# GPU‑Accelerated Performance Gains With AI-RAN powered by Nokia and NVIDIA (Presented by Nokia)

## Abstract

Get a look at how Nokia’s co‑innovation with NVIDIA marks a deep integration of GPUs in the radio access network (RAN), bringing high‑performance computing to the edge and accelerating the path to 6G leadership. This session will discuss the latest performance benchmarks and new use cases for AI‑RAN.

## AI Summary

- Presented an overview of the evolution of network workloads highlighting how AI is becoming the dominant workload with a focus on the challenges and opportunities it brings
- Explained the importance of reliable and deterministic connectivity for AI applications especially in critical areas like autonomous vehicles and robotics
- Described the collaboration with NVIDIA to integrate AI capabilities into existing 5G networks and prepare for 6G which will be AI-native by design emphasizing the use of GPUs and CUDA libraries to enhance network performance

## Transcript

**[00:00:12 – 00:00:13]** Okay.

**[00:00:13 – 00:00:15]** Hello and welcome.

**[00:00:15 – 00:00:18]** I hope you can hear me well.

**[00:00:18 – 00:00:20]** Thank you for the feedback.

**[00:00:20 – 00:00:22]** My name is Arne Schelicke.

**[00:00:22 – 00:00:26]** In the agenda you might have seen the name of Arji Ed,

**[00:00:26 – 00:00:28]** our head of AI-RAN with Nokia.

**[00:00:29 – 00:00:33]** I'm stepping in on his behalf here as he had an issue

**[00:00:33 – 00:00:35]** with a flight today.

**[00:00:36 – 00:00:40]** I'm from the strategic marketing of Nokia Global and my focus

**[00:00:40 – 00:00:42]** is in the AI-RAN area.

**[00:00:42 – 00:00:45]** Before I go into the presentation, I'd like to get a little

**[00:00:45 – 00:00:48]** bit more information about you in the audience.

**[00:00:48 – 00:00:51]** So, anybody coming from the robotics side,

**[00:00:52 – 00:00:54]** please raise your hand.

**[00:00:56 – 00:00:57]** Anybody coming from the telco side?

**[00:00:57 – 00:01:01]** OK, some here.

**[00:01:01 – 00:01:02]** And I speak your language.

**[00:01:02 – 00:01:05]** I'm trying to learn the AI language.

**[00:01:05 – 00:01:09]** Anybody coming from the application side, AI application side?

**[00:01:09 – 00:01:12]** Now I expect a lot of other people to raise their hand.

**[00:01:12 – 00:01:15]** Otherwise, some of the audience are sleeping.

**[00:01:15 – 00:01:17]** Oh, OK.

**[00:01:17 – 00:01:20]** Good, let me start.

**[00:01:21 – 00:01:24]** From the network side, and I know these people here know

**[00:01:24 – 00:01:29]** it, the other ones maybe not, we always have a dominant workload.

**[00:01:29 – 00:01:33]** So 30 years ago it was all about making voice services

**[00:01:33 – 00:01:35]** mobile and ubiquitous.

**[00:01:35 – 00:01:37]** Then there came 3G and data.

**[00:01:37 – 00:01:42]** When there came 4G and the dominant workload there was obviously video.

**[00:01:42 – 00:01:45]** Video either as a streaming, as a download,

**[00:01:45 – 00:01:47]** or embedded in social media.

**[00:01:47 – 00:01:51]** And now, we are really at the cusp of

**[00:01:51 – 00:01:56]** A new dominant workload for networks, and that's AI.

**[00:01:58 – 00:02:02]** And in AI, I learned that everything comes in trillions.

**[00:02:02 – 00:02:09]** So, 1.3 trillion independent AI interactions per year,

**[00:02:09 – 00:02:11]** sessions you might call it.

**[00:02:11 – 00:02:17]** That translates into 100 trillion AI tokens per day.

**[00:02:17 – 00:02:23]** It translates into 77 exabyte of traffic per month generated

**[00:02:23 – 00:02:27]** by AI already today over networks.

**[00:02:27 – 00:02:31]** And more than half of it has today a mobile endpoint, so

**[00:02:31 – 00:02:32]** that's mobile traffic.

**[00:02:32 – 00:02:36]** And that's why we talk about AI-RAN, the radio access network

**[00:02:36 – 00:02:40]** fusing with AI technology.

**[00:02:40 – 00:02:46]** But this is not only about AI traffic now becoming dominant.

**[00:02:46 – 00:02:49]** It's about different quality of traffic.

**[00:02:49 – 00:02:54]** On the left, you see a traditional, conventional AI traffic, mobile

**[00:02:54 – 00:02:58]** traffic, which is quite predictable and evolving smoothly.

**[00:02:58 – 00:03:00]** So if you have a video stream of a couple of seconds, for

**[00:03:00 – 00:03:05]** that couple of seconds, everything is very stable, predictable.

**[00:03:05 – 00:03:09]** If you have AI traffic, there are a couple of things that change.

**[00:03:09 – 00:03:14]** You see in the animation, it's more bursty, it's less predictable.

**[00:03:15 – 00:03:18]** And what you don't see in this animation is, it is more

**[00:03:18 – 00:03:25]** uplink heavy, where a lot of data pulled from the cloud,

**[00:03:25 – 00:03:30]** from the devices, camera, conditioning input, whatever it is.

**[00:03:30 – 00:03:35]** But it's also not only the user traffic, the payload,

**[00:03:35 – 00:03:37]** it's also that the sessions, if they become smaller, it's

**[00:03:37 – 00:03:42]** the control side of how you manage the traffic in your RAN network.

**[00:03:42 – 00:03:48]** And we need to be ready, and we are ready as Nokia, to have the

**[00:03:48 – 00:03:59]** platform to reliably connect the AI agents, robots, and cloud services.

**[00:03:59 – 00:04:02]** Another aspect is

**[00:04:03 – 00:04:08]** We are talking about high value services in the physical AI.

**[00:04:08 – 00:04:15]** It's, if you don't transmit reliably, physical AI devices,

**[00:04:15 – 00:04:20]** autonomous vehicles of all sorts, robots, humanoids,

**[00:04:20 – 00:04:22]** they could cause damage.

**[00:04:22 – 00:04:24]** They could cause harm.

**[00:04:24 – 00:04:28]** So the connectivity behind them needs to be extremely reliable.

**[00:04:28 – 00:04:30]** And we are talking here about a journey to

**[00:04:30 – 00:04:34]** Deterministic connectivity coming from a five-nines

**[00:04:34 – 00:04:37]** environment, moving towards a six-nines environment

**[00:04:37 – 00:04:41]** in terms of availability of a service, enabling then high-value

**[00:04:41 – 00:04:47]** services for the AI world through our reliable connectivity.

**[00:04:48 – 00:04:50]** How do we do this journey?

**[00:04:51 – 00:04:54]** It starts nowadays with 5G already, and we are putting

**[00:04:54 – 00:04:57]** AI on top of the existing system.

**[00:04:57 – 00:05:00]** We are doing that in the base stations, but we are also

**[00:05:00 – 00:05:04]** doing it in more centralized systems, and that's reality today.

**[00:05:04 – 00:05:07]** When we talk about the evolution in a couple of years, we talk

**[00:05:07 – 00:05:12]** about 6G, and 6G, the next cellular standard, will be

**[00:05:12 – 00:05:15]** AI-native by design.

**[00:05:15 – 00:05:19]** And that means that we have the compute, the connectivity,

**[00:05:19 – 00:05:22]** the control, all aligned around the AI paradigm.

**[00:05:22 – 00:05:26]** It's a totally different way to build networks.

**[00:05:26 – 00:05:30]** And for our network customers as Nokia, we make this

**[00:05:30 – 00:05:31]** evolution very smooth.

**[00:05:31 – 00:05:35]** So on the one hand, it's extremely revolutionizing the thing and

**[00:05:35 – 00:05:38]** the value of the mobile networks.

**[00:05:38 – 00:05:41]** On the other hand, we try to make the journey for network operators,

**[00:05:41 – 00:05:46]** telecom service providers, As evolutionary as possible.

**[00:05:46 – 00:05:48]** And how are we doing that?

**[00:05:48 – 00:05:55]** Last year in GTCDC, that was October 2025, we set the stage

**[00:05:55 – 00:06:00]** for smarter, faster and more energy efficient networks by announcing

**[00:06:00 – 00:06:02]** the collaboration with NVIDIA.

**[00:06:02 – 00:06:06]** It was really a fundamentally important moment for

**[00:06:06 – 00:06:07]** our whole industry.

**[00:06:07 – 00:06:10]** We are very pleased to work with NVIDIA on that, bringing

**[00:06:10 – 00:06:14]** together the best of both sides, from the connectivity,

**[00:06:14 – 00:06:20]** all the decades-long experience in mobile networks of Nokia, together

**[00:06:20 – 00:06:25]** with the dominant design in AI platforms with NVIDIA, not only

**[00:06:25 – 00:06:30]** the GPUs, but also CUDA libraries and everything which comes on top.

**[00:06:30 – 00:06:35]** On the right side, what you see is how we make this evolution

**[00:06:35 – 00:06:39]** to AI RAN very, very smooth for our customers.

**[00:06:39 – 00:06:42]** Let me explain that a little bit more in detail for those

**[00:06:42 – 00:06:48]** who are not familiar with base stations and radio access networks.

**[00:06:48 – 00:06:54]** So today, the compute of base stations is purpose-built,

**[00:06:54 – 00:06:55]** dominantly purpose-built.

**[00:06:55 – 00:06:59]** There are some cloud networks, cloud RAN networks, but the

**[00:06:59 – 00:07:03]** majority, more than 99% globally, is purpose-built.

**[00:07:03 – 00:07:08]** And what we do, we bring in the NVIDIA Arc Pro.

**[00:07:08 – 00:07:13]** into our existing platform so that you have basically

**[00:07:13 – 00:07:16]** just to introduce a plug-in unit at the time of convenience

**[00:07:16 – 00:07:19]** as a mobile operator and you bring in all the benefits

**[00:07:19 – 00:07:24]** of multi-purpose capabilities of a compute and of a further network

**[00:07:24 – 00:07:29]** evolution via software by just introducing once with this plug-in

**[00:07:29 – 00:07:33]** unit and by that we decouple the

**[00:07:33 – 00:07:38]** Capability evolution speed of the RAN, of the radio access network

**[00:07:38 – 00:07:41]** from the hardware introduction because once we are on the

**[00:07:41 – 00:07:46]** GPU platform, on the CUDA platform, the further thing is all software.

**[00:07:47 – 00:07:51]** Of course, we have a million plus installed base here so

**[00:07:51 – 00:07:56]** this is not negligible and that translates also into a

**[00:07:56 – 00:08:01]** A significant portion of the AI grid, which was introduced by Ronny

**[00:08:01 – 00:08:06]** Bacista this morning in his speech, AI grid being, you have a lot of

**[00:08:06 – 00:08:11]** distributed AI compute capabilities around the world, and in larger

**[00:08:11 – 00:08:16]** networks you have, in an individual network, more than 10,000 more.

**[00:08:16 – 00:08:20]** Maybe 50,000 day station sites, each of them being a point

**[00:08:20 – 00:08:24]** of access in proximity of the robotic systems, of the

**[00:08:24 – 00:08:25]** users, of everything.

**[00:08:25 – 00:08:30]** So that is one of the things which make the introduction

**[00:08:30 – 00:08:34]** of AI RAN into commercial reality extremely fast because

**[00:08:34 – 00:08:36]** we have an evolution path.

**[00:08:36 – 00:08:42]** Then for those networks where Cloud RAN is a reality, we

**[00:08:42 – 00:08:49]** also have an opportunity to use commercial-off-the-shelf servers.

**[00:08:49 – 00:08:54]** And at our booth here in the exhibition hall, you can see

**[00:08:54 – 00:08:59]** one of them from Quanta, which uses the Arc Pro, and you can

**[00:08:59 – 00:09:01]** have a look how this looks like.

**[00:09:01 – 00:09:09]** So very nice edge server design with Arc Pro inside to run AI.

**[00:09:09 – 00:09:11]** Workloads and RAN workloads.

**[00:09:11 – 00:09:14]** And we have a live demonstration of that multi-purpose

**[00:09:14 – 00:09:16]** capability of a platform.

**[00:09:16 – 00:09:21]** With that, we turn the cellular grid into an AI grid.

**[00:09:21 – 00:09:26]** But AI RAN is, for us as Nokia, more than just the base station.

**[00:09:26 – 00:09:31]** Yes, on the base station side, there are a lot of opportunities to

**[00:09:31 – 00:09:32]** increase the spectral efficiency.

**[00:09:32 – 00:09:36]** And spectrum is so extremely expensive that you need to

**[00:09:36 – 00:09:39]** be very efficient in using it.

**[00:09:40 – 00:09:43]** We talk about the AI native air interface with 6G.

**[00:09:43 – 00:09:45]** A couple of years ago, I think now three years ago,

**[00:09:45 – 00:09:49]** we have demonstrated the first prototype of that.

**[00:09:49 – 00:09:53]** Machine learning based algorithms everywhere, it's a no-brainer.

**[00:09:53 – 00:09:56]** Advanced channel estimation, for example, we have shown

**[00:09:56 – 00:10:00]** that on NVIDIA GPU this year at Mobile World Congress.

**[00:10:00 – 00:10:03]** Advanced beamforming, another combinatorial problem which

**[00:10:03 – 00:10:07]** you can better solve with AI than with any other heuristics.

**[00:10:07 – 00:10:10]** But this is spectral efficiency and many other efficiency gains,

**[00:10:10 – 00:10:13]** performance gains, in the base station because this is real-time.

**[00:10:13 – 00:10:18]** When it does not need to be real-time, you might benefit

**[00:10:18 – 00:10:21]** if you do the AI in a more centralized location, in the

**[00:10:21 – 00:10:25]** cloud or on-premises somewhere, but not at every base station site.

**[00:10:25 – 00:10:31]** And that is where we have the Manta Ray suite of solutions.

**[00:10:31 – 00:10:36]** And the SMO is the latest one here, which allow to automate

**[00:10:36 – 00:10:39]** the operations of the network, make the network way faster.

**[00:10:39 – 00:10:40]** What do I mean with that?

**[00:10:40 – 00:10:48]** With Manta Ray Autopilot, we do 15,000 autonomous, closed-loop,

**[00:10:48 – 00:10:52]** autonomous operations in an hour, which is a breakthrough

**[00:10:52 – 00:10:56]** compared to traditional ways how you optimize your network.

**[00:10:56 – 00:10:58]** The fact we could increase the...

**[00:10:58 – 00:11:02]** The resource utilization in the air interface by 30%,

**[00:11:02 – 00:11:12]** that's tremendous, giving results in more than double-digit user...

**[00:11:12 – 00:11:16]** Speed experience gains while the overall traffic is growing.

**[00:11:16 – 00:11:18]** And that's all by software.

**[00:11:18 – 00:11:23]** So automation helps for increasing the utilization of the resources

**[00:11:23 – 00:11:24]** in the base station.

**[00:11:24 – 00:11:28]** We increase the amount of resources available by increasing

**[00:11:28 – 00:11:29]** the spectral efficiency.

**[00:11:29 – 00:11:30]** They both go hand in hand.

**[00:11:30 – 00:11:34]** And this year we announced that we open up the marketplace for

**[00:11:34 – 00:11:40]** our apps, AI apps, running on the Manta Ray Suite in an announcement

**[00:11:40 – 00:11:43]** around the SMO just recently.

**[00:11:44 – 00:11:48]** So, at our booth, you can see the live system, how it

**[00:11:48 – 00:11:51]** works, the AI RAN today.

**[00:11:51 – 00:11:54]** Here you see with whom we are already working together,

**[00:11:54 – 00:11:58]** the public engagements which we have communicated so far.

**[00:11:58 – 00:12:01]** You see in the USA, it's T-Mobile.

**[00:12:01 – 00:12:04]** We are very proud about that collaboration because T-Mobile

**[00:12:04 – 00:12:08]** is one, if not the leading operator in the world.

**[00:12:08 – 00:12:12]** But we expand also to other companies like SoftBank

**[00:12:12 – 00:12:14]** and Docomo in Japan.

**[00:12:14 – 00:12:19]** to a couple of the leading operators in Europe with Deutsche

**[00:12:19 – 00:12:24]** Telekom, British Telekom, Vodafone, Telia in the Nordics

**[00:12:24 – 00:12:28]** and Elisa in the Nordics and just recently we expanded also to

**[00:12:28 – 00:12:30]** TIM in Brazil and you have very...

**[00:12:30 – 00:12:35]** From the very beginning, Indosat, Aurelio Hutchinson, IOH was

**[00:12:35 – 00:12:37]** also among our engagements.

**[00:12:37 – 00:12:41]** They have very strong and bold plans for turning their

**[00:12:41 – 00:12:46]** network into an AI grid, which can close the digital divide,

**[00:12:46 – 00:12:50]** or shall I call it now the AI divide, in their country.

**[00:12:50 – 00:12:54]** Provide AI capabilities to everybody.

**[00:12:55 – 00:12:57]** Now, what is the way forward?

**[00:12:57 – 00:13:00]** I started this presentation by saying, yes, last year in October,

**[00:13:00 – 00:13:01]** we announced the collaboration.

**[00:13:01 – 00:13:07]** Then in the fourth quarter last year, we have demonstrated

**[00:13:07 – 00:13:11]** that the software of the 5G

**[00:13:11 – 00:13:15]** Here it's called layer 1 software, that's a technical detail

**[00:13:15 – 00:13:19]** I would call, but that's the most compute intense thing of the 5G

**[00:13:19 – 00:13:23]** software, the RAN software, running on NVIDIA GPU, and we had that, and

**[00:13:23 – 00:13:28]** additionally we have simultaneously another workload, and we are

**[00:13:28 – 00:13:33]** using the GPU virtualization, the multi-instance GPU, which

**[00:13:33 – 00:13:38]** our platform provides, so you can split the GPU into a RAN part

**[00:13:38 – 00:13:41]** and a part for other workloads.

**[00:13:41 – 00:13:43]** Yeah, and we had a live demonstration

**[00:13:43 – 00:13:44]** at Mobile World Congress.

**[00:13:44 – 00:13:46]** We have a live demonstration here.

**[00:13:46 – 00:13:49]** Sorry that this is not on the slide.

**[00:13:49 – 00:13:51]** We go into field trials.

**[00:13:51 – 00:13:53]** So proof of concept we have.

**[00:13:53 – 00:13:59]** But field trials on a larger scale with multiple challenging

**[00:13:59 – 00:14:04]** aspects, like massive MIMO, multi-user MIMO, and higher

**[00:14:04 – 00:14:09]** data RATES FOR HIGHER ORDER.

**[00:14:09 – 00:14:15]** QAM, we introduce later this year, then next year is the year where

**[00:14:15 – 00:14:20]** we go with the first customers of us commercial, that is the time

**[00:14:20 – 00:14:25]** when this AI grid, AI RAN is an AI grid, will become available for AI

**[00:14:25 – 00:14:29]** workloads of the larger industry, of the larger AI ecosystem

**[00:14:29 – 00:14:36]** in those countries, in those areas where we launch it commercially.

**[00:14:36 – 00:14:39]** And then, as I said earlier,

**[00:14:39 – 00:14:44]** With the platform already introduced here, when the evolution

**[00:14:44 – 00:14:49]** to 6G, and for those who are not in the radio access or in the cellular

**[00:14:49 – 00:14:53]** industry, 6G is expected to come.

**[00:14:53 – 00:15:01]** Sometime between 2029 and 2030, consensus being 2030,

**[00:15:01 – 00:15:06]** we can then go to 6G largely by software evolution which

**[00:15:06 – 00:15:12]** accelerates the time to market for our customers to activate 6G.

**[00:15:13 – 00:15:17]** And with that, I'm closed.

**[00:15:17 – 00:15:19]** I don't know if there's an opportunity for some

**[00:15:19 – 00:15:21]** questions or answers.

**[00:15:21 – 00:15:27]** If not, I say thank you very much, and I give one or two

**[00:15:27 – 00:15:28]** minutes back to you.