Telco open RAN 'accelerator' wish list is from cloud cuckoo land

Operators want full hardware and software independence and the ability to migrate seamlessly from lookaside to inline. Good luck with that.

Iain Morris, International Editor

July 9, 2024

7 Min Read
Deutsche Telekom office
Deutsche Telekom, Europe's biggest telco, has thrown its weight behind open RAN. (Source: Deutsche Telekom)

Europe's big telcos are unhappy about the schism that has developed between Ericsson and Nokia in the open radio access network (RAN). Far from being open to all approaches, the Nordic vendors, which account for most RAN expenditure outside China, have gone tribal, picking technologies that are seemingly incompatible with each other.

The telcos' latest gripe is about the conflicting choices for a technology known as RAN hardware (or HW) acceleration. In their most recent list of technical priorities (available here), Deutsche Telekom, Orange, Telefónica, Telecom Italia (called TIM by its friends) and Vodafone make their displeasure known. Ideally, they would like the ability to make a "seamless transition from one option to the other," they say. One technical expert within the industry likened it to a "Christmas wish list." Another judged some demands unrealistic.

Acceleration relates to a potential redesign of telco networks, incorporating broader IT rather than telecom-specific technologies. Besides using open interfaces to combine vendors at the same mobile site, some telcos want a virtual or cloud RAN. Dispensing with customized technologies, designed solely for the RAN, these would migrate network functions to general-purpose processors and compute platforms. Telcos, in theory, would benefit from the scale economies of this bigger ecosystem. They could, say advocates, host RAN functions on the same infrastructure platforms deployed for other workloads.

So much for theory. Virtual RAN (or vRAN) has made limited progress, partly because the standard central processing units (CPUs) found in common, off-the-shelf servers are unable to cope efficiently with the unique demands of more advanced 5G networks. Acceleration works by offloading the most resource-hungry RAN software (or SW) to custom silicon. Even Intel, the CPU king, has acknowledged the need for it with vRAN Boost, an accelerator that forms part of its RAN portfolio.

The two tribes of acceleration

The dispute, then, is largely about how much of the RAN software should move from the CPU to the accelerator. An approach the industry calls "lookaside," associated with Intel and Ericsson, previously kept everything on the CPU except forward error correction (FEC), an especially voracious bit of Layer 1 code, the most demanding stratum of RAN software. Intel stumps up both FEC hardware and FEC code in this approach.

The "inline" alternative, championed by Nokia, shifts the whole of Layer 1 from the CPU (normally Intel's) to an accelerator provided by someone else. In Nokia's case, that silicon vendor is Marvell Technology, the same company that provides Layer 1 technology for Nokia's purpose-built 5G products.

But in a further sign the RAN is defying attempts to virtualize it, the boundary between lookaside and inline has grown more blurry. At FYUZ, an industry event held in Madrid last year, Joel Brand, a senior director of product marketing at Marvell, provocatively suggested Intel had abandoned lookaside and swung behind inline. Publicly, Intel continues to insist "the future of RAN is on standard compute." Yet in future products such as Granite Rapids and Diamond Rapids, it plans to run more Layer 1 software than just FEC on its accelerator. Some of the beamforming, an advanced 5G technology, will also leave the CPU.

Whether lookaside or inline, the accelerator is where the virtual RAN effectively ends. While it could still be managed by a cloud computing player, it is a product purpose-built for the RAN in which hardware and software look as inseparable as an old married couple.

The Layer 1 code written for Marvell's chips, accordingly, could not be used with Intel's, Nokia admits. Intel's FEC software, undoubtedly, would not survive a transplant to Marvell's silicon. In their push for "SW portability and integration across different HW," as they put it, telcos are likely to be frustrated. One technical source said this would be like getting the Android operating system to work on the iPhone.

Indeed, the real problem is not the clash between lookaside and inline. It is the incompatibility of one vendor's technology with another's, regardless of the acceleration techniques they have chosen. In Marvell's Layer 1 chip, much of the software has been designed for hardware blocks based on Marvell's own intellectual property. It could not be ported to an inline accelerator developed by Qualcomm.

Arm wrestling with x86

Even the CPUs are not based on an industry standard. It may appear otherwise because the market for server CPUs is dominated by x86, a chip architecture, and the ecosystem that has sprouted around this. Until a few years ago, there was essentially no alternative. Yet x86 is supported by only two big semiconductor companies – Intel and AMD. Arm, a rival architecture used in most smartphones, has been able to carve out only a few percentage points of market share and has made no significant headway in the virtual RAN.

In a white paper published in 2022, Intel championed the use of CPU cores for all the Layer 1 functions except those on its accelerator. With this approach, it argued, the entire RAN stack could be written in standard C/C++ programming languages. Software could be "easily ported to other CPUs." But this portability does not appear to be as straightforward as Intel makes out if the CPU is based on Arm.

If it were, then AWS would already be hosting RAN software on the Arm-based Graviton chip it first developed several years ago. "We are still working on it," said Ishwar Parulkar, the chief telecom technologist for AWS, when he met Light Reading last month. "The challenge is porting software to Arm. Today, it is all primarily Intel-based, and there is work to be done, and it is engineering work. It is not rocket science. It is porting it to a different architecture. But it's a stack that has been built and validated on x86, so there is some engineering work."

There is perhaps a greater chance of that "seamless transition" between different accelerators in digital signal processing. Marvell, Qualcomm and others use this technology in their accelerators for some of the RAN-specific algorithms to do with channel estimation and beamforming. And just about all the underlying intellectual property originates with two companies – CEVA and Tensilica (owned since 2013 by Cadence Design Systems).

"That part of the software can still be implemented in a pretty open way in an inline accelerator," said Gerardo Giaretta, the general manager of 5G infrastructure for Qualcomm, during a previous interview. "Our solution uses DSPs [digital signal processors] that are not Qualcomm-specific and have a toolchain that is very common to the others. Every vendor knows how to program those DSPs. We are not bringing a Qualcomm-specific DSP that makes the entire platform difficult to code on."

Nevertheless, there are far fewer people with expertise in this type of software programming than there are coders for C/C++, Intel points out in its white paper. The other problem is that chipmakers might not align neatly in their reliance on DSPs from CEVA and Tensilica. Marvell, for instance, runs only a small portion of Layer 1 on DSPs. The majority goes on its optimized hardware blocks.

More slicing and dicing

Does all this make the acceleration abstraction layer (AAL), as telcos call their initiative, an entirely futile endeavor? Not necessarily. According to a source active within the O-RAN Alliance, the group defining open RAN specifications, one objective is to standardize the interface between Layer 1 and the rest of the RAN software. This would allow an operator to combine software vendors more easily, using Qualcomm for Layer 1, say, with Ericsson for Layers 2 and 3. While disaggregation of this nature has already happened in the core network, RAN software is typically provided as a monolithic block.

There have been some moves to combine software vendors, however. In Japan, Rakuten Mobile, the youngest network operator, is working to pair Qualcomm's Layer 1 technology with its own SymRAN software for other RAN functions. The Qualcomm code runs on a Qualcomm hardware accelerator delivered on a PCIe card, which Rakuten can slot into an x86 server hosting SymRAN. Nvidia, meanwhile, is pitching its graphics processing units as RAN hardware accelerators. Aerial, the brand for its own Layer 1 code, could theoretically be coupled with another vendor's software for Layers 2 and 3, hosted on a CPU.

Overcoming the resistance of incumbents to this pick-and-mix approach is the familiar challenge. But even if this mission were to succeed, it would fall some distance short of the full hardware and software optionality telcos seek. Work on AAL, they say in their technical priorities document, "is still in early stages." They may be heading for disappointment.

Read more about:

Europe

About the Author(s)

Iain Morris

International Editor, Light Reading

Iain Morris joined Light Reading as News Editor at the start of 2015 -- and we mean, right at the start. His friends and family were still singing Auld Lang Syne as Iain started sourcing New Year's Eve UK mobile network congestion statistics. Prior to boosting Light Reading's UK-based editorial team numbers (he is based in London, south of the river), Iain was a successful freelance writer and editor who had been covering the telecoms sector for the past 15 years. His work has appeared in publications including The Economist (classy!) and The Observer, besides a variety of trade and business journals. He was previously the lead telecoms analyst for the Economist Intelligence Unit, and before that worked as a features editor at Telecommunications magazine. Iain started out in telecoms as an editor at consulting and market-research company Analysys (now Analysys Mason).

Subscribe and receive the latest news from the industry.
Join 62,000+ members. Yes it's completely free.

You May Also Like