“Reproducing kernel Hilbert house” (RKHS) isn’t a sequence of phrases that will be strung collectively in lots of conversations, even by somebody immersed in expertise at Nokia. It describes an algorithm utilized in machine studying, little identified outdoors skilled circles, and has been largely ignored by Nokia’s radio entry community (RAN) individuals just because the infrastructure could not deal with it. “We by no means used it as a result of it’s totally mathematically compute-intensive,” stated Pallavi Mahajan, the Finnish firm’s chief expertise officer. “It has loads of vector computation.”
That every one modifications now. “With GPUs, you are able to do that, and also you get effectivity,” stated Mahajan. These GPUs are the graphics processing items designed and offered by Nvidia, the world’s greatest firm by market capitalization. They’ve underpinned the generative AI growth, offering a way more highly effective platform than older central processing items (CPUs) to coach giant language fashions (LLMs). Nokia’s consultants consider they could possibly be equally beneficial within the RAN because the trade begins work on 6G.
Round ten months have now elapsed since Nokia accepted a $1 billion funding from Nvidia. That got here with a sector-shaking announcement of plans to design RAN software program based mostly on CUDA, the chipmaker’s platform, and make GPUs integral to Nokia’s future 5G and 6G merchandise. For a time, these will coexist with the customized silicon that Nokia obtains from Marvell Know-how, its longstanding 5G companion. However Nokia has been emphatic: the Marvell chips in its latest merchandise would be the final customized silicon it ever develops.
An Nvidia GPU will present a {hardware} basis capable of stand up to something 6G throws at it for a very long time – or so Nokia believes. “No matter we do any more, it is similar to six years of software program upgrades,” stated Mahajan. The power to make use of RKHS is merely one attraction of transferring to a GPU. “It permits us to exit and run totally different algorithms, which beforehand we couldn’t,” she defined.
Nokia’s declare that each one this might double spectral effectivity by 2028 has provoked various responses. Some have proven outright disbelief. A number of have reacted with drooling tongues, like ravenous diners impatiently awaiting a banquet. The promised enchancment dramatically and pointedly exceeds what rival Ericsson beforehand stated was doable on customized silicon operating AI software program. Like Marvell’s chips, that may carry Ericsson solely to date in 6G earlier than it wants changing, Nokia reckons.
What the FEC?
But the concept 6G will in the end demand a GPU is soundly rejected by the Swedish vendor. “As beforehand communicated, GPUs are one possibility for 6G and AI RAN, not the one one,” stated a spokesperson for the corporate by e-mail earlier than harping on the enchantment of customized silicon. “Goal-built silicon stays related the place operators want efficiency, decrease power consumption, and the flexibility to scale RAN workloads effectively. We count on future networks to make use of a mixture of compute architectures, relying on the use case and operator necessities.”
Ericsson isn’t completely hostile to the idea of a GPU RAN and would most likely not want to be too vital of a chipmaker that excites inventory markets like no different. Even when it has completed so with out a lot enthusiasm, it has for years supplied prospects a set of digital or cloud RAN merchandise based mostly on CPU structure, somewhat than customized silicon. Working an energy-hungry activity referred to as ahead error correction (FEC) on these CPUs gobbles an excessive amount of energy to fulfill telco calls for. Intel stumps up a {hardware} accelerator to deal with FEC and integrates that with its newest CPUs. Another that Ericsson has investigated would exchange Intel’s Xeon processor with an Nvidia Grace CPU and depend on a GPU because the FEC accelerator.
However Ericsson’s consultants evidently disagree that GPUs would confer AI-based RAN advantages unrealizable with CPUs or customized silicon. Largely, that’s as a result of there are constraints outdoors the info middle to do with energy consumption and latency, a measure of the time a community sign takes to finish its journey. RAN fashions should essentially be extra light-weight, and meaning they don’t require a GPU, based on Ericsson.
As an analogy, image a constructing website the place objects of varied weights have to be moved from one level to a different. Solely the strongest two or three staff can carry the heaviest gadgets. However using them to hold lighter items can be a wasteful misallocation of assets.
A RAN constraint that nobody can keep away from is named the transmission time interval (TTI). “Now we have 500 microseconds, half a millisecond, to get this completed,” defined Michael Begley, Ericsson’s head of RAN compute, when he met with Mild Studying at this yr’s MWC Barcelona tradeshow. “So, you’ll be able to’t do a million-parameter mannequin. You have to compress it. In any other case, there is not any level as a result of in case you miss the TTI, radio circumstances have modified and no matter you calculated is not related anymore.”
AI fashions that embrace tens of millions, billions and even trillions of parameters may want the muscle of a GPU. However when they’re restricted by laborious science to solely tens of hundreds of parameters, Ericsson’s AI-RAN fashions don’t. “These sorts of guidelines of physics and specs are the identical for Nvidia,” stated Gabriel Foglander, Ericsson’s head of RAN technique, throughout a newer assembly on the firm’s Stockholm headquarters. “Inferencing will must be met, no matter compute you wish to put in there.”
Sports activities vehicles and SUVs
Related opposition to Nokia’s argument that 6G will in the end necessitate a GPU RAN comes from Samsung, the world’s fifth-largest 5G vendor. “Samsung doesn’t consider that GPUs are a prerequisite for 6G, nor {that a} CPU-only or GPU-only method is the suitable alternative,” a consultant stated by e-mail. “We see CPUs persevering with to function the first general-purpose basis for RAN, with GPUs or different accelerators added the place extra efficiency is required and the workload can profit from it.”
However there are some essential nuances that distinguish Samsung from Ericsson. The Swedish vendor’s AI-RAN technique is geared closely to its personal customized silicon, which accounts for almost every little thing it sells as we speak. Samsung, against this, is now primarily invested in digital RAN merchandise based mostly on Intel’s CPUs. That would clarify why Samsung has expressed barely extra curiosity in utilizing GPUs for AI-RAN.
The South Korean vendor likens a CPU to a sports activities utility car (SUV) and a GPU to an precise sports activities automotive. “A sports activities automotive could also be quicker, however an SUV is extra versatile for on a regular basis use – carrying individuals, touring lengthy distances and working effectively throughout totally different wants,” stated the corporate in its e-mail. A CPU offers a general-purpose basis for the on a regular basis wants of RAN compute. However sure “intensive” RAN workloads may profit from the high-performance compute supplied by a GPU.
The bounds imposed by the TTI make CPUs a extra “pure match” for many duties, in Samsung’s opinion. On the identical time, it echoes the views of Nokia’s Mahajan concerning the usefulness of GPUs for computationally demanding workloads. Nvidia’s chips are, it elaborates, “essentially optimized for extremely parallel processing and may execute extremely parallel workloads in a short time.”
The problem appears to be the fixed must cram sufficient information right into a GPU’s hungry mouth to make system-wide use of them look economical. Extra information equals extra throughput, and extra throughput dangers rising system latency and overstepping the TTI mark. “This creates a trade-off between maximizing GPU efficiency throughput and assembly tight TTI necessities,” stated Samsung in its e-mail. “To satisfy the required TTI, a few of that throughput benefit could due to this fact must be sacrificed.”
Including intrigue is Nokia’s declare that it has been capable of notice the dramatic enhancements in spectral effectivity and different measures with out boosting power consumption. Its forthcoming GPU merchandise embrace a card designed to fit into an current Airscale-branded baseband chassis and execute the identical duties as as we speak’s Marvell playing cards for newer releases of the cell normal.
“They work with an influence envelope of 190 watts. Now that we’re bringing within the new card, it has to work in precisely the identical energy envelope,” stated Mahajan. “We can not have our prospects change their energy cabling, their racks and all of that stuff. Nevertheless it nonetheless brings 1.5 instances the capability, two instances the spectral effectivity.”
In keeping with one trade view, a GPU able to addressing later-stage 6G wants would essentially be over-specified for as we speak’s necessities. But Nokia’s roadmap consists of an array of merchandise, from the power-limited card usable with as we speak’s Airscale chassis to a extra superior equipment. “It is a compact pizza-box type issue the place you’ve each GPU and CPU compute inside,” stated Mark Atkinson, the pinnacle of Nokia’s RAN enterprise. “We’ll then have the ability to have the total advantage of GPU as a result of we can’t be restricted by any energy headroom anymore.”
Working with ASICs
Nonetheless, if GPUs are superior to CPUs for operating sure AI algorithms, the declare they are going to beat an application-specific built-in circuit (ASIC) is extra controversial. Google, notably, has been shifting from GPUs to what are referred to as tensor processing items (TPUs) for the coaching of LLMs. Co-designed with chipmakers Broadcom and Marvell, these TPUs are a sort of ASIC optimized for AI compute. To Google, they clearly maintain sights in contrast with Nvidia’s GPUs. Each Anthropic and OpenAI, the main US builders of frontier fashions, additionally reportedly now use TPUs for LLM coaching.
Apart from growing customized silicon for its baseband tools, Ericsson has just lately began introducing an ASIC it calls a “neural community accelerator” into its superior, large MIMO radios. That enables it to assist options together with an AI-native scheduler for hyperlink adaptation, AI-managed beamforming and AI-powered macro positioning, stated Foglander. “It is mathematics-heavy processing within the radio,” he defined.
There may be actually nothing within the emergent 6G normal that mandates use of a GPU, based on educated sources. No matter the professionals and cons of ASICs, CPUs and GPUs, market forces and chronic considerations about “vendor lock-in” could also be extra important elements in decision-making. In 5G, Nokia had at all times seemed extra dependent for its customized silicon on an exterior participant – specifically Marvell – than Ericsson ever did. Justifying the funding would have turn into more durable after Nokia’s lack of market share within the profitable US. Desirous to discover a shopper for its GPU RAN pitch, Nvidia may need been the one viable possibility.
The criticism is that Nokia has traded lock-in with Marvell for lock-in with Nvidia and CUDA, deemed incompatible with different silicon platforms. “If the GPU factor does not catch on, they do not have another,” stated Vincent Loncke, the founding father of a startup referred to as Kenyi Applied sciences and a former Qualcomm government.
In contrast, the flexibility to maneuver software program written for Intel’s CPUs to different x86- or Arm-based chips, with solely minor modifications, appears to offer Ericsson and Samsung a point of independence in digital RAN. To this point, nonetheless, almost all deployments are based mostly on CPUs from Intel, a enterprise that reported a $15 billion web loss for the primary half of 2026 and one whose workforce shrank by 46,800 workers, greater than a 3rd of the sooner complete, between 2022 and 2025.
Apart from Huawei, Ericsson is the one RAN vendor nonetheless investing in its personal customized silicon. Per Narvinger, its incoming CEO, sounds dedicated to that technique and has repeatedly rejected assertions from outdoors that Ericsson will finally must bow to economics and embrace general-purpose {hardware}. Because the world’s greatest RAN vendor excluding China, the Ericsson of 2026 can nonetheless make the numbers add up. But when telcos spend much less on 6G than they’ve on 5G, or Nokia can stage a comeback, that formulation might begin to look shaky.
