Energy Is the New Bottleneck in Live Video | Why VPUs Beat GPUs for Low-Res ABR at 4–6x Energy Savings

Live video is exploding, power budgets are shrinking, and the old “throw more GPU at it” mindset is breaking. We dig into the real constraint behind streaming at scale - energy - and share new data showing how VPUs can deliver 4–6x better efficiency than top-tier GPUs while holding quality where viewers notice it most. From the early days of CPU-only encoding to a modern, hybrid stack, we walk through the architecture that lets us stream more for less power without cutting corners on quality.
 
 We break down head-to-head tests run in the cloud comparing an RTX 4080 (NVENC and CUDA) with a NETINT Quadra T1U across AV1, HEVC, and H.264. You’ll hear how watts per stream changes the math for ABR ladders, why low-resolution rungs dominate real-world viewing, and where GPUs still win - premium HD and UHD tiers, complex filters, and specialized compute. Then we zoom out to the big picture: if live traffic is already the majority of internet bandwidth and is set to triple by 2030, scaling responsibly means optimizing for both density and sustainability. The numbers at 1,000 concurrent streams are stark: roughly 22.6 kW on a GPU path versus around 5 kW on a VPU path, with similar throughput and better low-res quality in many cases.

Our takeaway is a simple, pragmatic strategy. Use GPUs where their strengths shine and use VPUs for the heavy lifting across low and mid ABR rungs. That hybrid approach cuts operational cost, reduces carbon impact, and increases resilience under peak load. Along the way, we share how our live control, encoder, origin, and editor fit together across on‑prem and cloud, and why energy-aware orchestration is now a core feature, not an afterthought. Want the full benchmarks, VMAF curves, and methodology? Grab the white paper and put the data to work in your roadmap.

Enjoy the conversation, then help us spread the word - subscribe, rate, and share with a teammate who’s planning next year’s streaming capacity. Your feedback keeps these deep dives sharp and useful.

Download presentation: https://info.netint.com/hubfs/downloads/IBC-Peak-performance-minimal-footprint-Cires21.pdf

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.

Chapters

• 0:00 Setting the Energy Problem
• 0:19 From CPU to GPU to VPU
• 1:45 Building a Live Streaming Stack
• 2:08 Hybrid On‑Prem and Cloud Approach
• 2:08 Partnering with NETINT VPUs
• 3:08 Why GPUs Struggle on Efficiency
• 3:46 Test Setup and Method
• 4:23 Efficiency Results: 4–6x Gains
• 5:30 Scaling Impacts at 1K Streams
• 6:10 Quality vs Power Across Codecs
• 8:35 Key Takeaways and Strategy
• 9:56 White Paper and Sponsor CTA

Transcript

Voices of Video Intro: 0:07

Voices of video. Voices of video. The voices of video. Voices of video.

Nacho Mileo, Head of Product (Cires21): 0:17

The idea is talking about performance. Right now, um, we have been working with NETINT for a while, and we started with them because of density, but then we realized they were really good at energy too. And accounting that uh 60% of right now traffic in the internet is live streaming and it's growing. And we are supposedly tripling its size by 2030. We we started checking uh how uh how responsible it was when it comes to to energy. And right now we don't have just uh an like a hardware cost problem anymore. We have an energy budget problem. So a bit of story. Uh this is a flag chip. Uh this is C21 live encoder. Uh we built it on 2008 with CPUs. Uh then we added GPUs on top of that. Uh, this was even before the cloud was like like a thing. So uh we keep the CPU and the GPU and and we started growing from here. So at this point, we uh grow an ecosystem and we created an ecosystem that includes live control, live editor, and live origin. Uh again, as I was telling, the we even with with the GPU appearance, we stick to CPU and we we kept that architecture. So this is what it looks like right now. We have like a whole uh live streaming ecosystem that's formed by live control, which is an orchestration platform. We have live encoder I was mentioning, we have live origin, and we have the live live editor for for clipping and and editing. We have all of this uh both on-prem and and in cloud, so we we can approach this on a hybrid uh way. Um and the thing is, as I as I was telling two years ago, we met with NETINT and we added NETINT on on our encoders. So right now we have CPU, GPU, and VPU. And in in a context where the AVR are even more important is they that sports are requiring Ultra HD and 4K and 60 frames per second, and sustainability is is getting more and more pressure. We uh we saw that adding netting was like uh uh the way to go because they have an unmatch density. And and as I was telling you, uh we decided to test how it works when it comes to electricity. I will show a bunch of data right now. Um the thing with this data is that it's on a white paper, so I will do like a flyover, a quick flyover, but the real data is in the white paper that we will release after this talk. So uh first, let's do a reality check on GPUs because uh don't get me wrong, we love GPUs. Uh but the thing is we cannot pretend GPUs to work perfectly for gaming, for mining Bitcoin, and for video. It's not the way it works. It's it's it can do all of that, but it cannot do it like uh efficiently. So they deliver great performance but at a heavy energy cost. And the thing is, how sustainable is that brute force that that we use when when we use uh GPUs? So we did this test with a couple of folks from Ceros 21 and a couple of folks from um Akamai2, and of course with the help of of NetInt. And we did this uh on as I was telling Akamai Cloud, Frankfurt, in in Lino. We compare uh an RTX 4080 with QDA and M bank versus the Quadra TU1. We did different workflows with this, but one to one and one to N on ABR. And finally we measure throughput, power, utilization, and BMAF. This is what we see on efficiency gain. So, as you can see, for for netting, it's like in between four and almost six times uh more efficient when it comes to watts per stream. So we have 8264, HTVC, AV1. You will see that the the you know they are similar in terms of of savings, but in general, we never go over 0.7 on on net end, and we go almost uh to four in some cases, 3.5 in the case of of NVIDIA or of the GPU. So that's uh an efficiency gain of again average four to five times. So this is another way of saying it. It's way clearer that that the GPUs are super energy intensive when it when you compare it to to the VPUs. In and you can see that this is sustained in in every codec. So it's not a matter of of the encoding or the codec you are choosing, it's it's really, really uh different when it comes to energy consumption. You may say that these numbers are not huge, but then when you think about what happened when you put that numbers in 1K, this is what happens. It's like 22.6 kilowatt versus 5 kilowatts. So it's huge. The difference in in consumption is huge, and as as as I was saying, you cannot keep this up uh uh for for long. You know, it's on a world that's uh aiming to become greener and with an energy crisis and with AI consuming a ton of energy, we cannot uh you know skip this. Uh we should be responsible of this. We we didn't also measure this, but we also measured trans code capacity. And what we have seen, you know, forcing everything to to the very end, is that at 180p both GPU and CPU uh sorry GPU and VPUs uh work on 1920 streams on H264, but the GPU consumes four times more power from the same throughput. At 720p, Nvidia scale is slightly higher, 24 versus 22, but with five times more energy. And at lower runs, NVIDIA reaches up to 30 streams, which is 10 more than than net ins, but efficiency reverses, you know. Uh NVIDIA requires three or even three and a half times the energy that required that that was required by the VPU. And this is very nice, but what about quality? So this is what we've seen when we measure VMAF. I'm I'm I chosen I've chosen ever uh AV1 for this because uh they are very similar, and I wanted to go with AV1 because it's the newer, but you will see later that it's it's very similar. So at 180p, you will see that you know the winner is still NVIDIA, but for a little. 720 got some depicts, but in lower qualities, which are super important these days, especially if we consider mobile, we see that um you know we don't we don't only consume less energy with netting, but we also get better CMF scores. So as I was telling, this is AB1, and you see that we are always on the same path. When you see the white paper, you see that the it the path is almost identical for every every possible collect, but in the lower resolutions, netting performs better. Then we are almost even and five at 576. And we are almost similar again at 720, uh, at 180. But at 720, we got another DPEG that's super costful when it comes to energy. For AV1, NVIDIA is again best at HD but uh works worse in in low resolutions. In H E V C Nvidia wins in 180p, but netting wins in low resolutions again, and in H264, we are similar on the 180p. So the key takeaways for for this that for everything that that we measure the TW TU1, T T1U, sorry, it's between four and six times more efficient per stream than the than the similar GPU. The MVD ITX 4080, it's high in HD quality, yes, but it sometimes even six times, it requires six times the power that that the netting uh requires to work. So in low resolution, netting is superior both in quality but also in efficiency. So it's a win-win absolute in what when it comes to low resolution ABR. And finally, uh our recommendation is or the approach we are seeing is the hybrid deployments. Use GPUs for premium HD, but use VPUs for scale. So the bottom line is more streams for less power, lower cost that comes with green streaming, and scalable without having to compromise energy or space. That's all. Uh thank you very much. We will be on that corner, and we will release this white paper like in maybe 10 minutes or so. So feel free to download or look for us. Thank you.

Voices of Video Outro: 9:56

This episode of Voices of Video is brought to you NETINT Technologies. If you are looking for cutting edge video encoding solutions, check out NetInts products at netint.com.