That's a good question. If I look at percentage and then if I look basically not on the actual numbers but in terms of value, like cost, it's more fair, I think. But now basically not on the actual numbers but in terms of value, like cost, it's more fair, I think. But now of course the majority is CPU and then traditional. Gpu, Of course yeah, what I see, definitely agree with you that, let's say, five years from now, I think I would say probably at least a third are going to be inferencing chipsets.
Mark Donnigan:So like accelerated basically ASICs that are accelerating these large language model inferencing.
Stefan Ideler:In my opinion from a technology expert, I know every game server component in the next few years will start needing some local inferencing capacity because they're also going to incorporate lots of models. All your agentic models will need it, preferably where they are and preferably locally if possible. It's going to be part of every single business process. So I think that will see a major, major growth and people will also figure out that it needs to be cost optimized so it doesn't need to buy the Ferraris Now. You need specialized devices that are really really good at just these things. Similar to that. I see the same for VPUs, but in our case also expect more and more on the network cards as well, to get a lot more intelligence there, mainly because of the need to protect everything.
Mark Donnigan:That's fascinating that you say so. I want to ask you for a moment to elaborate on the network card having more intelligence built in. Do you see it going all the way to even some lightweight inference engine support in the network cards, or what sort of intelligence would be there?
Stefan Ideler:So what you're saying now could be definitely the future use case. What is our current use case is that to protect everything that we have everywhere, we have locally, in every single site, capacity which can, within 0.1 millisecond, appear in line to all of your traffic. And then it goes through all these big networking cards and all these big Linux servers, right. But we want to then look into every single packet to make sure that your byte matching matches the whitelisting matches and everything else. But for that we rely a lot on these network cards and the grammaticality of these network cards to push as much traffic from dirty traffic towards clean traffic. And that for us, specifically for our use case, is a big, big, growing part of our infrastructure.
Stefan Ideler:It might not apply to all providers or general providers, but that's why I also said in the other category, on the networking side, that's probably going to evolve further, because once you have all that extremely quick specialized programability on those network cards as well, it opens up also many other things, yeah, to do, to explore and think about. Yeah, um, so uh, for me, inferencing cards, basic dedicated inference cars, probably on number one, closely followed by vpus, I think traditional gpus, as in non-ai based gpus might be stable or decline a bit, and then, if the cpus, which of course you have, the holy war between x86 and ARM I wouldn't dare to put a bet on which will win at that point.
Mark Donnigan:Do you have a lot of ARM in your network or is it all x86?
Stefan Ideler:So for the traditional gaming use case it was all x86. It's much gigahertz per core as you can, because we don't want to spend any time on multithreading and we just want to not optimize and just give me as much gigahertz, yeah, yeah, yeah, yeah.
Mark Donnigan:Raw horsepower, exactly, you know.
Stefan Ideler:so we would say American muscle, Like, if you're talking cars, you know it's the old, like 1968, you know like terrible gas mileage, they're noisy, they pollute the environment, but you know, but they're powerful yes, but also there we see a movement towards uh, these companies also realizing what, if I optimize this a little bit better, yeah, there's a tremendous amount of cost that I can also save, regardless if it's from cloud or from brand, because if I make an application more efficient, it lowers my costs per session. Right, correct. So that's why we're also seeing some ARM. We're seeing some applications improving the threading, which is why I think we'll probably be having more ARM in terms of percentage ratio five years from now.
Mark Donnigan:Interesting. Yeah, well, it is interesting that Ampere was acquired by SoftBank and rolled into ARM, which totally made sense. I mean, it was one of those like, well, yeah, is anybody surprised? But yeah, it was one of those like, well, yeah, is anybody surprised? But yeah, it'll be interesting.
Stefan Ideler:Note that the initial ARM demand to us also came because of the fact that at a certain cloud provider that Granatone ARM range is priced a lot more interesting compared to the x86.
Mark Donnigan:So that drove also the initial demand to us about arm and like yeah, and and then since then it's sprouted and and we see the receipt yeah, well, and you know, we, we we fully support arm and have, and I know we've, you know, uh, had conversations and I think may even be doing some work around that as well. But yeah, it's the architecture. I think one summary of the conversation we just had is that, first of all, I would say that network architectures not are changing, they have changed, not are changing, they have changed. You know. The only question is where are providers in the continuum of the change process? Is, you know, is kind of how I would look at it. Maybe that's too simple of a view, but I don't think you know, and if you look at again, I I earlier I use the term the traditional CDN, and that's not a derogatory thing, that's just. You know. There's companies that have been out there delivering bits in a very traditional way that was well-known, well-understood for 20 years, 25 years.
Stefan Ideler:Those use cases that might still be the best choice, right.
Mark Donnigan:Exactly, and that's a very good point. It's not that all of those applications and use cases have gone away, but the fact of the matter is, even those service providers, those video streaming services, are now changing and now they're starting to push their processing to the edge. You know, I know of some very, very large consumer services that are really, by all rights, operating a full edge network. You know, and you know it's interesting because still there's other parts of the business. It's like oh, the edge. Everybody's talking about the edge but it doesn't really exist. It's like, no, it very much exists and that's where things are going. So, yeah, this is great. Well, I want to just conclude with this and invite all of the listeners and the viewers for those of you that are actually watching this on video, if you're listening on a podcast platform, thank you very much.
Mark Donnigan:Stefan and a number of other folks from i3d will be in the NetEnt VPU ecosystem booth at IBC. Vpu ecosystem booth at IBC. We're literally right in between hall one and hall five, very easy to find. It's quite large. There's going to be a lot of excitement there. We have a number of partners there all talking not only about what they're doing with VPUs but also what they're doing in general. So if you want to just talk about i3D's network, what they can do for you for cloud gaming, et cetera, make sure that you stop by and see them. So, stefan, it was great having you on again. You're becoming a regular guest, you are doing absolutely fantastic, Mark.
Stefan Ideler:It's always a pleasure to be here. All right, Well good. Absolutely fantastic, Mark. It's always a pleasure to be here.
Mark Donnigan:All right, well, good, well, thank you again, and thank you to all the listeners. We really do value and cherish your support. If it weren't for you, we would be talking to ourselves. So thank you again and, as I like to conclude, happy encoding and until next time, be safe, be well. This episode of Voices of Video is brought to you by NetInt Technologies If you are looking for cutting-edge video encoding solutions check out NetInt's products at netintcom.
