Summary Bullets:
• NVIDIA introduced a new virtual RAN platform based on its GPUs
• The company argues that a converged RAN/AI network can make the RAN more efficient and enable new services to further monetize mobile networks.
Earlier this year, when AI giant NVIDIA and others announced the formation of a new industry consortium, the AI RAN Alliance, the lack of detail surrounding their plans was understandable. After all, they had promised not much more than to begin exploring the ways that AI could be used to transform RANs.
Last week, NVIDIA took another significant step. In addition to announcing an AI RAN innovation center with US operator T-Mobile, NVIDIA articulated an updated vision for the AI RAN as well as a new offering meant to support that vision. The big picture is: monetizing mobile networks by converging them with AI infrastructure. AI could potentially make RANs run more efficiently and improve performance, including reducing jitter and latency enough to enable new services like industrial automated vehicles and so on. In addition, when network traffic is low, NVIDIA imagines operators using their new AI capabilities – supported by NVIDIA GPUs – to sell new services using generative AI – perhaps even GPU-as-a-service.
NVIDIA’s offering for helping operators achieve this vision, branded AI Aerial, breaks down into three computing elements:
• AI Radio Frameworks, which is used for developing and training AI models (deployed on the same networks that AI will be tasked with improving).
• Omniverse Digital Twin, which allows operators to test their AI models in simulations of real-world networks and
• CUDA-Accelerated RAN, the platform to deliver AI RAN on the network.
That last element is new and significant. NVIDIA calls it ARC-1 – a RAN computing platform meant to handle the full virtual RAN software stack (the physical layer and above), including software-defined baseband processing, fronthaul and backhaul. ARC-1 is built on NVIDIA’s MGX servers and powered by its Grace Blackwell chips (a combination of GPUs and CPUs). It also uses the company’s Compute Unified Device Architecture (CUDA) parallel computing platform – a kind of operating system for accessing the GPU. Operators can deploy it in centralized locations or at the cell site, for example.
NVIDIA said existing RAN vendors like Ericsson and Nokia – who work with NVIDIA in the AI RAN Alliance – could use ARC-1 hardware and CUDA libraries and add their own RAN software on top. It’s also compliant with Open RAN (specification 7.2), so that the platform could work with radio units from other vendors (NVIDIA noted Fujitsu, Ericsson’s Open RAN radio supplier, as an ecosystem partner in this area).
The vision is bold and potentially disruptive, which makes it all the more notable that NVIDIA is working with major RAN vendors on AI RAN. And it addresses a top concern of mobile operators: increasing network monetization. But achieving it is a very tall order. At this early stage of AI RAN exploration, it remains unclear to what extent efficiency gains made in the RAN may be offset by AI’s own considerable power requirements, for example. And as it suggests a fundamental overhaul of network architecture, operators disappointed with the returns on their 5G network investments may find the expense and complexity of achieving this vision extremely daunting. They’re likely to see AI RAN as a 6G technology to implement at scale no earlier than the 2030s.
At the same time, NVIDIA can argue that (a) increasing network traffic – including a rising wave of AI application traffic that NVIDIA itself is helping to generate – will require operators to keep investing in their networks no matter how reluctant they may be to open their wallets and (b) AI can deliver a greater ROI than the first 5G networks did by increasing efficiency and agility (in theory) and enabling new services, thereby countering the cause of operators’ conservative stance. In any case, a company with NVIDIA’s resources and influence can move mountains. Moving the mobile networking industry, though, might take a little longer.