Parallel Wireless is back, boasting a cloud RAN breakthrough

On Bristol harborside, not far from where slave trader Edward Colston's statue took an inelegant dive at the hands of antiracism protesters, is a small but unusual cluster of radio sites. Commercially used kit in the UK is typically provided by giant equipment vendors like Ericsson and Nokia. The trial waterfront equipment is powered by the software of Parallel Wireless. After a difficult year spent in relative obscurity, the small US company is back, and it claims to have done something unique.

Take any commercial radio access network (RAN) and the chances are that it uses custom chips and software coded specifically for their needs. But for several years, parts of the industry have been working to virtualize or cloudify the RAN. As elsewhere, this would mean producing code deployable on any general-purpose processor and being able to share IT resources with other workloads. But it is easier said than done.

Today, most cloud or virtual RAN software sits atop central processing units (CPUs) built by Intel. This is no surprise. Last year, more than 70% of the CPUs sold for data center servers came from the Santa Clara-based chipmaker, according to Counterpoint Research. With FlexRAN, a reference design dating back to 2013, Intel has remained the dominant force in the small cloud RAN market. What has not helped is that software written for Intel's CPUs cannot easily be redeployed on other hardware platforms.

Earlier this year, Ericsson did claim to have redeployed the code originally produced for Intel's chips on hardware built by AMD. But this was facilitated by AMD's adherence to x86, the underlying architecture first designed by Intel. Unfortunately, x86 is not an industry standard and no other big chipmakers use it. Ideally, Ericsson would be able to inject the same software into CPUs based on Arm, the main architectural alternative. But Ericsson has not yet managed it. In September, Matteo Fiorani, the head of Ericsson's distributed unit and infrastructure business, said the effort will require "tweaks and optimizations."

But Parallel Wireless now reckons it has a code base fully compatible with any general-purpose CPU, regardless of the architecture. At its UK office in Bristol, before a small group of government officials and employees from UK telco incumbent BT, it this month showed off the fruits of collaboration with Arm-licensing chipmakers including Ampere Computing, Marvell Technology, Nvidia and NXP. In the equivalent of a lab environment, products were seen to work.

They'll never take our freedom

It is a potentially significant move by Parallel Wireless, which reportedly laid off hundreds of employees last year in difficult market conditions. Before then, it had looked tied to Intel as a prominent client of FlexRAN, licensed to be used only with Intel's CPUs. "Even if you could run FlexRAN on another platform, it wouldn't work, because FlexRAN is heavily utilizing the AVX instruction set that is machine-specific and only on x86 architecture," said Nicolas Scheidecker, the head of UK research and development for Parallel Wireless. Ditching FlexRAN effectively frees it to work with any chipmaker it wants.

That meant taking a few steps back and designing software for the physical layer (or Layer 1), the most computationally demanding part of the RAN, from scratch. Much of the development has taken place at the Bristol office, which today employs about 40 people. Parallel Wireless has also benefited from UK government funding provided through the Department for Science, Innovation and Technology. Still, long-time RAN watchers will be curious to understand how the small company apparently did something that has eluded its bigger rivals.

Wary of disclosing company secrets, Scheidecker told Light Reading that Parallel Wireless acquired some of its expertise by observing other sectors. "Telecom is less mature in this space than other industries, and if you take the cloud industry and the video-processing industry and look at how Netflix and Google run their data centers, they don't run exclusively on x86," he said. "They have found solutions to deploy on different types of CPU."

Essentially, Parallel Wireless has tested some of those solutions, found the ones that are most suitable and then adapted them for what Scheidecker calls "the real-time embedded signal-processing problem." Despite last year's upheaval in the workplace, it has also continued to invest in talent. Scheidecker himself, who joined in mid-2019, previously worked for Intel and other chipmakers.

Cloudy vision

From his perspective, the latest breakthrough is important after too much industry backtracking and compromise on cloud RAN. This year, in his view, has been marked by a reversion to custom silicon in some parts of the ecosystem. Because Layer 1 is so hungry for resources, several vendors are pushing a technique called "inline" acceleration, which moves this software off the CPU and back onto custom chips. But these resources cannot be shared, and the software must be separately managed, according to critics. "It is regression," said Scheidecker. "It is a step back and I think a bit short-sighted."

The objection Parallel Wireless faces is that CPUs are less power efficient than customized chips. The company's reasonable response is partly that general-purpose processors are improving all the time. Older CPUs were "mostly scalar" and therefore unable to handle RAN workloads, according to Scheidecker. But the introduction of "vector" processing changes all that, he said. AVX512 ticks the box for x86, with SVE2 more recently doing the same for Arm.

Even Ericsson, which still invests heavily in custom silicon, has said general-purpose processors are now closing the performance gap. Arm, moreover, is prized for its energy efficiency. Aided by platform-agnostic software, competition to Intel from both Arm and RISC-V, a third architecture that Parallel Wireless claims to support, may spur further development.

Yet even Intel has begun using accelerators in its products to handle an especially demanding Layer 1 function called forward error correction. This would apparently not impede the use of Parallel Wireless for other functions provided companies stick with standardized approaches. An interface called BBDev, for instance, would allow Parallel Wireless to relinquish some functionality to custom silicon. "It is an abstraction layer between the software domain and a hardware accelerator," said Scheidecker.

For all the momentum on the vectoring side, he doubts a CPU-only approach would fly with massive MIMO, an advanced 5G technology. "The cost of running forward error correction on the CPU would probably not be the right solution," he conceded. But the key for Parallel Wireless, he emphasized, is to have maximum flexibility, allowing it to support acceleration where it is deemed essential. "What we are trying to do as much as possible is use general-purpose processors."

BT's involvement in all this is interesting. The operator has hardly been the biggest advocate of open and virtualized RAN infrastructure and has downplayed any need for an equipment refresh before the late 2020s. But Richard MacKenzie, a BT engineer present at the Parallel Wireless update, sounded relatively impressed with what he saw.

Few will expect to see the software used at scale in a commercial network anytime soon. But telcos are under pressure to prove they meant what they said. If open RAN just means rebranding giant deals with Ericsson, as AT&T did last week, then vendor plurality will be nothing more than a pipe dream.