Worried about the cost of developing custom silicon for a single RAN vendor, Marvell Technology reckons the market needs a common approach.
Jensen Huang, the boss of Nvidia, is known mainly for his leather jackets, AI evangelism and graphics processing units (GPUs), the semiconductors planted in the world’s data centers that have turned most of his employees into millionaires. But Huang has also become a constantly whirring cash dispenser for tech companies in his vicinity, spraying dollars that are gratefully snatched by Nvidia’s gang of allegiants.
Some of those are involved in telecom, an industry that seems to enthrall Huang. And a $2 billion investment in Marvell Technology, Nvidia’s latest move, is one of its most intriguing yet. It comes as Marvell urges 5G network vendors outside China to adopt a single silicon platform for the radio access network (RAN) products that each of them sells, Light Reading has learned. That silicon role is one Marvell is keen to play.
It does, however, seem to clash with the strategy pursued by Nvidia. Outlays that would have other investors nervously sweating into their suits are pocket change for the chip behemoth, with its fat margins, $60 billion in net cash and market cap north of $4 trillion. Nvidia’s Marvell deal followed investments worth $5 billion in Intel, $2 billion in each of Coherent and Lumentum and, notably, $1 billion in Nokia. The latter makes the Finnish vendor of network products Nvidia’s primary (some might say only) vehicle for the concept of AI-RAN, which proposes the use of GPUs at 5G and future 6G sites. Today, those are powered by Marvell’s custom silicon whenever Nokia sells RAN products. Nvidia’s partnership with Nokia has looked threatening to that arrangement.
Chip trade-offs
But even before Huang’s dollars showed up, there were doubts Nokia’s contract with Marvell could survive long into the 6G era. Inside Marvell itself, concern has risen about the expense of developing custom silicon products for a single player in a relatively small market. Last year, telcos spent just $35 billion worldwide on RAN products, according to Omdia, a Light Reading sister company, and the figure is down from $45 billion in 2022. Outside China, largely closed to foreigners, Nokia served just a quarter of all business. And silicon accounts for a small fraction of spending. In the 12 months leading up to November last year, Marvell earned just $542 million in revenues from all telco customers.
Samsung, its other major RAN customer, has already predicted the demise of custom silicon in the RAN’s baseband units (BBUs), where Marvell sits. “We’re still selling our purpose-built BBUs to a number of customers, but I do believe that it’s a matter of time,” Alok Shah, the vice president of network strategy for Samsung Networks, told Light Reading when asked the question at MWC Barcelona a few weeks ago. The bulk of what Samsung now sells is based instead on Intel’s central processing units (CPUs), the general-purpose x86 technology also sold to data centers and clients outside the telecom sector. In Samsung’s view, this “virtual RAN” looks more economically sustainable than the purpose-built variety.
The GPUs of AI-RAN fall into the same bracket, attractive partly because the same underlying technology supports other needs, appears far more profitable and can draw on Nvidia’s huge resources. Yet CPUs and, especially, GPUs are not an ideal fit for the RAN. Despite improvements, their general-purpose nature makes them less efficient at performing certain baseband tasks than custom silicon, designed specifically for that job.
On a scale, CPUs and GPUs lie at one extreme, with more compelling investment returns but suboptimal performance measures. Silicon customized for one RAN vendor is at the opposite end: While it excels at the job, it is growing harder to justify economically. Marvell’s idea is that a baseband chip designed for the whole industry, excluding China’s Huawei and ZTE, would be an effective compromise.
Different strokes
Today, however, the silicon strategies of the main RAN vendors are diverging. Ericsson, the largest, remains wedded to the custom silicon it develops internally and rejects the argument that its investments are unsustainable. “If you were to start with a blank sheet of paper and say I’m going to design the world’s greatest purpose-built silicon, it would be very expensive,” said Michael Begley, Ericsson’s head of RAN compute, at MWC. “But when you have 30 years of investment and you’re improving it in each and every generation, it’s not the same investment level that you’re talking about.”
Some analysts are unconvinced. Separate from this, Ericsson has developed its own virtual RAN software that runs, like Samsung’s, on Intel’s CPUs. After some tweaks, Ericsson believes the same code could be deployed on CPUs from alternative vendors. They include AMD, which uses the same x86 architecture as Intel, and Nvidia, whose CPU (not GPU) product is based on Arm, the main alternative. Virtual RAN addresses different customer demands, says Ericsson. Others think Ericsson may be hedging its bets.
Before Nvidia announced its investment in Marvell, Nokia appeared to have swung around to Shah’s view that it’s “a matter of time” before custom silicon is phased out. At its capital markets day in November, weeks after it pocketed $1 billion from the Huang cash dispenser, CEO Justin Hotard told analysts he was “moving away from a legacy hardware model” and making a “shift from proprietary to general-purpose hardware.”
Marvell is understood to have contracts in place with Nokia for the shipment of commercial RAN products until the mid-2030s. But it is alert to the risk of being dropped. Technically, this would become more feasible if Nokia is eventually able to use Nvidia’s GPUs in the radio products for massive MIMO, an advanced technology that relies on bigger volumes of antenna elements.
Nevertheless, while Ericsson still professes commitment to custom silicon, Nokia has appeared to be keeping it in the portfolio as a backup should Nvidia not work out. It is under far more pressure than Ericsson to slash costs, too. Despite removing thousands of jobs, and having a smaller budget for RAN research and development than its Swedish rival, Nokia watched the operating margin at its mobile networks business group halve last year, to just 2.8%.
And now for something completely new
For all its efforts to build and acquire RAN expertise, Nvidia might see considerable value in Marvell’s assets and experience. Those include intellectual property for multiple codecs and algorithms used in Layer 1, the slice of the RAN stack that is the most dependent on the underlying hardware. Among them are features with such arcane names as Polar, Viterbi and Turbo.
This does not neatly align with Nvidia’s mission of turning Layer 1 into code that is compatible with CUDA, its software platform, and running it on GPUs. To Nvidia, Marvell’s main attraction could be its expertise in developing connectivity chips for data centers. Its digital signal processors for optical equipment are mentioned in Nvidia’s statement about the deal. Yet as a new investor, Nvidia has an interest in seeing Marvell succeed in other areas, too. AI-RAN is also specifically referenced in the statement.
An Nvidia-backed Marvell offering a custom silicon platform to everyone would be hard for RAN vendors to ignore. It would, of course, require them to collaborate on defining a common set of requirements. If widely adopted across all products, it would kill off any opportunities to differentiate through hardware. But an industry pivot to CPUs or GPUs would potentially do the same.
Disregarding the Chinese vendors, the Layer 1 silicon used in today’s BBUs comes almost entirely from Ericsson, Intel, Marvell and Qualcomm, which figures to a much smaller degree. Commercial realities and their own parsimony have thwarted telcos determined to broaden the choice. And no one is thriving. With Nvidia at its side, and its share price enjoying the connection, Marvell has given operators something new to consider.
https://www.lightreading.com/5g/nvidia-backed-marvell-pitches-one-chip-to-rule-the-ran
