For Radio Access Network (RAN) software developers, "testing" chips is a never-ending process. RAN chip connoisseurs are constantly trying out hardware that isn't part of their usual product line, but few truly delve into the details. It seems everyone is worried that tasting a new product will cause some kind of "gastrointestinal upset." 

In the development of 5G baseband chips, Ericsson has consistently used a hybrid approach, combining its own chips with Intel chips. Nokia has long favored Marvell Technology's products. Samsung uses both Intel and Marvell products. While chip companies like AMD and Ampere are happy to experiment, they have consistently failed to truly participate in the development of mainstream chips.

Large-scale Radio Access Network (RAN) software developers

Ericsson, Nokia, and Samsung, covering all major Radio Access Network (RAN) software developers, will become the primary choice for European and North American operators once the current geopolitical landscape settles and Chinese suppliers are completely excluded from the European and North American markets. It's relatively certain that Huawei and ZTE will not provide any 6G base stations in the EU, and they have already been banned from these markets in Canada and the US. This situation is unlikely to change in the short term.

Unfortunately for Western operators, the RAN chip market used by Ericsson, Nokia, and Samsung is less competitive. This effectively forces telecom operators to rely on a few chip manufacturers for critical network components. Meanwhile, Ericsson and Samsung appear to be increasingly reliant on RAN chips from Intel. Nokia, on the other hand, seems to be relying on Marvell. The Open RAN, introduced by telecom operators to introduce competition, has not brought about greater diversity in chip choices.

Part of the reason for this lack of choice is the recent decline in sales. According to research from Omdia, a sister company of Light Reading, the RAN product market revenue was approximately $45 billion in 2022. By 2024, this figure had plummeted to $35 billion. However, chip development remains a costly business. The rationale for investing specifically in developing chips for this shrinking RAN sector is increasingly difficult to justify. For many stakeholders, leveraging existing platforms in personal computers and data centers is a more sensible approach.

Unstable ASICs

Virtual RAN, which replaces custom chips with general-purpose processors, currently holds a small share of the RAN computing or baseband sub-segment. Omdia predicts this will be only 10% by 2023, but also expects it to more than double by 2028. With recent developments, RAN may more aggressively erode the custom chip market. Intel's latest Granite Rapids processor series demonstrates that telecom operators and suppliers have realized the performance and cost gap with dedicated chips has narrowed significantly.

This has a significant impact on Ericsson. The Swedish company still invests a portion of its annual R&D spending (approximately $5.7 billion last year) in application-specific integrated circuits (ASICs) for Layer 1 (the most critical part of the baseband). In other areas, it relies on Intel. If RAN gains a larger share in a slow-growing or stagnant market, the return on Ericsson's equivalent investment in ASIC R&D will be affected.

ASIC chip expenditure

Ericsson executives have not publicly stated that the company plans to cut spending on ASIC chips, but their wording has subtly shifted. Per Narvinger, head of Ericsson's Mobile Networks business group, said regarding trials of different chips, "We are experimenting on multiple platforms. But more importantly, we have actually achieved hardware and software separation." The ultimate goal is to have a single radio access network (RAN) software that can be deployed on multiple hardware platforms.

However, ASIC chips cannot achieve this because they tightly integrate hardware and software. In an interview with Light Reading last month when Ericsson released its third-quarter earnings, Narvinger pointed out that general-purpose 5G solutions include AMD, Intel, and Nvidia. But Intel remains Ericsson's only commercial partner. Despite Ericsson's claims of enthusiasm for Open RAN, its current business revolves almost entirely around dedicated 5G chips.

In contrast, Samsung's exit from the custom RAN chip market appears remarkably swift. Undoubtedly, it is the largest mainstream supplier of virtual RAN products, while its dedicated 5G technology, developed in collaboration with Marvell, has received little attention. Samsung built a completely virtual RAN for Verizon in the US and is about to adopt the same solution for Vodafone in parts of Europe. Canadian telecom operator Telus sources both virtual and dedicated 5G products from Samsung. However, Bernard Biro, Vice President of Wireless Strategy at Telus, stated that the performance of virtual RAN has surpassed that of traditional RAN, and the cost has also been significantly reduced. Similar to Ericsson, Samsung's processors are all sourced from Intel.

Just as simple as AMD

For Ericsson and Samsung, the most natural alternative to Intel is AMD. AMD is the only major chipmaker using the x86 architecture, meaning software written for Intel chips requires almost no modification. Switching to processors from Arm licensees, due to their different instruction sets, is far more difficult.

AMD's financial situation is also far better than Intel's. In the recently concluded third quarter, AMD's sales grew 36% year-over-year to approximately $9.2 billion, while Intel's sales grew by only 3% to $13.7 billion. Last year's massive losses forced Intel to lay off thousands of employees and severely dragged down its stock price—despite a recent $5 billion investment from Nvidia, Intel's stock price has still fallen by a third since the end of 2023. During the same period, AMD's stock price rose 78%.

Ericsson's main problem is likely the hardware accelerators provided by AMD to support forward error correction (FEC), a resource-intensive Layer 1 task. Granite Rapids and earlier Intel platforms integrated the FEC accelerator into the main processor, while AMD's accelerator resided on a separate card. Ericsson had previously stated its preference for integrated accelerators over discrete cards, criticizing discrete cards for increasing costs.

However, Samsung has been experimenting with a Virtual Radio Access Network (VRAN) software suite that, when deployed on AMD processors, requires no hardware accelerators. These processors typically have more "cores" (the fundamental building blocks of a processor), leading Samsung to believe they can achieve purely software-based front-end error correction (FEC). A commercial offer may be imminent.

Marvell

The emergence of Granite Rapids isn't the only reason why general-purpose processors will play a more significant role in future radio access networks (RANs). For two years, Nvidia has been pushing its graphics processing units (GPUs)—chips aimed at mainstream gaming and artificial intelligence (AI)—to also support RAN software. After two years of relative silence from kit vendors, Nvidia last month acquired a 3% stake in Nokia for $1 billion. The condition was clearly that Nokia would use the funds to design 5G and 6G products that could run on Nvidia GPUs.

Currently, telecom operators generally resist Nokia's GPUs, largely due to their perceived high cost and power consumption, making this move extremely risky. This also puts Marvell at risk. Nokia's ability to develop general-purpose software for both Nvidia and Marvell platforms is ultimately limited. Moreover, Nokia clearly cannot use the funds from Nvidia to specifically develop 5G Advanced and 6G software for Marvell chips. Furthermore, Nokia's latest earnings report shows that its mobile networks business group is facing financial pressure, with an operating loss of €64 million ($74 million) and sales of €5.3 billion ($6.1 billion) in the first nine months of this year.

Nokia and its customers may also rely on AMD for future chip choices. While the Finnish company has not been committed to building a Layer 1 stack for x86 processors, Nokia believes that the technology AMD developed for Nvidia GPUs is easier to port to other GPU platforms than ASICs. Moreover, for companies outside of China, AMD appears to be the only viable GPU alternative to Nvidia.

The era of globalization has created some thorny technological dependencies in several sectors, and 5G networks are one of them. Doubts remain about whether the Arm ecosystem is ready to support virtual RAN, and AMD's continued absence may leave RAN vendors and their customers with only Intel and Nvidia as general-purpose chip providers in an era of limited custom chip options. Telus's Bureau stated, "This will be a viable option in the near future," while also noting AMD's participation in RAN pilot projects. As for other manufacturers, it's hard to say at the moment.

Source: Content compiled from lightreading

View more at EASELINK