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
Watts Up With Your Encoder? Akamai & Cires21 Benchmark VPUs vs. GPUs
Energy efficiency is quickly becoming the new battleground in video processing infrastructure, and a groundbreaking benchmark study has revealed just how dramatic the differences can be between competing technologies. In this eye-opening conversation, Chris Milstead from Akamai and Dennis Mungai from Cires21 share findings from their joint research comparing Video Processing Units (VPUs) to traditional GPUs for encoding workloads.
The benchmark shows that VPUs deliver 4.7X better energy efficiency than GPUs for video encoding, while maintaining equivalent quality.
Highlights from the study:
- VPUs consumed only 12 watts while running 19 simultaneous encoding jobs, compared to GPUs using 59 watts for 16 jobs.
- Custom silicon is back: modern VPUs offer both efficiency and flexibility.
- NVMe interface makes VPUs exceptionally easy to deploy in cloud environments with no special drivers.
- Predictable, linear scaling across resolutions and codecs enables precise capacity planning.
- Energy efficiency is critical as data centers face power constraints in many regions.
- Simplified deployment shortens the gap between R&D prototyping and production use.
- Benchmark used Netflix’s Meridian film to test both H.265/HEVC and AV1 encoding performance.
- Akamai now offers cloud instances with NetInt VPUs starting at $0.42/hour.
The results are staggering: VPUs achieved a 4.7X efficiency advantage while delivering equal quality. Running 19 simultaneous jobs at 12 watts versus a GPU’s 16 jobs at 59 watts is not just incremental - it signals a fundamental shift in how future video platforms can be architected. As Chris notes, with regions like the Netherlands halting new data center construction due to energy limits, these gains are becoming essential for continued growth.
What makes this discussion particularly valuable is the depth of technical insight. Dennis explains how the NVMe interface dramatically simplifies deployment, creating a “level of certainty with speed” that narrows the gap between prototype and production. Predictable scaling across codecs and resolutions means operators can plan capacity with confidence - something GPU-based systems can’t match. As Dennis puts it: “Cost savings is not the goal. It’s the outcome of systems so well designed that it becomes an inevitability.”
Whether you’re scaling a video platform, navigating data center power constraints, or simply looking to cut operational costs, this conversation offers crucial insights into how purpose-built silicon is reshaping the video processing landscape. Listen now to understand why custom ASICs are making a comeback - and how they might fit into your future infrastructure.
READ THE BENCHMARKING STUDY:
https://www.linode.com/blog/compute/benchmarking-vpus-and-gpus-for-media-workloads/
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. Voices of Video.
Dennis Mungai:Voices of Video.
Mark Donnigan:Well, welcome back to this exciting edition of Voices of Video. So I am joined today by two companies that don't really need any introduction Chris Milstead from Akamai. Welcome, chris.
Chris Milsted:Thank you very much, Mark. Everyone love his beer.
Mark Donnigan:Yeah, awesome, awesome, awesome. And, dennis, it's great to have you back. Dennis Mungai from Cirrus21, welcome, thank you very much. We, of course, had a nice interview leading up to NAB Dennis and I did. We talked about everything that Cirrus21 is doing is doing. So today we came together to talk about some recent benchmarking and a study that both companies have worked on, akamai and Cirrus21. Chris, why don't you give us a quick overview of what this study was comprised of and what you did? Yeah, thank you very much, mark.
Chris Milsted:The study was really trying to get a different angle and different look at kind of the traditional benchmarking and kind of approach that people take to transcoding and encoding and people tend to look at, you know, just basically raw power.
Chris Milsted:What I was really interested about and what we've had, you know, a number of companies kind of contact us about, is what are the other aspects they could look at. And one of the really hot ones at the moment is energy efficiency. And what we were interested in and what we're being asked about is I know there's some kind of political, you know, tides maybe against some of this green reporting and stuff, as as was maybe a couple of years ago, some of this green reporting and stuff as was maybe a couple of years ago. But companies are still interested in their scope three emissions and like how actually energy efficient and how actually, you know, bearing in mind all this GNI stuff that's going on and sucking all the electricity out, and we're starting to see, you know, six gigawatts at campuses involved being projected by some of these big hyperscalers how do we actually get better energy efficiency and how do we actually get more bang for our kilowatts when we're actually doing all these working and transcoding video.
Mark Donnigan:Yeah, energy efficiency. I know you're in Europe, dennis is working in Europe and other parts of the world as well through Cirrus 21. There is no doubt that, whether it's through which rightfully so we still need to compare about care about carbon emissions. So you know it is true that you know there's a little bit of a ebb and a flow in terms of the focus on, you know, net zero by 30 and those sorts of initiatives, but you know the environment does matter. I don't think anybody would disagree with that. Other side of that is just the cost and the fact that in many countries or regions in Europe you simply can't build data centers. You know, like in the Netherlands from what I understand, you know, talking to people who are running data centers in the Netherlands you can't build one. So if you need to increase capacity, you either need to find a more efficient way, you know, to get that computing done, you know, or you just simply can't grow, which is not good.
Chris Milsted:Yeah, exactly that, and I think you know whether it's to do with the reduction in carbon emissions it's. You might have a fixed power budget for your data center that's right and if you need to do transcoding, you might have to sacrifice, you know, some of the capacity you might not be able to replace. You know a high-powered GPU with an even higher-powered GPU to do this work. So what are your other options and what choices do you have in the market? So that was really the idea behind the study and the investigation.
Mark Donnigan:That was the idea. Yeah, that's great. Well, dennis, you know Cirrus21, I know, played a real pivotal role in supporting Akamai. Working with Akamai, of course, you know this is what you guys do is run video workflows. So how did you support Akamai in this process? And then let's get into what actually did you study and how did you approach the study?
Dennis Mungai:The approach we used with Akamai was default. First of all, provide the reference software stack that we can run the production grid benchmarks on using flows that are similar or mirror our encoding settings. The second and the third way was on collecting the necessary metrics and comparisons between the respective platforms, that is, for the NVIDIA GPU and for the NetInt card that's installed on one of the appliances, just analyzing the results in detail and also making sure that any edge cases we remain to clarify through either repeated tests or a further reanalysis of the same.
Mark Donnigan:What software Were you using? Cirrus 21 media processing framework, or were you using something a little bit more generic, one of the open source, or what were you using something a little bit more generic, one of the open source or?
Dennis Mungai:what were you operating we? Have tested some of these workflows with both 0.31 implementation and also for the reference pipelines. A bit of FFLPEG's Linux code integration. Yes.
Mark Donnigan:So I guess you answered the question, but I just want to make sure it's clear, even for the listeners. So were these you know I'll call them encoder level evaluations or were you really attempting to emulate, like here's, what the power consumption and of an end to end workflow would look like? How did you approach that?
Dennis Mungai:Our approach is actually taken from a very common workflow type we see, whereby you would want to take a very high-resolution content and specifically I mentioned the samples we used. We used Netflix's Meridian Film, which is a common benchmark for VMA scoring and transcoder benchmark evaluation, and then, taking that source, you downscale the content into various resolution ladders and then you measure the maximum simultaneous learners you can run per job, per resolution, per frame rate, until you saturate the respective GPU to the point where you no longer have real-time playback. And the main focus here is that we consider Atlassian Gold successful if you can maintain at or above real-time playback, because the target scenario here was also set to around 60 fps, which is very common for sports and other live broadcasting scenarios, so that even if you have lower frame rates for any of the ladders, you can then extrapolate, based on the available data, the kind of density you can run on each implementation side by side.
Mark Donnigan:Excellent, excellent, okay, great. So, chris, I'm not going to assume that all of our listeners know how Akamai is supporting VPUs. Listeners know how Akamai is supporting VPUs.
Chris Milsted:Maybe it would be helpful to back up and even just explain the architecture and the topology. Yeah, sure, I think, just a little bit of a background. So about three years ago now, akamai acquired a company called Linode. They were a kind of a cloud provider, one of the kind of mid-tier ones, I guess, to begin with. And I I think what we looked at internally was actually, you know, what were we spending on our cloud computing budget. So you know I think it's a it's a story we told as well within akamai that we've saved, you know, a very large amount of money by moving a lot of our services in-house. And one of the other reasons we're doing that is we thought, well, we have this cloud computing platform, we should start offering services on there and building on that Lino platform. So what we're doing now is, you know, you've seen a couple of years of, as that Lino acquisition has happened, we brought that onto the Akamai backbone. So you're seeing the full power of the Akamai network. So, you know, a content distribution network and a global reach in there.
Chris Milsted:And now we're starting to build kind of these specialist services on top of that. And we've got this great partnership with NetInt. So the NetInt cards, they're fantastic. They just fit into the PCI slots in the front of your servers. So literally we've got a bunch of servers in a bunch of data centers, all the cards, and then we basically work to offer plans that customers can just you know by the hour come along and consume.
Chris Milsted:We put out some initial plans. That's one thing I would say. So there's some initial kind of T-shirt sizes with the NetIn cards with, you know, various amounts of CP, ram and number of cards, I think for anyone who is interested. Yeah, I think we're still exploring you know exactly which data centers to next launch additional services in, and also we're open to feedback on exactly what those T-shirt sizes would be more desirable in the future. I think they're very nice to kind of cookie cutter out. But if there's other components that people want to do and kind of combine the NetInk cards and the CPU, we are getting some feedback that maybe we need to be a bit more flexible there. So you know, if anyone is interested, please get in touch and give us some feedback on that.
Mark Donnigan:Yeah, it's great, we really appreciate our partnership with Akamai and you know we're seeing tremendous interest from our side For several years. Kind of the number one question that we got and this was before we installed a lot of VPUs in the Akamai network in Linode the number one question was what cloud are you on? Because it is absolutely true that many, many, many platforms are looking to move of the usual suspects, shall we say but at the same time they're not necessarily wanting to go out and take on building, certainly building a data center, but even maybe taking on the management of a colo. You know so they're, you know they're looking for that like infrastructure, as a service, and I think that is so valuable in the industry, I believe is responding very positively to Akamai's offer. So I want to get back to you know some of the findings of the study and you know I'll let one of you jump in or both to talk about this 4.7 X, if I'm recalling correctly.
Mark Donnigan:Advantage, power advantage, power, efficiency, I guess, advantage of VPU over GPU. Anytime, you know, you see a big number. That is really impressive. I think it's natural to have one or two responses, you know, oh, wow, tell me more. Or you know, sort of eyebrows raised like how can it be? You know, that sounds a little too good to be true. So, um, what? What can you, you know, tell us about this 4.7 X advantage? You know, what does it actually mean? You know?
Chris Milsted:um, yeah, just explain that you're going to have a go at that, dennis, and then you jump in without getting anything modeled up absolutely that we can do, thanks.
Dennis Mungai:So our observations here and this is the part where we say sometimes apples to oranges can be a bit misleading due to context. What we observe with the latent cards is that you can run the same workload density to a degree not to a fault, but a degree as you would on the N? Nvidia cards for a fraction of the cost and with near matching vmaps course, even for the most recent codec types, specifically everyone and hivc. And yes, the power difference that we observe between the nvidia and the netode cards is actually correct. This is something we have run about a couple tens of times just to be sure. So that claim holds. That's amazing that claim holds.
Mark Donnigan:yeah, and I'm glad you referenced AV1 and HEVC because, you know, anytime we're talking about encoder studies, you know that's a very natural question. Well, which codecs did you use? You know, and which implementations. And also, just to be clear, and, by the way, we'll link up to this study so that people can download the paper, go to the blog post etc. Uh, in the show notes, um, but uh, correct me if I'm wrong, it was the rtx 4000 gpu, right, that was benchmarked the, yeah, the rtx 4000, 80 the generation yeah, yeah, okay, um, and and, and that's also important because you know.
Mark Donnigan:then the next natural question is well, which gpu, gpu did you compare? You know, compare with? And also, what's important is that that is a GPU that also is available in Linode and for certain applications, would be an excellent choice. Um, so I uh. You know one of the things you know, we're all engineers here. I'm, I'm the least technical of uh, of the from the two of you, um, but uh, but you know I am, I am a technical person, um, and so I always appreciate, and even in in marketing, you know, we're always um.
Mark Donnigan:You know, I tell my team and I think to myself. You know, like, okay, let's get to the ground truth, you know, let's make sure that we're not just sort of talking and hand waving, and you know, and so I like the fact that you benchmarked on the same platform. You know there's no tricks here. You know it's not like. You know running a GPU in some different environment, and then you know it's on the node, it's their instances that are available. Somebody could go out and replicate this study if they just want to prove it for themselves. Remind me again, what video set did you use?
Dennis Mungai:Specifically Netflix's Meridian. It's part of the open content platform where our video engineers can take samples and then use it for analyzing video pipeline flows, for performance and conformance.
Mark Donnigan:That's right, and the objective metrics was VMath. Did you also use PSNR or even SSIM?
Dennis Mungai:To an extent, our main focus was on initially VMUF, but you would be pleased to know that, as far as PSNR is confirmed, the results are hold. Yeah that's great.
Mark Donnigan:Yeah, that's great. And so what I'm pointing out out here again for the listeners, for some of our maybe slightly more skeptical engineers, which you know we do primarily, our audience, or engineers or engineering, you know, oriented. You know this, this, this was done, you know this is a proper study. So I think that you know you will find, if you go do your own independent analysis, that these results hold and there's not going to be some wild variance, which we've all been a part of, those right when you go fact check something, you're like, hmm, how come my numbers are so much different than what you know this company is claiming?
Chris Milsted:And I'll say that for Dennis as well. So we did, we did repeat, you know, with kind of gaps in the testing and stuff. So you know, the cloud instances we spun up were returned to the pool. So we definitely got different machines each times. And yeah, I saw with my own eyes over and over and over again, you know, at the very heavy workloads the N nvidia cards would peak in the low 80 watts I think. I saw maybe 12 watts as a maximum for the, for the um, the vpus, and I saw that over and over and over again. There was no real variance.
Mark Donnigan:I mean those numbers were pretty solid time after time after time so I have, uh, you know, just because, uh, my memory isn't what it used to be. So I wrote, somepu was able to run 19 jobs at 12 watts. The GPU ran 16 jobs, so a little bit less, but at 59 watts, so it's pretty significant. Yeah, yeah, yeah, it's amazing. So, you know, I think you know we can really dig into the numbers. And again, I would just encourage everybody to, you know, go download the paper, you know, read the blog post, you know, and even, if you want, you know, replicate some of this yourself, go study it yourself. But I'm curious for each of you, and I want both of you to answer this from your respective positions, you know. So, chris, from the Akamai perspective perspective, what surprises were there for you, you know, in running this study? And then, dennis, I'll you know, then you can answer. Next, from the cirrus 21, but, chris, you know, what were the any surprises? Was there anything that you're like whoa?
Chris Milsted:yeah, I mean, I think the massive surprise was just how low the power started and how low the power stayed, and I think you know that really demonstrates it's. You know, we seem to be going full circle now with we're back to designing custom silicon.
Mark Donnigan:Isn't that interesting. I'm glad you point that out, because 20 years ago that's all we had was custom silicon, you know, and then nobody wanted it.
Chris Milsted:And now yeah and we're wanted it. And now here we are and we're kind of going there and it's amazing when you look at you know quite the energy efficiency you can gain with that. And I think you know we also ran some other numbers around. You know what would be the cost per stream as well, and those are good numbers as well, depending on your resolution as well. Depending on your resolution.
Chris Milsted:So I was just like blown away and said, well, actually this is kind of a no-brainer for these.
Chris Milsted:You know these av1 hvc workloads if you've got the right workload, then this is just by far the best way to do.
Chris Milsted:And also, when you think about the emissions and the density and how you could get these into a day center, I mean, you know we're talking about people who want to put some of that I think it was. I think they're talking about megawatt racks or something absolutely insane in the future and I'm just and I'm looking at that going. This is far more sensible, air-cooled, much more sensible way to go. So I think it really just made me stop and think and re-evaluate kind of number of design decisions and just kind of play that off against you know kind of the AI market, what they're doing and thinking. You know there's got to be a better way to do things. So that was my kind of main original and you know I love the fact that they're available in the Akamai cloud and, you know, buy them by the hour if you need them, or as long as you want, and that's the nice thing about out that consumption-based computing as well.
Mark Donnigan:Yeah, by the way, starting at 42 cents an hour for the like you say there's t-shirt sizes, so that's the current. I think that's the small instance, but I mean it's just yeah, the economics are incredible. So, dennis, you know what surprises for you. So, dennis, you know what surprises for you.
Dennis Mungai:My initial surprise especially came around with, I would say, the little stuff.
Dennis Mungai:First, because, out of bias, I tested it last. Okay, and I did this specifically because I was trying to see how it would, how a card of this form factor and power draw would match to a flagship. And what we noticed is that, perhaps due to the fabric or the nature of its design, its level of parallelism as far as simultaneous real-time jobs are concerned, it's a very, very consistent number. That number doesn't dip, it doesn't clip, whereas with NVIDIA, despite having a lot of these GPO resources, the main limiting factor are its n-venc unions. You basically cannot offload a lot of this video processing past NVENC. Nvenc is what will handle the encoding and the output of the mixture. Sure, you can offload scaling and everything else to the GPU, but at the end the primary bottleneck is the number of NVENC engines you can have in a GPU like this, and that limit will hit you hard where you least expect it, because the expectation here would be that the NVIDIA GPU would naturally fly through the workload, but it doesn't mirror expectations. Yeah, especially at the resolutions you tested at.
Mark Donnigan:Yeah, yes, yeah. So thank you for pointing that out. There actually was some incredibly intentional, as you would expect, design behind how the ASIC scales in terms of the number of threads or jobs that are happening. Obviously, an ASIC is a different architecture. It's not correct to talk about, like CPU cores, which you know, a GPU, you know you kind of can think of it that way, in terms of just a tremendously high density of compute cores. But there is the notion of threads and one of the beautiful things about the way the Quadra is designed is that it scales. So this is something that everybody has to sort of first wrap their head around.
Mark Donnigan:We're all used to in software. You know, if we're running, let's say, x264, you're going to get the highest density, obviously because it's not as compute intensive of a codec that you know. Everybody knows that. Then you go to x265, hevc. Fortunately, x265 has many, many, many years of the open source community optimizing it, and so you know it's now maybe only, you know, 2x or 3x more compute intensive than x264, which really, when you consider the benefits of HEVC, is really not bad. In fact it's actually pretty amazing. But then you go to AV1 and I mean it's oh my gosh, it's 10X. You know the compute requirement. Even you knowVC or AV1, which means you're completely predictable. You're able to plan your capacity, plan your throughputs. You don't have to worry about the encoder falling down because, oh, my top profile actually have you know, I'm choosing to encode that at AV1,. You know, with AV1, and I've got a couple of HEVC and then the rest are H264. It's completely predictable and it scales on resolution. So a 1T1U will do ADP 30 streams.
Mark Donnigan:You say, well, what would it do for 4k? You literally just divide by four, it's eight. But GPU doesn't quite scale in a CPU, forget about it. You know, like you definitely don't just divide by, you have to go test it and you know you might find. You know you have to divide by, you know some higher numbers. So it's a real benefit, you know, to both the ASIC architecture and then the way that Quadra was designed.
Mark Donnigan:Well, I think you know what would be interesting is to talk about, you know, studies are great and data is important and data helps validate. You know technical decisions and obviously you know, chris, for Akamai, data is very important because you have customers coming and saying, hey, we've heard about these things, but show us, show us that the claims are real, and so you need that. But I'm curious what someone could do with this data. What does it mean? Because numbers are awesome Wow, 4.7 energy, x energy advantage.
Mark Donnigan:If I switch to VPUs On one level, you're like well, of course, what it means is you know they're going to save money, they're going to, you know, not be polluting the environment with as much CO2, et cetera. But what does it really mean? You know, dennis, maybe you're in a good position to comment on this. And then, chris, I'll let you give your perspective, because, dennis, you've been building software for many years, and so when you have a tool like a VPU, how is it changing the way that you're building systems, the way maybe customers are thinking about you know what you're building, their requests.
Dennis Mungai:What does it mean For us specifically? It has brought about possibly the fastest turnaround time between prototyping something in R&D and having it in prod. And this is because the whole VPU stack is very simple. You only deal with a single fabric, which is NVMe. You want to query the status of the device. It's NVMe. It's a very simple, plain NVMe namespace so that even when you port this to something like a containerized instance, provided the NVMe namespace is available, regardless of the privilege level you're running at, you will have the coder sessions in those instances.
Dennis Mungai:Then the second part here is also the function of a resource abstraction. You view an NVMe namespace as if it were a single object. So, regardless of the number of VPUs you've passed to a given container, you will still be able to independently allocate jobs to them without affecting the specific runtime. The specific runtime affecting what's happening on one VPU versus another, and this also stretches towards what you would call resource monitoring and allocation. This is for when you need jobs to failover from an available VPU to another, regardless of how the VPU instance was instituted, be it on a cloud or on-premise. The VPU instance was instituted, be it on a cloud or on-premise, provided the namespace is valid, you're going to have a limits code session and this is the kind of it's a level of certainty that does move some design decisions forward, and at a much faster pace than we would have anticipated.
Dennis Mungai:So a bit of context here, just to step out a bit from what you're talking about. For you to evaluate a GPU-based encoder, depending on the various models you have out here and this also goes beyond outside NVIDIA you're going to have to first of all separate the kind of environment you're on, because bare metal does not translate to the same thing as an MIT instance, even when it simulates bare metal. But a VPU is a VPU, it's a single slash dev NVMe and as long as that instance is available, regardless of how it was started up, you have that level of persistence in that deployment time that you asked for. So for us, in terms of design decisions, you move faster, but without breaking currents. With GPUs, you will need to not only prepare extra layers for provisioning with a lot of safety checks, but you also have to be aware that some of the pipelines that you deploy for tests in R&D will not be one-to-one mirrors in production. So with VPUs, we have moved the gap between certainty and predictability to certainty with speed.
Mark Donnigan:Yes, yeah, I really like that. You're pointing out certainty. You know, I think for anybody who's operating, even if it's a small video service or platform, you know certainty is, above all else, what we want and need. Yes, you want it to look good. Yes, you know there's a lot of other factors, but you want to know that the system is going to behave and operate the way that you predict it to the way it should designed it to not predict. So, yeah, that's really great.
Chris Milsted:Yeah, yeah yeah, yeah, just to echo on the back of that. I mean, I think I think me and dennis jumped on a webex or something when we were first agreeing to do this and I think after we'd found out you know which lab he was trying to dial in from to the machines, I think three minutes after I clicked the button it's like there you go, they're ready to go off. Let's get the pipeline on there and off we go. I just just that kind of fungibility around the cloud computing, cloud resources. I think that should add to the mix as well.
Mark Donnigan:Yeah, and I think what's also important to point out is there still is, you know, the walls coming down on this belief, but there still is a little bit of a belief of, okay, I hear you, asics are more efficient. I've seen the quality or I've heard from colleagues the quality is good, but it's hardware. It's not flexible. It's sort of that feeling of like hardware is like hardened, sort of that feeling of like hardware is like hardened, you know. And yeah, if you own a data center, great, use hardware, put it, put it in your data center. You know, and what you guys are pointing out in these, you know, in your observations, is that actually it's very VM and container friendly, you know, and it's engineered for that, in fact, and it's engineered to scale and all those things. So some of the mindsets can actually shift. Not that the mindsets were wrong, because other hardware that isn't built around NVMe, as you pointed out, dennis, you know it's one of the major advantages. You know our engineering team really struck gold when you know it's one of the major advantages. You know our engineering team really struck gold. When you know they literally just went well, why don't we just use the NVMe interface? You know, like every well, not every machine, but nowadays every modern machine supports it. You know, we plug the card in, whether that's through a PCIe electrical or mechanical interface, or through a U.2, or even now the new M.2, which we have a card coming out on a super, super small M.2 form factor and it just, it works, it's available, All the resources are there. There's no drivers, there's no OS compatibility, all of those things which other hardware really struggles with, you know, including GPUs a lot of times. So yeah, it's very interesting.
Mark Donnigan:This has been a great discussion. Thank you for sharing. You know I'd be slightly remiss if I didn't, you know, say, hey, what's what's coming? You know, you know I'll try and avoid. You know what's in the future? You know, are you tracking? But?
Chris Milsted:but seriously, I think you know, with IBC coming up, if we say anything ahead of IBC we will be shot. So we'll keep that quiet for the show, I think so.
Mark Donnigan:So I'm certainly not, I'm certainly not looking for any pre-announcements from akamai. Uh, absolutely not. Uh, I, I like you, chris, I want you to keep your job, so, so, um, but um, but, but from, uh, you know I, I want to phrase the question and I would love to get your thoughts around how systems are being built. You know we went from the era it is interesting, you point out, chris you know, 20 years ago. You know the original MPEG-2 encoders were silicon, because that was the only way to do it. You know MPEG-2, it was silicon, right, you know, dedicated ASICs.
Mark Donnigan:And then we went, you know, through this period of everything was CPU and then the you know, gpus, you know, started to. You know, and we really are welcoming others, you know, who are saying, hey, there's, there's an opportunity here to build very, very high quality silicon, you know, that is flexible, that's performant, that is high quality. You know that can bring a lot of advantages. And so what does this mean for how we're going to be building our video platforms in the future? What are your perspectives?
Chris Milsted:Well, I think, from the Akamai point of view, I think what we're looking at is how do we bring these devices closer to where they're needed? So we talked about the live streams and you talked about sport. I think at the moment, when you look at the Akamai kind of cloud, we have these kind of distinctions internally, where we have these kind of core sites where we're building out with, you know, all the kind of the object storage and all the kind of other core components you'd expect. But of course those are big investments, so they're going to be in a limited number of facilities because you're going to need a lot of energy density in them, you're going to need a lot of power and cooling. So then we've got these other models where we've got, like you know, down to our content distribution network and a kind of our edge sites and in between we're looking at kind of this hybrid model, kind of distributed site, which is kind of how do we you know, for example, if you've got a live sports feed that's from a country that doesn't have one of these very large data centers, could we actually bring these netting cards closer to where you'd want to be streaming that live feed?
Chris Milsted:And then looking at kind of technologies. You know everyone's been using satellite to distribute those feeds as they go around traditionally. Well, we have the Akamai backbone. Could we start look like, looking at maybe we could pipe this over that as well. So I think from the ecami side we're looking at more. You know, rather than taking kind of the data to a site, we're looking how do we kind of take the capabilities closer to the, the end users and where they're actually needed. I think that's kind of our main focus at the moment and I'll hand over to dennis on the kind of encoding side and observations there so yeah, so dennis, you know what.
Mark Donnigan:What does um having you know really highly performant vpus, you know hardware in the network. What does that mean for you in terms of how video systems are being built, or what do you see? You know, is there some major shift or trend? Or, you know, does it open up possibilities? Or is it just about cost savings, which is fair enough? You know, could be just like it's business as usual architecturally, and we're just going to be able to deliver video at a much lower cost. What do you see? And we're just going to be able to deliver video at a much lower cost. What?
Dennis Mungai:do. You see, I have never viewed cost savings as the goal. Cost savings is a side effect of well-designed systems and I believe the goal here is how to call not just togetherness but separating the distance between iteration to prod. Bringing systems this closer not only shortens the product lifecycle gap, but it means that with potential firmware updates, especially for the VPUs, you not only get newer codec support down the line, but you also get better rate control methods for the various encoding schemes you have. So essentially you get like a hardware fabric that behaves like a flexible software fabric.
Dennis Mungai:Small, incremental and meaningful upgrades only when they're needed, and if they are not needed you can have these rolled out in separable revisions such that, regardless of the distance or deployment type you have, we never have to constantly break production with dependencies. We do not not need an encoding layer, a simple example here being how the NetIntvKs work. All you have is an NVMe device and there is typically no quote-unquote host-side driver. All code texts are initialized through LibExcoder and then from LibExcoder. You then buff, or should I say create a single namespace from which you can then derive multiple encoder instances depending on resource availability and scaling. So this kind of compactness also allows us to reduce the kind of resources needed to push the same workload.
Dennis Mungai:I'll give you a simple example from production we have.
Dennis Mungai:We have hard containers as small as sub-100 MP3s, and then compare this to something that needs a whole CUDA stack to build. When you're pulling down the CUDA stacks from the NGC, that's going to be about 3 point something gigabytes in size versus 100 megabytes for the same appliance, for the same instance, for the same deployment type. So not only are you improving integration times, but you are also improving deployment scenarios for places that may be so remote you may not have the bandwidth you assume a lab facility has. So for us, the kind of work that NetEase is doing, where you can move a lot of this stuff to the edge if you need to and keep them on cloud as needed, is a level of flexibility appreciated, not as needed is a level of flexibility appreciated. I think this is the part that we will have to essentially learn how to live with and also pay for with time, because the real, like I said, the cost savings is not the goal. It is the outcome of systems so well designed that it becomes an inevitability.
Mark Donnigan:Yes, yeah, that's great an inevitability. Yes, yeah, that's great, incredible insights, and we totally agree with you. You know, it's interesting Just to circle back around on the environmental aspect of reducing your hardware footprint. I've been in many, many, many conversations where, you know, sort of behind closed doors, people were forced to kind of admit and say, look, at the end of the day, we're going to end up buying whatever's lowest cost. You know, first, you know, and my response is always well, that's great news. Because when you're able to, you know, get a four X, a five X, a six X or 10 X density advantage, it's lower cost, you're consuming less energy and, guess what, you're not polluting the environment with as much CO2 because you're consuming less energy. So it's fascinating how, um, you know, when you are focused on doing things, when we collectively, the industry, is focused on really driving efficiencies through every step of the end-to-end workflow, it has residual effects. Our operational costs generally are reduced, you know, we're using less energy, et cetera, et cetera, and those are all good things. It's wonderful, all right.
Mark Donnigan:Well, chris Dennis, thank you again for joining me on Voices of Video. I look forward to seeing you both at IBC in Amsterdam, just like five, five, five and a half weeks away now. So it's, it's coming up soon. It'll be a great time, and if any of our listeners are going to be in in Amsterdam at IBC, make sure you stop by the booth. Make sure you stop by the booth, say hi, dennis. Cirrus 21 actually has a stand in the VPU ecosystem. So does Akamai. So Akamai is going to be there, and I'm sure Chris will be, you know, hanging out and in and out and all around.
Mark Donnigan:So yeah so this is this is going to be going. This is going to be really good, great opportunity. So again, thank you for listening to Voices of Video.
Dennis Mungai:We appreciate you all and as I always say, happy encoding. 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.
