Your browser is not supported. Please update it.

6 March 2020

Marvell gains from Nokia’s enforced U-turn on base station chips

Nokia’s challenges, like those of its main rivals, are complex and multi-faceted, and cannot be assigned to any one decision or setback. However, the failure of its strategy to develop a homegrown base station system-on-chip (SoC) has taken center stage in the analysis of the company’s sharp loss of market value over the past year, the departure of its CEO, and the reports that it is “exploring strategic options”.

Certainly, the chip decision has proved fateful. Initially, it seemed positive that Nokia was doubling down on its long history of developing chipsets for its infrastructure inhouse. There is a general trend, after a period when vendors were tending to save cost by using merchant chips for more of their products, to reverse that pattern and seek differentiation with a homegrown design. Samsung and Huawei both say their base station chips (and Huawei’s new ARM-based server processors) are key competitive advantages in terms of their product performance and their control of their supply chain.

Nokia was going the same way, developing the ReefShark architecture to underpin a range of network processors and SoC products, as well as high end silicon like the FP4 processor to power very high performance routers. The initial launches were well received, but in its core product area, the base station, the strategy has backfired.

Of course, base station chips have always been proprietary, inhouse affairs. All the vendors develop platforms based on a combination of dedicated ASICs, FPGAs (field programmable gate arrays) and other elements. Until recently, they mainly turned to merchant suppliers for device chips, for non-core products such as small cells, or for very specialized components.

But the migration towards disaggregated RANs, in which the network is split into central units (CUs), distributed units (DUs) and radio units (RUs), has changed the landscape, and it may shift again when many baseband functions are fully virtualized.

The CU, which will host non-latency sensitive functions in a central cloud, will run on common cloud infrastructure, with all the efficiencies that implies. Nokia and Ericsson are working with Intel x86-based platforms for this; Huawei has an ARM-based alternative; Marvell is also a major player.

The complication arises in the DU, which will run many of the most demanding, latency-sensitive and processor-intensive RAN tasks. General purpose processors (GPPs) such as Intel Xeon are not (yet) able to support these to the same levels of performance as dedicated chips, without the help of accelerators, based on FPGAs, graphical processor units (GPUs) and ASICs, which can offload the most demanding processes. This combination of chips makes the solution more expensive than the old proprietary solutions, something operators are seeking to address by making cost efficiencies elsewhere in the network – for instance, Deutsche Telekom’s work with Cohere Technologies’ algorithms to boost air interface efficiency (see Wireless Watch February 26 for a full discussion of GPP economics in base stations).

Nokia had designed its own souped-up solution based almost wholly on FPGAs, but it was very costly to develop and deploy, squeezing margins if the firm wanted to price the base stations in line with others based on old-style chipsets, or on merchant platforms. It also failed, at least in the first instances, to deliver the performance gains it had promised, as Sprint’s teething problems with its Nokia 5G sites highlighted rather publicly.

In the company’s third quarter earnings call last October, CEO Rajeev Suri acknowledged that 5G profit margins had been hit by the high cost of the homegrown base station chipset – gross margin for the Networks business fell from 34% in 2018 to 30.6% in 2019. Suri said the figures would improve once it completed its transition to cheaper SoC platforms, relying more on merchant partners. It will be too early for that transition to the new Reefshark SoC be visible in quarterly results – that will happen, if all goes well, in the first half of 2021 –  but improved margin figures in Q4 2019 did indicate that the inhouse chipset costs have been reduced already.

Suri said Nokia will provide quarterly updates on how many of its 5G shipments are powered by the new ReefShark in order to show its progress in reducing 5G product costs. “These new products made up approximately 10% of our 5G product shipments in the fourth quarter 2019, and we expect that percentage to increase progressively over the course of 2020, ending the year at more than 35%,” he said. “We expect to end 2021 at approximately 70%, and to essentially complete this transition in 2022.”

He added: “We have increased our SoC R&D capacity by about 60% over the past year, and we’ll continue to ramp up in 2020. And for 2020, we have three ReefShark SoCs under development.”

Nokia’s two main partners in developing the ReefShark SoC are Intel and Marvell. Marvell is the most established supplier of ARM-based processors in the telecoms network space, powering infrastructure from cloud servers for vRAN CUs, to specialized elements like base stations and small cells (some of this courtesy of its acquisition of Cavium two years ago).

Meanwhile, Intel dominates the server space with its x86-based Xeon platform but faces the challenge, outlined above, that this is not sufficient on its own for very advanced tasks such as RAN baseband processing or advanced AI/ML. Intel has been fleshing out its platform with acquisitions in the areas of FPGA (Altera), structured ASIC (eASIC) and is even developing a GPU (see separate item below). For the first time, it has a real shot at capturing significant share of the base station chip market, to complement its presence in servers and in switch-chips (the latter should be improved by another acquisition, of start-up Barefoot Networks).

Both these companies will bring contrasting expertise and approach to address Nokia’s challenge. Marvell said it will contribute customized chips based on ARM cores, which will be used in several elements of Nokia’s Airscale RAN platform, targeting reduction in size and power consumption, as well as improvements in capacity and overall performance.

The Marvell chipsets will gradually replace Nokia’s own FPGAs. This move, like the work with Intel, will reduce the ability to reprogram chips after they are deployed in the field (the great benefit of FPGA, especially when technologies are immature and subject to frequent changes).  But Nokia says it will reduce cost and speed up deployment time.

Although Nokia’s problems seem to have landed Intel with a shiny new customer for its newly launched base station SoC (along with Ericsson and ZTE), the world’s largest chipmaker has also been reported to have been part of the failure of Nokia’s homegrown approach. Nokia said, when the problems first came to light last autumn, that it had been let down by one of its suppliers, and several analysts pointed the finger at Intel.

“We believe the key problem area for Nokia has been the development of the 5G baseband, due to both Nokia’s internal development issues and Intel’s 10nm [10-nanometer process] delays,” said Barclays in a client note published in November. It also predicted that Marvell’s processors would gain an important spot in the new-look ReefShark SoC, writing: “We think these delays have led them to work with Marvell for their 5G baseband. We expect Nokia to begin the transition away from FPGA to Marvell’s baseband from 1Q20.”

The analysts also believe that Nokia has also been working more closely with Broadcom and Taiwanese ASIC design house Global Unichip, on certain radio and analog functions.