Samsung, Nokia и Ericsson ведут борьбу за вычислительную основу 6G: выбор между GPU и ASIC

The race to define 6G's computing foundation has become a three-way standoff between the world's biggest radio access network vendors — and a detailed analysis published by Light Reading on August 25 reveals that Samsung's specific objection to the GPU-versus-ASIC framing has been almost entirely missing from the public debate. Nokia has committed to building its 6G radio access network software on Nvidia graphics processing units. Ericsson is betting on custom application-specific integrated circuits it designs in-house. And Samsung — the world's fifth-largest 5G vendor — says neither GPU-only nor ASIC-only is the right answer, and invokes a hard physics constraint that applies equally to Nokia and Nvidia. Operators who must choose a long-duration infrastructure path before 6G standards freeze in March 2029 are watching all three arguments simultaneously and committing to none. The divergence is not primarily about technical disagreement in the abstract. It reflects three fundamentally different assessments of what the physics of radio processing will and will not allow — and a strategic wager on which bet will prove correct before operators lock in infrastructure commitments that will outlast the standards-empty window between 5G Advanced and the 6G specification freeze. Nokia's GPU Bet: An Algorithm Previously Left on the Shelf In October 2025, Nokia accepted a Nvidia $1 billion equity investment from Nvidia, giving the chip company a roughly 2.9% stake at $6.01 per share, and announced plans to build its radio access network software on Nvidia's CUDA platform. The deal, accompanied by a T-Mobile commitment to trial the resulting platform, was widely read as a generational pivot — and Nokia has since been explicit that the Marvell custom silicon in its newest 5G products will be the last custom silicon it ever develops. Nokia's technical argument centers on a class of machine-learning algorithms called reproducing kernel Hilbert space methods — RKHS, in shorthand — that are mathematically intensive enough that RAN engineers have simply never used them, because prior baseband hardware could not handle them. "We never used it because it's very mathematically compute-intensive," said Pallavi Mahajan, Nokia's chief technology officer. "It has a lot of vector computation. With GPUs, you can do that, and you get efficiency." The commercial ambition that follows is substantial. Nokia launched its anyRAN commercial AI-RAN platform on July 15, 2026, targeting more than 100% spectral efficiency gains by 2028 — meaning operators could serve roughly twice their current traffic volume from the same licensed spectrum without buying additional airwaves. Its near-term product roadmap includes a GPU card designed to slot into an existing Airscale-branded baseband chassis, operating within the same 190-watt power envelope as the Marvell silicon it replaces while delivering 1.5 times the capacity and double the spectral efficiency. A more advanced form-factor combining GPU and CPU compute — a dense appliance Nokia calls a "pizza-box" — is also in development for deployments where power headroom is not constrained. Nine tier-one operators have engaged on Nokia's AI-RAN platform. T-Mobile holds the only confirmed timeline: pilots are scheduled to begin by end of 2026, with commercial service expected in 2027 and mass deployment targeted for 2028. The other eight operators are engaged but have not been publicly confirmed to specific milestone dates. Ericsson's Counter: Physics Doesn't Care About GPU Marketing Ericsson, the world's largest radio access network vendor outside China, is unmoved by Nokia's GPU thesis — and its objection is rooted in a hard constraint of radio physics rather than product rivalry. Every RAN system must complete its signal-processing cycle within the transmission time interval: approximately 500 microseconds, or half a millisecond. In 5G New Radio operating at the 30-kilohertz subcarrier spacing most common in mid-band deployments, one slot — one complete transmission unit — lasts exactly 0.5 milliseconds. That ceiling has a direct implication for AI model size. "We have 500 microseconds, half a millisecond, to get this done," explained Michael Begley, Ericsson's head of RAN compute, speaking at MWC Barcelona this year. "You can't do a million-parameter model. You need to compress it. Otherwise there's no point, because if you miss the TTI, radio conditions have changed and whatever you calculated isn't relevant anymore." AI models at data-center scale — the kind that benefit most from GPU parallelism — run to billions or trillions of parameters. RAN AI models, constrained by the timing budget, run to tens of thousands. At that size, Ericsson argues, a GPU's advantages largely disappear. Gabriel Foglander, Ericsson's head of RAN strategy, was more direct about the implication: "Those kinds of rules of physics and specifications are the same for Nvidia. Inferencing will need to be met, whatever compute you want to put in there." Ericsson's preferred answer remains application-specific integrated circuits built in-house under its EMCA (Ericsson Modular ASIC) architecture. Its latest generation, the RAN Processor 6672 built on Intel's 4 process node, doubles baseband capacity at half energy of its predecessor while consuming less than half the energy. As of late July 2026, NTT Docomo's deployment of the 6672 completed a sweep of all three major Japanese carriers onto Ericsson's custom-silicon platform. Ericsson is not philosophically opposed to GPUs for every RAN workload. It has investigated pairing an Nvidia Grace CPU with a GPU as a forward error correction accelerator for its virtual RAN products. And it recently began embedding a chip it calls a "neural network accelerator" ASIC into its advanced Massive MIMO radios, enabling an AI-native scheduler for link adaptation, AI-managed beamforming, and AI-powered macro positioning. What Ericsson flatly rejects is the claim that GPUs are a prerequisite for 6G AI-RAN — the specific language Nokia and Nvidia have used to describe the market. Ericsson launched its AI in RAN software commercially in June 2026 as a software-only subscription that operators can activate on hardware already deployed in their networks, without a hardware replacement cycle. More than 15 carriers have deployed it commercially, including SoftBank, Bell Canada, SK Telecom, and Rogers, delivering up to 10% better spectral efficiency and up to 20% higher downlink throughput in live networks. Those figures represent roughly one-tenth of Nokia's 2028 roadmap targets — but Ericsson's numbers exist in real commercial networks today, while Nokia's await pilots that are not scheduled to begin until the end of 2026. Samsung's Third Path: Why the SUV Beats the Sports Car in Everyday Traffic Samsung's position is the least widely covered of the three, and it represents the most structurally nuanced argument in the debate. "Samsung does not believe that GPUs are a prerequisite for 6G, nor that a CPU-only or GPU-only approach is the right choice," the company said in a written statement to Light Reading. "We see CPUs continuing to serve as the primary general-purpose foundation for RAN, with GPUs or other accelerators added where additional performance is needed and the workload can benefit from it." Samsung describes the CPU as a sport utility vehicle and the GPU as a sports car. "A sports car may be faster," Samsung wrote, "but an SUV is more versatile for everyday use — carrying people, traveling long distances and operating efficiently across different needs." The metaphor captures a real engineering reality. Samsung has invested more heavily in virtual RAN products built on Intel and AMD CPUs than any other major vendor, and its architecture separates radio workloads from AI inference workloads in a way that allows hardware to be upgraded incrementally rather than all at once. Samsung's position is partly a function of its product mix. Unlike Ericsson, whose revenue depends heavily on custom silicon, Samsung has built its RAN business on software that runs on Intel Xeon 6 and AMD EPYC processors — and has validated all three simultaneously alongside Nvidia's ARC Compact platform. That hardware-agnostic approach is specifically what the "refuses to choose" framing reflects: Samsung is not making the same all-in architectural commitment that Nokia has made to CUDA. Samsung also agrees with Ericsson on the TTI constraint, and makes the specific mechanism explicit. The challenge with GPUs in RAN, Samsung notes, is that they are most efficient when fed large volumes of data simultaneously — but maximizing data throughput risks increasing system latency and exceeding the TTI timing budget. "This creates a trade-off between maximizing GPU performance throughput and meeting tight TTI requirements," Samsung stated. "To meet the required TTI, some of that throughput advantage may therefore need to be sacrificed." That is not an argument against ever using a GPU in RAN. It is an argument that the conditions under which a GPU's advantage is fully realizable are narrower than Nokia's marketing implies. Does CUDA Lock-In Trade One Problem for Another? Beneath the technical debate runs an uncomfortable strategic question: who ends up locked in to whom? Nokia's critics point out that replacing Marvell silicon with Nvidia CUDA is not an escape from vendor dependency — it is a change of landlord. CUDA is a proprietary software platform widely considered incompatible with other GPU vendors' silicon, meaning that Nokia's growing library of CUDA-optimized RAN software would require a substantial rewrite to run on AMD, Qualcomm, or any other chip platform. "If the GPU thing doesn't catch on, they don't have an alternative," said Vincent Loncke, founder of startup Kenyi Technologies and a former Qualcomm executive. Ericsson and Samsung, by contrast, argue that software written for Intel x86 or Arm-based CPUs can be ported to alternative CPU vendors with relatively minor changes — a degree of hardware independence that CUDA-based development does not offer. There is merit to this argument in principle. In practice, however, nearly all virtual RAN deployments currently run on Intel silicon, which creates its own form of de facto concentration even if the theoretical switching cost is lower. Intel itself is in a period of structural transition. Its workforce shrank by approximately 46,800 employees between 2022 and 2025, and the company posted a combined GAAP net loss of combined $14.7 billion H1 GAAP loss, driven primarily by a $12.5 billion non-cash mark-to-market loss tied to its CHIPS Act government agreement rather than operational failure. On a non-GAAP operating basis, Intel's revenue grew 25% year over year in Q2 2026, its strongest quarterly growth in more than fifteen years, led by Data Center and AI demand. But the structural workforce contraction and the company's long-running manufacturing gap relative to TSMC represent genuine platform risk for RAN vendors who have built their virtual RAN products around Intel's roadmap. Samsung, notably, has hedged this risk by simultaneously qualifying AMD EPYC alongside Intel Xeon — hardware-vendor agnosticism made real, not just claimed. What the 6G Standard Actually Requires: Silence on Silicon One point where all parties agree: the emerging 6G standard mandates nothing about processor type. The International Telecommunication Union and 3GPP standards body are focused on spectrum requirements, latency targets, and throughput specifications in Release 21 — not on the silicon vendors use to meet them. 3GPP locked the Release 21 timeline in June 2026, setting a specification freeze targeted for no earlier than March 2029. That is the point at which chip vendors can begin designing certifiable 6G silicon. Until then, every vendor's position is a bet made in a standards-empty window. That ambiguity is precisely what makes the current divergence so consequential. Carriers evaluating 6G roadmaps are making decade-scale infrastructure commitments before the architecture question is settled. Base station hardware qualification cycles run 12 to 24 months per component change, meaning an operator that deploys Ericsson's ASIC-based platform at scale faces significant practical barriers to switching architectures mid-cycle even if Nokia's GPU approach delivers its promised gains. The same logic applies in reverse: an operator that begins writing AI-RAN workloads in CUDA and builds operational tooling around Nokia's platform faces non-trivial migration costs if the spectral efficiency promises prove undeliverable in live dense urban networks, or if Nvidia's strategic priorities shift in ways that affect the RAN software roadmap. Samsung's platform — hardware-agnostic vRAN running on Intel, AMD, and selectively on Nvidia GPUs — represents the lowest-switching-cost position of the three, but also the one with the most modest near-term performance claim. What Operators Actually Have to Decide For a carrier CTO evaluating these three paths, the decision variables are not primarily about which architecture is theoretically correct. They are about timing, risk, and lock-in depth. Ericsson's proposition: AI-optimized RAN software with measured commercial gains (10% spectral efficiency, 20% downlink throughput) is available today, on hardware already deployed, at 15+ carriers globally. The physics argument against GPU superiority is made by the company that has the most to lose if GPUs do eventually deliver Nokia's claimed gains — which means its credibility depends entirely on whether the TTI constraint actually limits model complexity the way Ericsson's engineers claim. If Nokia deploys at T-Mobile in 2027 and measures 50% real-world spectral efficiency gains, Ericsson's ASIC thesis will face exactly the kind of commercial pressure it currently dismisses. Nokia's proposition: A more computationally ambitious AI-native RAN that cannot be proven until pilots begin at end of 2026 and commercial deployments start in 2027. The claim is five to ten times the gains Ericsson has measured, backed by a $1 billion chip-company investment and nine named tier-one operators engaged but not yet committed to deployment dates beyond T-Mobile. If the RKHS algorithms and deep-learning receivers deliver in live dense urban networks, Nokia will have closed the argument. If they deliver only a fraction of the roadmap targets in real commercial conditions — as dense-cell physics and fronthaul split constraints may dictate — the lock-in costs will come due. Samsung's proposition: Maximum hardware flexibility at the cost of maximum performance ambiguity. The CPU-first architecture delivers measurable AI-RAN gains today; selective GPU augmentation for specific parallel workloads adds performance where the TTI budget allows. No platform commitment that forecloses future silicon options. The tradeoff is that Samsung is not making a headline architectural claim that generates the same operator urgency as Nokia's 100% spectral efficiency target. Ericsson's custom silicon strategy remains viable at current market scale — with the annual RAN market at $35 billion and declining from a peak of $45 billion in 2022, the economics of sustaining multi-generational ASIC development depend on Ericsson's continued market leadership. Per Narvinger, Ericsson's incoming CEO who takes the role in October 2026, has repeatedly rejected assertions that Ericsson will eventually bow to economics and embrace general-purpose hardware. Whether his position remains tenable depends significantly on how Nokia's 2027 commercial deployments perform. The silicon bets being placed now, in the standards-empty space between 5G Advanced and a 6G specification that will not freeze until March 2029, may prove harder to undo than the vendors — or their customers — currently appreciate. Three architectures. Three sets of promises. One set of physics. And a half-millisecond window that applies equally to every chip in the contest. Frequently Asked Questions Does 6G require Nvidia GPUs, or can it use other chips? The emerging 6G standard mandates nothing about processor type. Nokia argues that GPUs are necessary to run computationally intensive AI algorithms — including reproducing kernel Hilbert space methods — that ASIC and CPU hardware cannot execute efficiently. Ericsson and Samsung both reject this claim, pointing to the transmission time interval constraint: the half-millisecond window within which RAN processing must complete limits AI models to tens of thousands of parameters, not the billions that require GPU-class compute. At that model size, both companies argue, purpose-built ASIC or general-purpose CPU architecture performs as well or better than a GPU, with lower energy consumption and lower lock-in risk. Nokia's commercial AI-RAN platform is expected to begin pilot deployments by end of 2026, which will provide the first real-network evidence for both sides. What is the TTI constraint and why does it limit GPU advantages in 6G networks? The transmission time interval is the minimum time unit during which a base station completes a radio scheduling cycle. In 5G New Radio at 30-kilohertz subcarrier spacing — the most common configuration in mid-band deployments — one slot lasts 0.5 milliseconds. All signal-processing computations, including any AI inference, must complete within that window before radio conditions change and the calculation becomes irrelevant. That timing budget limits how many parameters a RAN AI model can contain: where data-center AI models run to billions of parameters and benefit enormously from GPU parallelism, RAN AI models are limited to tens of thousands of parameters by physics. At that size, the throughput advantages of a GPU are largely neutralized — though Nokia argues that GPU parallelism enables classes of algorithms (including RKHS methods) that ASICs cannot run at all, regardless of model size. What makes Samsung's position different from Ericsson's in the 6G chip debate? Both Samsung and Ericsson reject Nokia's argument that GPUs are a prerequisite for 6G. But their architectures are different. Ericsson remains the only non-Chinese RAN vendor investing heavily in custom application-specific integrated circuits, which it argues deliver superior energy efficiency for the specific real-time constraints of radio processing. Samsung has invested primarily in virtual RAN products built on commercial CPUs — Intel Xeon 6 and AMD EPYC — while leaving the door open to selective GPU augmentation for workloads that are highly parallel and can exploit GPU throughput within the TTI timing constraint. Samsung has simultaneously validated its vRAN software on Intel, AMD, and Nvidia platforms, a hardware-agnostic approach that avoids the CUDA lock-in risk Nokia has accepted. The practical tradeoff is that Samsung has made no headline spectral efficiency claim comparable to Nokia's 100% target, and its near-term performance gains are more modest. What happens if Nokia's GPU approach fails to deliver in live networks? Nokia's commercial AI-RAN platform is built on software optimized for Nvidia's CUDA architecture. If the platform's real-world spectral efficiency gains in commercial dense urban networks prove significantly below the 100%+ target Nokia has published for 2028, the lock-in cost becomes acute: CUDA software written and optimized for Nvidia GPUs cannot be ported to AMD, Qualcomm, or Intel silicon without substantial rewrites. Nokia's own Marvell-era custom silicon will have been discontinued by that point, leaving no fallback. The Marvell chips in Nokia's newest products are the last it has said it will ever develop. An operator that has built AI-RAN workflows around Nokia's CUDA platform would face non-trivial migration costs if performance disappoints. T-Mobile's expected 2027 commercial deployment will provide the earliest real-world benchmark against which Nokia's roadmap claims can be measured. ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.