Nokia says it is undecided about using Nvidia’s GPUs in future massive MIMO radio units but claims they have major benefits in AI-RAN.
There isn’t much spectral efficiency in the movies. Conventions of the ghost-story genre dictate a gradual ratcheting of tension, like the slow twist of the executioner’s garrotte, so that the phantoms have toiled invisibly for hours to disturb and disorient their first victim before she finally emits a TV-rattling shriek. Spectral efficiency in telecom, while of a less creepy nature, has been just as woeful. Such, at least, is the implication of Nokia’s latest claim that network capacity, rather than a specter’s scariness, can be doubled by 2028.
It would require only the insertion of a new card that hosts one of Nvidia’s graphics processing units (GPUs) into Nokia’s existing Airscale chassis. Designed to be less powerful and energy-hungry than the GPUs sold to data centers, it could be slotted in, like a food tray in a kitchen trolley, alongside other cards that host radio access network (RAN) software without frying the equipment.
Older cards run algorithms written by people. It is by switching parts of those algorithms for AI on its GPU cards that Nokia believes it can realize those spectral efficiency gains, especially in Layer 1 (or L1), the most computationally demanding slice of RAN software. “What we’re demonstrating with AI-RAN now is visibility that we can get to more than 100% spectral efficiency gain by 2028, doubling the capacity operators can get from their existing spectrum assets,” said Justin Hotard, Nokia’s CEO, on a call with reporters last week.
Silicon supremo
A promised doubling is a huge raise on rival Ericsson’s “up to 10% better spectral efficiency” AI-RAN pitch last month. Unlike Nokia, Ericsson is sticking with custom silicon conceived in-house. In its antenna-rich, massive MIMO radios it has also begun to introduce new internally designed chips it calls neural network accelerators for AI-based improvements. GPUs have not yet entered Nokia’s radios, as opposed to its baseband gear. But they could as Nokia works on a successor to Habrok, the brand name for its massive MIMO range.
“We’re currently working on the definition of the next-generation Habrok for massive MIMO, which will be in the 2028 timeframe horizon,” said Mark Atkinson, the head of Nokia’s RAN business. Various options are on the table. “No reason why it should, and no reason why it shouldn’t,” he told Light Reading when asked if a GPU would be used.
“Essentially, you need an L1 beamforming chip in massive MIMO, and there are many ways to do that,” Atkinson explained. Beamforming is an advanced technique that allows the mobile network to shoot signals directly at user devices rather than blanketing a wider area.
While Nvidia has neither confirmed nor denied this publication’s earlier report on plans to develop a GPU product for massive MIMO radios, the company has argued that GPUs would have benefits in this part of the network. A possibility in 6G is that even more L1 processing will move from baseband appliances and servers into radio units, a trend already observable in 5G. Functions might include “massive MIMO precoding, uplink receiver processing, and latency-sensitive functions,” said an Nvidia spokesperson in emailed comments. Such workloads rely on “dense matrix operations” where GPUs excel, according to the chipmaker.
Bettering Stradivarius
Talk of spectral efficiency may sound either arcane or bizarrely supernatural to the layperson, but it could be worth billions to the average big telco. Operators invest enormous sums in spectrum, a finite resource, as well as the networks that use it. A doubling in capacity is essentially a doubling in the value of those frequencies.
Yet in the absence of commercial proof, some analysts are dubious. “The 2x capacity claim using AI-RAN with no roadmap on how Nokia will get there remains a TBD item for EJL Wireless Research,” said Earl Lum, the founder of the analyst company he identifies, in a LinkedIn note. Some wonder what the baseline is for comparison. A doubling on a badly performing network might not be that impressive.
Industry experts have also voiced doubt that AI can make such a big difference. For one thing, neither Ericsson nor Nokia appears to envisage any replacement of the underlying waveform, orthogonal frequency division multiplexing (OFDM), used in 4G and 5G networks and likely to be retained, with modifications, for 6G. The contrast is with Cohere Technologies, a noisy startup backed by several big telcos that links spectral efficiency gains to the use of an alternative waveform, orthogonal time frequency space (OTFS), or elements of it.
Regardless of that, RAN algorithms to their inventors are like Stradivarius violins, refined over years through trial and error to deliver the optimal results. AI would probably fail to design a better-sounding instrument than those made in the early eighteenth century by the famous Italian luthier. The performance of any RAN is also constrained by the Shannon-Hartley theorem, a hard limit (in principle) on how much information can be shoved into a given amount of bandwidth.
In comparing the different strategies of Ericsson and Nokia, the big question is whether a GPU offers benefits that custom silicon cannot. Their ability to do multiple calculations simultaneously has clearly given Nvidia’s chips an advantage over other forms of silicon in the training of AI’s large language models (LLMs). Yet the AI models deployed in the RAN need to be much smaller to avoid signaling delays. That could make custom silicon just as suitable as a slimmed-down GPU.
Marvell-less Nokia
Few, though, are prepared to dismiss Nokia’s claims, and the Finnish company has a growing roster of telco trialists for its AI-RAN products, including the giant operators T-Mobile US and Indonesia’s Indosat Ooredoo Hutchison. Nokia has framed its strategy as a pivot from custom to merchant silicon and promised the Airscale GPU cards will be available for the same price as older products that do not feature GPUs. Eventually, its custom silicon range will be phased out.
This has important ramifications for Marvell Technology, Nokia’s existing supplier of baseband silicon, and has generated some confusion over what Nokia will continue to support in the future. Essentially, what it means is that Marvell will remain an option for customers in the Layer 2-plus parts of the RAN software stack, where it offers chips under its Octeon brand that Atkinson regards as merchant-silicon central processing units (CPUs). In L1, where he does not see Marvell providing a merchant-silicon offer, its latest custom silicon chip will be the last that Nokia supports.
Within Nokia, that product is codenamed ABIP, in which the first three letters stand for “Airscale Baseband Indoor” and the fourth simply distinguishes it from earlier versions. It has jumped back in the alphabet to name its Nvidia card ABIG, with the G apparently short for GPU. “We have a lot of customers with Airscale installed base. They have a lot of predecessor cards – ABIL, ABIO, ABIP – running everything from 2G up to 5G today, and ABIG becomes a plugin,” said Atkinson. “So, ABIP will be the final custom silicon product which we will bring to market.” It will, he emphasized, be supported for “many years.”
The uncertainty is whether Marvell is being steered toward the exit or has voluntarily decided to find it. The chipmaker is known to have concerns about the business case for developing bespoke silicon when its customer has recently lost market share in the relatively small and shrinking RAN sector. “If there is no chip guy that’s willing to give you a chip other than Nvidia, then you have to use their narrative,” said Lum.
He also wonders why GPUs, if they are so advantageous, do not prominently figure in the strategies of the four other big RAN vendors, including China’s Huawei and ZTE. “The two guys who support 40% of the global market and the three biggest operators in the world have not chosen to go down the GPU path,” Lum said, noting developments within China. “Neither has Ericsson and neither has Samsung.”
The transition to Nvidia’s merchant-silicon GPUs would appear to have ramifications for another company in Nokia’s baseband mix, too. In the L1 domain of digital signal processing (DSP), Nokia has long relied on intellectual property from a company called CEVA. It is one of two relatively obscure players dominant in that space, the other being Tensilica (acquired by Cadence Design Systems in 2013). As DSP moves to GPUs, CEVA might no longer be needed.
“In the past, we had CEVA DSP cores which were integrated into the Marvell chip,” said Atkinson. “Of course, now we’re moving completely across to GPU paths, so that is where the heavy lifting is now ongoing in the R&D. There’s a large element of rewrite that we are doing.” AI tools are assisting Nokia with that rewrite, which Atkinson does not expect to be an overly cumbersome task.
Splits and chills
ABIG is also being designed to work with that existing Habrok range of massive MIMO radios. In today’s products, Marvell supplies L1 silicon for radio unit functions such as beamforming, just as it does for baseband’s distributed units (appliances and servers), while Broadcom’s radio unit chips are used for the digital front end, a function that converts analog signals into zeroes and ones. With ABIG, the challenge is using different silicon providers, and different software platforms, across the distributed and radio units.
Even so, Atkinson thinks it can be overcome by relying on open fronthaul interfaces, linking distributed and radio units, defined by the O-RAN Alliance. “We made a choice in the past to have the same silicon in the baseband as we have in the massive MIMO, but it was a marriage of convenience as opposed to a marriage of necessity,” he said. “You don’t need to have the same chip on both ends of the pipe.”
Baked into Habrok radios is support for a variant of the O-RAN Alliance’s open fronthaul interface called 7.2e. Perhaps less well known than the 7.2a and 7.2b variants, it would shunt more of the L1 processing into radios and is, according to Atkinson, the formal name for 7.3, a split option proposed by parts of the industry that was not O-RAN-compliant, he says.
Intriguingly, it is this 7.3 split that Nvidia reckons would shift functions like massive MIMO precoding and uplink receiver processing into radios. It “underscores the value of GPUs in handling the computational demands of these advanced functions,” said the company’s spokesperson. A massive MIMO unit that includes 128 transmitters and receivers – a so-called 128TRX configuration – would require about 32 times the processing load at the “lower PHY layer,” the part of L1 handled in the radio, the spokesperson added, noting that the load may grow exponentially “as we march toward 1024TRX and 7GHz in 6G,” demanding “robust and scalable compute architectures.”
If all this sounds more frightening to the telco executive than the average Hollywood horror, the conundrum is the choice between what are now radically divergent RAN paths offered by the two main vendors outside China. Nokia’s alignment with the world’s biggest company by market cap is certainly a more gripping story than Ericsson’s continued promotion of custom silicon. But the wrong choice could be a chiller.
https://www.lightreading.com/6g/nokia-weighs-nvidia-for-6g-radios-in-spectral-scrap-vs-ericsson
