Having already pitched its GPUs for use in RAN compute’s servers and appliances, Nvidia is working on an offer that would go directly into radios.
Long before the current AI boom, Nvidia’s graphics processing units (GPUs) were barely known outside the computer games industry. Their ability to do parallel processing – multiple calculations at once – later captured the interest of AI’s developers, transforming Nvidia into the world’s most valuable company. These days, CEO Jensen Huang seems determined to insert his chips into everything. A PC chip, codeveloped with Taiwan’s MediaTek, is promised this year. There are GPUs for cars, GPUs for robots, even GPUs for turning a home into a mini data center. And as part of this GPU colonization, Nvidia is working on a chip for the 6G radio.
The move, confirmed by Nvidia to Light Reading, marks a dramatic new development in the GPU giant’s “AI-RAN” strategy. Previously, it had demonstrated how Grace Hopper, a superchip combining an ordinary central processing unit or CPU (the Grace part) with a more powerful GPU (Hopper), could replace the custom silicon in a radio access network (or RAN). Grace Hopper and related products would take over the RAN compute that happens in appliances or servers, commonly referred to as central units (CUs) and distributed units (DUs) by the industry. This left the antenna-hosting radio unit (RU) clamped to a mast or sited on a rooftop – the other half of the RAN equation.
Unlike the CUs and DUs, those RUs have not formerly been a target for Nvidia or the makers of general-purpose processors. Years before Nvidia waded into the RAN, Intel had similarly pitched its CPUs as an attractive substitute for RAN compute’s application-specific integrated circuits (ASICs). This “virtual RAN” forerunner of AI-RAN would theoretically allow the telecom industry to ride on chips sold to a much larger audience and therefore enjoying better returns on investment. But Granite Rapids, Intel’s latest virtual RAN product, includes no RU component, the company has confirmed. Nor does Intel have plans to design one.
Massive MIMO changes everything
Why would it need to? With simpler 4G and 5G radios, the Layer 1 or physical layer (also called PHY) processing, the slice of RAN compute hungriest for IT resources, happens in the DU. Anyone dissecting RUs with a relatively small number of antenna elements would find transceivers, data converters and digital front ends for turning analog signals into zeroes and ones. But they are unlikely to encounter an ASIC that plays a significant Layer 1 role.
That changes in massive MIMO, an antenna-rich technology included in today’s more advanced 5G radios. Here, the Layer 1 functions are divided between the DU and the RU to avoid any performance degradation. Among other things, the RU must guarantee support for beamforming, a clever technology that allows signals to pinpoint devices instead of wastefully carpeting bigger zones. On the hardware side, the conventional approach is to include an ASIC in the RU for beamforming and other such “low PHY” network functions.
It is these ASICs in the RU that a GPU would replace. In detailed commentary shared with Light Reading by email, Nvidia says the move is necessary as RUs become increasingly sophisticated. Basic radios include four transmitters and receivers. 5G Advanced and 6G could boost the number to 128, demanding 32 times as much processor power for low PHY compute. With so-called ultra-MIMO, deployed in 6G’s higher spectrum bands, RUs featuring up to 1,024 transmitters and receivers are even now imaginable. “As ultra-MIMO, 7GHz and AI algorithms are introduced in 6G RU, GPUs would become essential to meet the compute demands,” said Nvidia.
Any lack of an RU chip would conceivably limit the 5G and 6G opportunity for Nvidia. In massive MIMO, the same silicon provider has tended to bridge Layer 1 across the DU and the RU. Using different providers would require the Layer 1 software developer to work with two separate platforms.
In open RAN, a DU from one company could, in principle, be joined to an RU from another via a standardized interface. But this arrangement would look even more complicated. It would also require the software developers on either side to disclose their closely guarded algorithms to each other, according to one RAN expert who spoke on condition of anonymity. Their reluctance to do so partly explains why almost no multivendor massive MIMO has been commercially deployed, he said.
But anyone using Granite Rapids to support massive MIMO would be forced to pair an Intel processor in the DU with something else in the RU. Intel does not see that as a problem. Defined by a group called the O-RAN Alliance, open RAN’s 7.2x interface is designed to address any interoperability concerns and means “performance is not dependent on the same silicon being used in the DU and the RU,” it said. Samsung, the vendor that has deployed more virtual RAN products than anyone else, has accordingly relied on a homegrown ASIC for low PHY processing in its RUs, it confirmed to Light Reading.
Under the strict definitions previously applied by Intel, however, the use of custom silicon for part of the compute would disqualify this as a fully virtualized RAN. In the past, experts at Intel were critical of Nokia for advertising a virtual RAN product that kept all Layer 1 functions on custom silicon from Marvell Technology. What the industry took to calling virtual RAN with an “inline” accelerator used an Intel CPU only for the higher layer functions. Samsung, similarly, has had to carve out part of the Layer 1 that could not be virtualized.
What Nvidia now proposes would, effectively, change all that, giving vendors more “flexible, software-defined compute platforms rather than fixed-function silicon,” said the GPU maker in its emailed remarks. Given the nature of custom silicon, with its tight coupling of hardware and software, this part of the RAN compute has remained off limits to the broader developer ecosystem. Nvidia claims to have thrown open the doors to any expert familiar with CUDA, its software platform, which already boasts about 6 million developers. It has gone even further by designing its own CUDA-based RAN compute architecture, branded Aerial, with which anyone is free to play.
But GPUs are power hogs, aren’t they?
There is, nevertheless, industry skepticism about introducing GPUs into the RAN. Even some of the people who have overcome their reservations about putting GPUs into DUs doubt they would be economical in RUs. According to one estimate, RUs account for up to 90% of the energy consumption by a mobile network. Anything that increases consumption even slightly could render them unviable, and GPUs still have a notorious reputation as power hogs.
But Nvidia and its allies say that the energy-hungry GPUs aimed at data centers cannot be compared with those it is developing for the RAN. In similarly “constrained environments such as automotive and robotics,” it already has “embedded systems capable of operating at power levels below 100W and temperatures of 100 degrees Celsius,” said the company. The GPU that goes into an RU is more likely to resemble the GPU designed for gaming, said one source.
There is intrigue, too, in the relationship that appears to have developed between Marvell and Nvidia this year. In March, Nvidia made a $2 billion investment in Marvell while sounding more interested in the latter’s optical expertise than its RAN smarts. Marvell, however, is a vendor of RAN silicon to both Samsung and Nokia, which has also secured a $1 billion investment from Nvidia and announced plans to develop GPU-compatible RAN products.
Then, in late May, came the following statement from Marvell CEO Matthew Murphy during a call with analysts about the latest financial results: “Marvell will enhance its existing Octeon basestation processors to work directly with Nvidia GPUs, integrating AI with wireless infrastructure on a single software-defined computing platform.”
“This will enable telecommunications operators to run both 5G and 6G radio workloads and high-performance AI applications concurrently on the same hardware,” he added. While the details are currently thin, it all implies Nvidia has seen the attractions of drawing on Marvell’s RAN expertise just as it has leaned on MediaTek in PCs.
Any general-purpose chips built to support multiple needs will always lag the performance of custom silicon designed for one thing, insist experts. Few are prepared to disagree. But investing exorbitant sums to develop ASICs for the RAN has become harder to justify when those chips have no audience beyond telcos. Last year, operators worldwide spent just $35 billion on RAN products, according to Omdia, a Light Reading sister company. The amount has fallen from $45 billion in 2022, and no one expects a recovery.
Should ASIC development for the RAN become uneconomical, there is a scenario in which newer general-purpose chips can outperform the older custom silicon that is still in use. “The key consideration is not a single power number, but as RU functionality becomes more complex and increasingly AI-native, the economic trade-off shifts toward programmable platforms that can evolve with standards and flexible deployment models, rather than fixed designs optimized for a single configuration,” said Nvidia. In the absence of general-purpose silicon for the RU from Intel or anyone else, Nvidia might be the only choice there is.
https://www.lightreading.com/6g/nvidia-has-a-radical-new-ai-ran-plan-a-6g-radio-unit-chip?shem=rimspwouoe,
