AI-RAN looked like it would be a big strategic topic at MWC – and so it proved over Day 0 and Day 1. Nokia made AI-RAN a key part of its pre-MWC press conference on Sunday, lining up three operator leaders with whom it is working on AI-RAN: T-Mobile USA’s President of Technology Ulf Ewaldsson, Indosat’s CEO Vikram Sinha, and SoftBank’s Ryuji Wakikawa, VP Head of Research Institute of Advanced Technology.
At MWC itself the expanded AI-RAN Alliance – it now has 77 members after launching a year ago – has its label on ten AI-RAN demos at partner booths. And vendors such as Nokia will also be showing their own demos, albeit not under the AI-RAN Alliance umbrella.
Why is the AI-RAN important? Well, first, it’s a possible route (not the only one) to enriching the operations of a network with the latest AI technology. Second, it could be a way to change the economics of deploying networks, by combining RAN deployments as software on the same platform as AI workloads.
As Ryuji Wakikawa said at the Nokia panel event, acknowleding previous failed efforts to leverage compute based in the mobile network via Mobile Edge Compute (MEC), “We’ve been speaking about edge computing for a decade. This is a last chance for operators to transform their premium edge into revenue streams”
AI-RAN introduces questions, as well as providing opportunity
This is all up in the air. Although Nokia has a clutch of announcements around AI-RAN, and Ericsson too is working with KDDI and T-Mobile, and had the AI-RAN Alliance’s Alex Choi on stage at its pre-MWC event, it’s not clear how AI-RAN might play out. It might indeed turn into a threat to the established vendors. Nvidia is developing its own radio software stack, from Layer 1 upwards. Other RAN software players, such as SynaXG are offering solutions for AI-RAN iterations, based on the Nvidia platform.
It’s also worth bearing in mind that the current established RAN vendors do not support the use of GPUS to run their RAN software. Ericsson did once look at Nvidia GPUs for Cloud RAN acceleration, back in 2019, and Nokia’s prototype that it is demonstrating at MWC with SoftBank does not place its vRAN on the GPU. Nokia said in a press release a year ago that it would be working with Nvidia on using GPUs for vRAN acceleration, but then seemed to row back from that statement in further briefings.
Nokia’s AI-RAN work on the GPU
However, TMN has learnt that Nokia is now indeed investigating the use of GPUs for vRAN accleration, in its work with T-Mobile.
Aji Ed, Head of Partner Cloud RAN Solutions, Nokia, said the company is conducting a feasibility sstudy with T-Mobile, to look at how it might run the physical Layer 1 on the GPU. “We are making progress, that’s what I can say at the moment. But we haven’t decided any productisation plan yet,” he said. (See below for more on this)
Nor is the architecture for AI-RAN a done deal. T-Mobile’s Ewaldsson said that he wasn’t sure that the architecture for AI-RAN is yet fully understood. “I’m not sure if the architecture is ready or where they should be, at the base station or further upstream, ” he said of the compute resources that would underpin the AI-RAN.
Thirdly, the economics are extremely unproven. Will telcos be able to monetise the GPU capacity they will have in the edges of their networks, by offering it as a service to other users? This model makes many nervous.
SoftBank’s AI-RAN prototype
SoftBank had already briefed media and analysts, under embargo, on some of the details on its work on AI RAN, and with its partners Nvidia, Fujitsu and Nokia.
Ryuji Wakikawa, VP Head of Research Institute of Advanced Technology, said that AITRAS, KDDI’s name for the AI-RAN system it has built, has been making progress on several fronts, incorporating orchestration and management of AI and vRAN elements on its GPU-based platform.
First off, what is AITRAS? Well, it’s built on the AI-RAN concept, so it incorporates vRAN and AI on the same GPU server, which is the Nvidia Grace Hopper 200. RedHat OpenShift handles the platform apps and cloud environment, whether they are AI apps or the vRAN running concurrently. SoftBank has developed its own L1 RAN software (*TMN understands this is in fact a co-development with Fujitsu on the Nvidia Ai-Aerial platform), and uses Fujitsu for L2/L3. SoftBank has also developed its own orchestrator, which allows it to control and manage the software and hardware elements sitting in the system.
Nokia bringing its own RAN, not on the GPU
One of SoftBank’s announcements was about work it has done to integrate Nokia’s RAN with the Grace Hopper and AITRAS platform, so that the operator can manage the Nokia vRAN workloads in concert with the AI workloads.
Wakikawa offered insight into how Nokia is integrating its vRAN with the Nvidia GPU platform. The Nokia RAN software is not running on the Nvidia GPU. For this integration SoftBank is removing its L1 and Fujitsu’s L2/L3, and bringing in Nokia AnyRAN accelerator as a SmartNIC in a slot in the Nvidia-based server. This is the part that offloads the vDU acceleration required for L1. The L2/L3 processing can be achieved in the CPU part of the platform.
“Unfortunately Nokia does not yet support GPU acceleration,” Wakikawa said, so we use their accelerator card.”
Here’s where the AITRAS orchestration comes in. “The GPU is fully available, so using our orchestrator we activate the GPU for AI allocation. It’s a big step to collaborate with a big vendor to prototype if the RAN and Grace Hopper can exist on the same platform.”
The trial has also integrated Nokia’s MantaRay network management with the AITRAS orchestrator. The Manta RAY system counters enable AITRAS to predict the demand of the radio, which it then uses to allocated CPU resources between AI and the RAN.
One thing to note, then, is that in this prototype SoftBank is paying for the Grace Hopper platform but is not getting the enhanced economics of using the GPU platform for L1 acceleration. Instead it is plugging in the Nokia accelerator as a SmartNIC and using the GPU solely for AI, whilst managing to combine CPU management according to RAN and AI requirements, using the MantaRay and AITRAS systems in combination.
Whilst Nokia’s AnyRAN approach has indeed allowed it to plug in to the server, does it feel like a halfway house that the GPU the operator has invested in does not support the RAN element as well? Does Nokia plan to go the full way and support the use of a GPU-accelerated RAN?
Nokia, T-Mobile and Nvidia
Nokia’s Aji Ed has confirmed, as reported above, that Nokia is working with T-Mobile to explore the use of GPUs for vRAN acceleration. He said the company is benchmarking the performance of different channels on the CPU, and how it can extrapolate this into feature parity with its integrated and Cloud RAN prodcuts.
“We still need to make a product decision but we are making the proper evaluations at this point to make sure we have the right understanding of the TCO and cost benefit analysis. This is what we will be doing in 2025.”
But even as Nokia makes its exploration,Ed is not convinced that the AI-RAN business case necessarily needs the RAN to run on the GPU.
“You don’t have to integrate fully with the GPU, but of course if you integrated L1 on the GPU you can get a bit more flexibility. You can probably rearrange a bit of resources. But that’s not the main thing, because the main thing is driven by the AI workloads.
“If we are talking about the real co-operative business cases coming from the AI workloads, the AnyRAN approach will enable, as an intermediate step, operators to really figure out what we can do with the AI workloads on their RAN systems.”
The other possibility is that AI-RAN serves as a potential break in the RAN supply chain, with other players leveraging GPU capabilities with RAN software. One of these could be Nvidia itself, with its AI-Aerial platform, which it says is a complete RAN L1 and above platform.
While Ed did not want to comment on Nvidia’s plans directly he did say, ” They have done some trials. Our view is that for any product to make it commercially, you need feature partiy compared to what we have today. For instance on beam forming, energy efficiency, carrier aggregation, to make it compatible with existing purpose built or cloud-native systems. So I think bringing all these features into Aerial will be the challenge – I don’t know exactly how much feature parity they have, but that would be the challenge I see – in order to bring such a system into real commercial implementation.”
Other clever things SoftBank is doing with AITRAS
For those who want to know more about what SoftBank is doing with AITRAS, read on.
First, it is demonstrating the use of AI to enhance the radio. The first is Uplink Signal Processing, where it has implemented an on-air (ie. not simulated) implementation of channel loss interpolation, seeing a 20% performance gain.
It has also deployed MU MIMO with AI enhancing the MAC scheduling on a system level simulator using multi-layer perception. Tests have shown a 9% gain – a “huge number” Wakikawa said.
And the third is carrying out beam forming predictions to keep the accuracy of beam steering.
It is also experimenting with deploying the CU element on a Grace-Grace server (so a double Arm CPU platform, no GPU), which it thinks is a more suitable platform for the CU, doubling capacity for the CU, and being able to support more DUs from the CU. (the DU says on the Grace-Hopper platform because it does need the GPU acceleration).
Power management and energy efficiency control using Red Hat’s open source Kepler. It is also using Kepler to figure out how much the energy consumption is per port on Red Hat OpenShift. This will manage and optimise across regional differences in power demand, as well as the sources of electricity powering different pods. SoftBank can then manage which app goes to which pod based on power availability and pricing.
Edge Routing
By integrating 3GPP features SoftBank is able to route specific user traffic to a specific AI server on the edge, smart switching between the internet and local breakout paths.
