Nvidia bid to reshape 5G needs Ericsson and Nokia buy-inNvidia bid to reshape 5G needs Ericsson and Nokia buy-in

Barring a few notable exceptions, today's mobile networks are "purpose built." In short, that means they are designed to do one job – provide connectivity for voice and data services – and do it as efficiently as possible. The technology platforms they use are not suited to other tasks, just as the powerlifting gym giant is a poor option for the pole vault.

For several years, Intel and other IT companies have been pushing an alternative "general purpose" story. It encourages telcos and their kit vendors to scrap the purpose-built technologies and replace them with the chips and platforms commonly found in data centers, on-premises servers and even ordinary PCs. The rationale is that the much bigger IT sector's economies of scale would trickle into and benefit telecom. The same servers used for the radio access network (RAN) could also support generic workloads. Succeeding the powerlifters and pole-vaulters is a more versatile athlete who wins overall.

But even before this form of virtual or cloud RAN has properly hatched, it risks being flattened by a third option that has taken shape in the last year. Nvidia is better known as trillion-dollar chipmaker providing the picks and shovels for the artificial intelligence (AI) gold rush. Yet the graphical processing units (GPUs) it has previously sold to the likes of Amazon, Google and Microsoft are now on offer as RAN chips, too.

Just as Intel does with its x86 architecture, Nvidia is positioning its technology as the foundation of a cloud RAN that would enjoy much bigger economies of scale than any of today's purpose-built networks can possibly realize. It has opportunistically pounced on what it sees as the inadequacy of older general-purpose technologies, and primarily x86, to support cloud RAN. For a long time, Intel was able to rely on Moore's Law, whereby the number of transistors that could fit onto a silicon chip of a given size would double every two years. But Jensen Huang, Nvidia's leather-jacketed boss, reckons Moore's Law is dead.

"Moore's Law has come to an end and if you want to continue developing this signal processing, which is very demanding with the high-performance workloads, you need to get onto an accelerated platform," said Soma Velayutham, the general manager of Nvidia's AI and telecom business. "If you look at the demands of a massive MIMO, 6G kind of architecture – being AI-driven – this needs a programmable, accelerated platform."

CUDA revolution

Nvidia clearly has a vested interest in depicting GPUs as a good option for the RAN. Amid signs of a slowdown in spending by hyperscalers, it would stand to earn billions of dollars if the world's telcos had a similar hunger for its chips. But Japan's SoftBank already plans a commercial deployment, and operators elsewhere are in talks to do the same.

When Mike Sievert, the CEO of T-Mobile US, complained a few months ago about the elusiveness "at scale" of "a more open RAN architecture," he did it on stage in conversation with Huang. T-Mobile's partnership with Nvidia, announced at the time, proves Sievert is taking the GPU maker's pitch very seriously indeed.

Equally important for Nvidia is a recognition by key figures within Ericsson and Nokia of the suitability of those GPUs for compute-intensive RAN workloads, including Layer 1, the most hardware-dependent code. "Graphics processing units, because of the vector processing, just fit Layer 1 compute very well," Tommi Uitto, the head of Nokia's mobile networks business group, told Light Reading early this year. In October, Fredrik Jejdling, his counterpart at Ericsson, said the Swedish vendor was exploring the potential of putting its RAN software on Nvidia's GPUs.

But hardware dependency means this would demand a rewrite of code previously generated for either purpose-built chips (application-specific integrated circuits, or ASICs) or an x86 platform. Ericsson, Nokia and others would have to acquire fluency in compute unified device architecture (CUDA), the platform in which Nvidia has invested. "They do have to reprofile into CUDA," said Velayutham. "That is an effort."

To help them get started, Nvidia has assembled its own stack of RAN software, branded Aerial. It is fully deployable as commercial-grade technology, insists the chipmaker, and already running on a distributed unit (DU) – essentially, a server hosting RAN software – in the live production network of Japan's NTT Docomo. But Nvidia has no ambition to be a DU company, according to Velayutham. Aerial, then, is mainly intended as a reference and starting point for the real experts.

"They can rewrite the code themselves, or they can adopt it as it is," said Velayutham. Fujitsu, a partner on the SoftBank project, appears to have integrated parts of Aerial with its own software for Layer 2 and turned that into a product. "Some believe they want to do it all themselves. We help them with resources and reference code. Some just want to integrate. We help them with just integration and give them the code," said the Nvidia executive.

He identifies Fujitsu, Mavenir and Radisys as "committed partners" while noting the earlier-stage tie-ups with Ericsson and Nokia. But the design of GPU-compatible RAN software would have clear ramifications for the big kit vendors and their existing partners. Today, Ericsson has two software tracks for Layer 1 – purpose built, for ASICs designed in-house, and virtual, for Intel's central processing units – while Nokia relies on a silicon partnership with Marvell Technology to deploy the same code for both purpose-built and virtual RAN. In either case, CUDA would mean investing in resources for a completely new track.

Adding multiple tracks like lanes to a highway is unlikely to suit either company in the long run. Ericsson sounds determined to be as hardware independent as possible. "Everything doesn't happen overnight, but if you build a virtualized stack then it should be able to run across multiple platforms," Jejdling told Light Reading in October. Nokia worries about maintaining feature parity when there are multiple tracks.

Power management

The objective for Nvidia, meanwhile, is to succeed where x86 has failed and show that CUDA – besides supporting the versatility of a cloud RAN – can deliver the performance of an ASIC. "What you're really getting is the performance of, or even better performance than, an ASIC that is actually in a software-defined architecture," said Velayutham. "CUDA is almost like a software-defined ASIC."

The hardest task for Nvidia is to silence the skeptics who insist GPUs are power hogs that will drive up RAN costs and decimate profits. Unfortunately, many telcos will have their eyes on the escalating energy consumption of GPU-buying hyperscalers, and Velayutham admits that Nvidia is essentially pitching the same GPUs deployed in those hyperscaler facilities to the telecom market.

There appear to be some nuances, however, that tend to be lost in the headlines. Nvidia's broad argument is that its GPUs in a telco facility can play a dual role of AI chip and RAN accelerator. But if a telco does not buy the AI pitch, it can invest in a less capable and less power-hungry GPU, according to Velayutham. "We have a range of GPUs, it's a sliding scale," he said. "You can have 20% RAN and 80% AI, or you can have 20% AI and 80% RAN."

The latest message from Nvidia is that such GPU choices should make no difference to RAN software vendors because of the optimization that CUDA provides. "CUDA abstracts from the underlying GPU," said Velayutham. "CUDA would actually optimize for the smaller GPU or the largest GPU." The range goes up to 1.5 kilowatts for the biggest, he said.

This also gives operators a range of deployment options, according to Nvidia. The pilot conducted with SoftBank demonstrated a single server supporting 20 cells in a configuration suitable for more densely populated areas. Velayutham believes this goes beyond what an x86 platform would be able to support.

The economics, though, evidently remain a concern for some telcos, including one in Japan. On a recent call with reporters, Sharad Sriwastawa, the CEO of Rakuten Mobile, said a distance limitation of 20 kilometers between the DU and the radio site made Nvidia's technology unviable. While he did not elaborate on the specific reason, it is probably the exorbitant cost of optical SFPs (small form-factor pluggables) once fronthaul connections exceed 20 kilometers, according to Earl Lum, the president of EJL Wireless Research.

"We cannot change the laws of physics of the 20 kilometers with better updates and better compute," said Velayutham, when challenged on that point. "What we are saying is we are providing you a compute that's highly efficient for this work, and that compute can come in different form factors." In less densely populated rural areas, Nvidia offers low-power GPUs that can be deployed at a radio site in a distributed RAN configuration, he insists.

Telcos and vendors may not like the idea of a RAN compute market dominated on the chips side by Nvidia. Indeed, while virtual RAN still accounts for only a tiny share of the total RAN market, Intel's dominance there remains a concern for parts of the industry. Even if successful GPU alternatives emerge, they seem unlikely to use CUDA as the underlying architecture, threatening more effort and complexity for software developers.

Unfortunately, the RAN market remains in decline and the capacity of the most advanced chipmaking foundries – led by Taiwan's TSMC – is increasingly being dedicated to Nvidia and Big Tech. Sales of RAN products fell 11% last year, to about $40 billion, according to Light Reading sister company Omdia. This year, it now forecasts another 10% to 15% drop. As a growing percentage of sales is reinvested in research and development, piggybacking on the AI and hyperscaler moves could seem like a sensible idea.