RAN Silicon Rethink- Part II; vRAN and General-Purpose Compute
Overview:
The global Radio Access Network (RAN) market has experienced a significant decline, dropping by nearly $10 billion in annual product revenue between 2022 and 2024, from roughly $45 billion to about $35 billion by the end of last year (source: Omdia).
- As the IEEE Techblog previously reported, Nokia is gradually moving away from its long-held reliance on custom RAN baseband (BBU) silicon from Marvell [1.] as it pivots to use Nvidia’s GPUs, as part of the latter’s $1B investment in Nokia in October 2025.
Note 1. Nokia uses Marvell RAN silicon in its 5G ReefShark portfolio. The companies collaborate to develop custom OCTEON SoC (System-on-a-Chip) and Infrastructure Processors, which are used to boost 5G AirScale base station performance.
- Samsung has long partnered with Marvell Technology on purpose-built 5G baseband silicon. However, rising development costs and a contracting market for proprietary RAN hardware are reshaping that strategy. The economic case for new, custom RAN chipsets is becoming weaker as operators accelerate network virtualization.
- In sharp contrast, Ericsson continues to defend its investment in proprietary silicon architectures while maintaining a flexible approach for operators that prefer virtualized or cloud RAN implementations running on standard central processing units (CPUs). At present, those solutions rely exclusively on Intel processors, though Ericsson notes its software is being engineered with portability in mind to support future hardware diversity.
Samsung’s Silicon Strategy:
Among RAN equipment vendors accessible to operators across North America and much of Europe, Samsung now stands as the principal alternative to the two Nordic RAN equipment suppliers, following the exclusion of Huawei and ZTE from many Western markets.
The South Korean conglomerate has become the global frontrunner in virtualized RAN (vRAN) deployments. Whereas custom silicon once dominated RAN infrastructure design, Samsung’s strategy has notably inverted that paradigm: vRAN is now its mainstream offering, and purpose-built hardware has moved to the periphery.
By the close of last year, Samsung reported supporting approximately 53,000 vRAN sites worldwide — a significant share of which lies within Verizon’s U.S. footprint. The company also disclosed major European developments, including Vodafone’s planned rollout across Germany and other markets, which will rely entirely on vRAN technology. For Samsung, discussions of bespoke, purpose-built 5G infrastructure have become increasingly rare.
According to Alok Shah, Vice President of Network Strategy at Samsung Networks, this transition reflects both the rising cost of developing custom silicon and the performance enhancements achieved by general-purpose CPU platforms.
“We’re still selling our purpose-built BBUs to a number of customers, but I do believe that it’s a matter of time,” Shah told Light Reading during MWC Barcelona, when asked if Samsung envisions an eventual phaseout of its proprietary baseband hardware portfolio.
Virtualized RAN Gains Momentum:
Transitioning to virtualized RAN (vRAN) allows network equipment vendors to capitalize on the scale economies of commercial data-center silicon. Samsung has established commercial vRAN contracts with Verizon and Vodafone, reflecting growing operator confidence in software-defined architectures.
“Virtual RAN performance has reached parity,” Shah said. “I know not all of our competitors feel that way, but that’s certainly how we feel. And the cost of building that modem is pretty high, even for a company like Samsung that’s really good at semiconductors,” he added.
Intel’s Granite Rapids Xeon platform exemplifies this shift to vRAN. The processor’s increased core density enables operators to cut hardware footprints; in many configurations, a single server can now support workloads that previously required two. Several network operators have confirmed this performance improvement during field evaluations.
Samsung and Ericsson continue to explore additional CPU suppliers. AMD’s latest multicore x86 processors offer up to 84 cores, compared with 72 in Intel’s Granite Rapids. However, offloading Forward Error Correction (FEC)—one of the most compute-intensive RAN processes—remains a challenge. Intel’s vRAN Boost feature integrates a dedicated hardware accelerator for FEC, while AMD currently lacks a direct equivalent.
Samsung has also evaluated Arm-based platforms, which increasingly support efficient software migration from x86. Nvidia’s Grace CPU, built on Arm architecture, has emerged as a potential candidate, especially when paired with its GPUs for selective Layer 1 acceleration.
Samsung’s roadmap aligns with a gradual and selective introduction of GPU acceleration. The company demonstrated GPU-based beamforming optimization during MWC, illustrating how AI can refine radio energy targeting. However, Samsung executives maintain that the latest Intel CPUs also provide sufficient capacity to host AI inference workloads directly. “Granite Rapids has plenty of capacity to support AI algorithms on-platform,” noted Shah.
While Nokia is building a GPU-compatible Layer 1 to accelerate computationally intensive baseband functions—including FEC—Samsung’s approach appears incrementally narrower, focusing on targeted AI for RAN optimization rather than complete GPU offload. GPUs may ultimately support AI at the Edge applications—so-called AI and RAN—where telecom operators leverage deployed GPUs for latency-sensitive inference services.
The degree to which such applications will reside within RAN sites remains uncertain. Some operators suggest that edge inference may instead remain within core network clusters that can meet latency requirements more efficiently.
Samsung’s architecture already supports GPU integration through commercial off-the-shelf (COTS) servers from manufacturers such as HPE, Dell, and Supermicro—aligning with broader cloud-native RAN trends. “It’s an off-the-shelf card that can be integrated directly into standard servers,” said Shah.
For now, Intel remains Samsung’s primary compute partner for commercial vRAN products. “We haven’t had an instance where customers are pushing for a second platform—it’s primarily a matter of commercial interest,” Shah added. The direction is clear: Samsung, like other leading vendors, is prioritizing scalable, general-purpose compute over bespoke 5G silicon as vRAN deployment accelerates.
……………………………………………………………………………………………………………………………………………………………………
References:
https://www.lightreading.com/5g/samsung-eyes-death-of-purpose-built-5g-but-has-no-ai-ran-fears
RAN silicon rethink – from purpose built products & ASICs to general purpose processors or GPUs for vRAN & AI RAN
Ericsson goes with custom silicon (rather than Nvidia GPUs) for AI RAN
Marvell shrinking share of the RAN custom silicon market & acquisition of XConn Technologies for AI data center connectivity
Intel FlexRAN™ gets boost from AT&T; faces competition from Marvel, Qualcomm, and EdgeQ for Open RAN silicon
Analysis: Nokia and Marvell partnership to develop 5G RAN silicon technology + other Nokia moves
2 thoughts on “RAN Silicon Rethink- Part II; vRAN and General-Purpose Compute”
Leave a Reply
NVIDIA and Marvell Technology have formed a partnership designed to link Marvell into the NVIDIA AI factory and the wider AI-RAN ecosystem using the NVIDIA NVLink Fusion platform. This collaboration gives customers who build on NVIDIA architectures greater choice and flexibility when developing next-generation hardware environments.
Backing this hardware integration is a massive financial endorsement, as NVIDIA has invested $2 billion into Marvell. Financial analysts monitoring the telecoms sector often look for concrete monetary commitments to validate hardware alliances. This massive injection of capital guarantees that Marvell has the resources to scale its production of custom silicon and networking components.
Beyond the direct capital injection, the two technology giants plan to collaborate extensively on silicon photonics technology. Advancements in optical interconnects directly address the growing need for high-speed, low-latency networking architectures necessary to support heavy AI workloads at the edge.
The foundation of this joint effort relies on the NVIDIA NVLink Fusion platform, a rack-scale architecture that lets developers create semi-custom AI setups using the existing NVLink ecosystem.
Within this arrangement, Marvell takes responsibility for supplying custom XPUs alongside scale-up networking gear that maintains strict compatibility with NVLink Fusion. NVIDIA, meanwhile, will deliver the underlying support technologies, specifically the Vera CPU, ConnectX NICs, BlueField DPUs, Spectrum-X switches, the NVLink interconnect itself, and the rack-scale AI compute.
This specific division of hardware components creates a heterogeneous AI infrastructure for engineers building custom XPUs, ensuring complete compatibility with NVIDIA systems. Operators can then seamlessly integrate their edge deployments with NVIDIA’s GPU, LPU, networking, and storage platforms, tapping into the broader global supply chain ecosystem that NVIDIA maintains. The inclusion of BlueField DPUs, for instance, allows operators to offload heavy security and networking tasks from the main processors, freeing up valuable compute cycles for revenue-generating AI applications.
Matt Murphy, Chairman and CEO of Marvell, said: “Our expanded partnership with NVIDIA reflects the growing importance of high-speed connectivity, optical interconnect, and accelerated infrastructure in scaling AI.
“By connecting Marvell’s leadership in high-performance analog, optical DSP, silicon photonics, and custom silicon to NVIDIA’s expanding AI ecosystem through NVLink Fusion, we are enabling customers to build scalable, efficient AI infrastructure.”
The ability to deploy specialized compute nodes within the Radio Access Network changes the economic model of cellular sites. By partnering to turn the telecommunication network into a distributed AI infrastructure using the NVIDIA Aerial AI-RAN framework for 5G and 6G, operators can host enterprise workloads directly at the cell tower.
This edge capability establishes new revenue streams entirely disconnected from consumer smartphone subscriptions. Enterprises require low-latency processing for automated manufacturing, autonomous logistics, and real-time video analytics. Network operators can lease this edge compute capacity to enterprises, thereby driving up Average Revenue Per User (ARPU) and reducing enterprise churn.
Deploying private 5G solutions provides another direct application for this newly-announced infrastructure. The integration of Marvell’s custom silicon and NVIDIA’s rack-scale compute equips operators with the precise hardware combination necessary to secure highly lucrative private networking contracts. The data never leaves the factory floor, satisfying heavy data sovereignty and compliance regulations.
Jensen Huang, Founder and CEO of NVIDIA, commented: “The inference inflection has arrived. Token generation demand is surging, and the world is racing to build AI factories. Together with Marvell, we are enabling customers to leverage NVIDIA’s AI infrastructure ecosystem and scale to build specialized AI compute.”
Telecom operators cannot drop rack-scale AI compute into existing mobile switching centers without encountering friction. Integrating NVLink Fusion hardware requires extensive coordination with legacy Operations Support Systems (OSS) and Business Support Systems (BSS). Legacy BSS/OSS platforms were primarily designed to meter voice minutes and megabytes, not continuous API calls or dynamic edge compute provisioning. Overhauling these billing engines to handle AI-RAN monetization represents a massive and multi-year undertaking.
Furthermore, spectrum management becomes increasingly complex under this model. Running multi-tenant AI workloads concurrently with high-priority 5G baseband processing demands precise resource isolation. Directors often estimate that ensuring exact resource isolation can consume ten percent of edge computing overhead.
Operators must navigate multi-cloud environments, ensuring that containerized network functions interoperate smoothly with enterprise AI applications sharing the same physical silicon. While Marvell’s XPUs and NVIDIA’s Vera CPUs provide the processing variety needed for these distinct tasks, the software orchestration layer remains a daunting hurdle for IT directors to clear.
When operators expose these new edge capabilities through network APIs, they invite third-party developers to write applications integrated directly with the radio network. However, creating a developer-friendly API portal demands heavy investment in software infrastructure.
Legacy systems frequently lack the agility to authenticate, meter, and bill thousands of concurrent API requests originating from enterprise software. Upgrading these backend systems requires navigating a complex web of vendor lock-in and customized software deployments. IT directors face the difficult task of modernizing the billing infrastructure without causing service interruptions for the existing subscriber base.
https://www.telecomstechnews.com/news/nvidia-and-marvell-alliance-ai-ran-infrastructure/
Huawei appears wedded to ASICs. Ericsson also continues to invest in custom silicon and fears that following Nokia would lead to “lock-in.” Software coded for Nvidia’s GPUs cannot be deployed with other hardware platforms.
Instead, Ericsson is working through its virtual RAN strategy to ensure that nearly all the software written for Intel’s processors can also be used with CPUs from AMD and Arm licensees – including Nvidia – after only minimal changes.
https://www.lightreading.com/5g/nvidia-lines-up-ai-grid-as-orange-cto-echoes-the-ai-ran-doubts