Qualcomm pumps up the baseband, betting on chips plus code

Even some laypeople know Qualcomm as one of the powers that be in mobile phones. For its most recent full fiscal year, the San Diegan business pocketed $44.2 billion in revenues, mainly by selling device chips and licenses. But it's a different matter on the mobile networks side. In the basestations that beam signals to your smartphone, Qualcomm is practically nowhere.

All that could change in the next few years if Qualcomm's plans work out. Telcos have tended to source ready-to-use network equipment for any given site from a single vendor. Today, that market is dominated by Ericsson, Huawei and Nokia. But technology executives at some of the biggest telcos want the freedom to buy and assemble parts from multiple suppliers. For Qualcomm, it is an opportunity to sell radio access network (RAN) chips based on its mobile expertise.

A lot of its attention is now focused on the chips installed in distributed units (DUs), the server boxes responsible for much of the RAN computing. In a traditional network designed by Ericsson or Huawei, these chips and the software they host are developed in-house. In the sort of disaggregated RAN that some operators prefer, Qualcomm could feasibly take over both the hardware and software roles while letting someone else handle other functions.

For now, Qualcomm is restricting itself to what industry parlance calls Layer 1 – the bottom-most slice of the software stack, and the one that demands the most processing power. This would leave other chipmakers to look after higher-layer software, housed mainly in server boxes known as central units (CUs). Intel, with its x86-based processors, currently leads the 5G field there in both traditionally built networks and newer "virtualized" ones, according to analyst data.

Partly because it seeks accommodation with x86-based computer equipment, Qualcomm has been installing its silicon on separate cards. These can be slotted into any server provided it offers support for PCIe, an industry standard. A technique known as inline acceleration would offload all the Layer 1 processing from the main server chip – whether based on x86 or Arm, a rival architecture – to the more customized Qualcomm silicon on that accelerator card.

Hitting the accelerator

It's a technique derided by Intel as costly, inefficient and incompatible with virtualization. Instead, Intel wants operators to rely more heavily on its main chips. Because these general-purpose processors are not ideal for the most computationally demanding Layer 1 software, Intel has developed its own customized silicon as well. But this handles only a subset of Layer 1, using a technique called lookaside acceleration. Intel previously offered it on PCIe cards, much as Qualcomm does with inline. Its preferred approach now is to put the lookaside accelerator on the same die as the main processor.

Qualcomm hits back at the Intel criticism of PCIe cards. "They didn't make this argument six months ago when they had a PCIe card," said Gerardo Giaretta, Qualcomm's general manager of 5G RAN infrastructure. With an inline card, he points out, central processing unit (CPU) cores can be switched off for energy and cost savings, because they are not required for any Layer 1 software.

"If you have a lookaside solution, at some point in time the number of CPU cores is so high that you need to put in two servers," he said. Using inline, an operator could simply introduce an additional card – the same approach the industry has followed with traditional, purpose-built networks.

Intel's other point of attack is about virtualization. Qualcomm, like other chipmakers building PCIe cards, uses Arm blueprints to make its customized silicon. The same cloud tools and software cannot be used to manage these dedicated functions and the broader x86-based deployment, Intel has previously argued. Instead of having a virtual or cloud RAN, an operator is back in the old world of physical appliances, according to this rationale.

While Giaretta concedes there would be "work to do" in this area the first time round, he vehemently rejects the view that Arm-based inline accelerators cannot form part of a virtual RAN. "You have this concept of physical and virtual functions, but the physical functions can be updated and managed by the virtualization layer and so that argument has short legs."

Qualcomm's five-year share price ($)(Source: Google Finance)

Still, the industry is currently divided over the issue of lookaside versus inline acceleration. The purveyors of inline insist theirs is the more energy-efficient solution, and the feedback Light Reading has received so far points to operator support for this boast. Yet Ericsson, the world's biggest supplier of 5G network equipment outside China, has built its entire cloud RAN strategy around use of Intel and lookaside.

"If you are an operator with Ericsson in the network and Ericsson is telling you it is only going to use lookaside, then what do you do?" said Giaretta. "Unless you are really religious, you are not going to change your vendor just to move from lookaside to inline."

Unlike most of its rivals, Qualcomm has come out with some clear numbers to support its claims about energy efficiency, although Giaretta says the dearth of other data makes strict comparisons difficult. During a demonstration involving an HPE server and a 64-antenna, 100MHz radio unit, it recorded total energy consumption of less than 18 watts when four mobile devices were simultaneously connected. Throughput of 24 Gbit/s is possible with peak power consumption of less than 40 watts, it said in a blog published shortly after this year's Mobile World Congress.

Special sauce

Qualcomm's other big differentiator with certain other vendors of inline acceleration is a major role in software. Marvell, the chipmaker to which it is most often compared in the 5G infrastructure market, contributes only the hardware. Nokia, its biggest 5G network customer, combines this with its own Layer 1 software, as others would have to if they chose Marvell's technology. But Qualcomm supplies both.

This seems to hold appeal for newer entrants that would otherwise have to write Layer 1 code compatible with Qualcomm's chips. Companies such as Mavenir and Rakuten-owned Altiostar (now branded SymRAN) began virtual RAN life as clients of Intel, using its FlexRAN reference design to produce their Layer 1 software. Companies cannot simply port this to another silicon platform. "The more you try to optimize, the more the Layer 1 tends to be tied to the hardware," said Giaretta.

Accordingly, Qualcomm has formed partnerships with Mavenir and Rakuten where it is providing both accelerator cards and Layer 1 software. The two other big clients cited by Giaretta are NEC, a Japanese equipment vendor, and Viettel, a Vietnamese operator keen – like Rakuten – to develop its own network technology.

The question is whether Qualcomm's software activities are a turnoff for established RAN vendors that view their own Layer 1 algorithms as a competitive differentiator. Qualcomm has opened its Layer 1 code so that it can be tweaked or even replaced with alternative software, according to Giaretta. But someone like Ericsson might not appreciate a third-party comparison of its software running on Qualcomm's chips against an all-Qualcomm set of products.

What's good for Qualcomm is that it has very little to lose in this sector and something to gain if new entrants are able to capture market share. The mobile infrastructure opportunity is a relatively small one for Qualcomm: Sales of all RAN equipment last year were roughly equivalent to Qualcomm's revenues. But the device market has been shrinking, and Qualcomm's sales fell by $2.5 billion, to $8.5 billion, for its recent, June-ending quarter. Anything extra is a plus.

Related posts:

— Iain Morris, International Editor, Light Reading