SAN JOSE, Calif. — After failing to gain much traction in servers to date, Arm renewed its foray into the cloud with a roadmap of cores designed for infrastructure gear. The news marks a re-commitment for the company that dominates smartphone and embedded cores but has yet to stake out a position in the cloud or the emerging market for AI accelerators.

Starting next year, Arm and partners will roll out cores, IP, and SoCs optimized for servers, networks, and storage systems. They will come at an annual cadence in step with new process nodes. Arm promised 30% annual performance gains for the cores bearing a Neoverse brand announced at its Arm Tech Con here.

The initiative comes at a time when smartphone growth has slowed to single digits and Arm has virtually saturated the microcontroller market. Meanwhile, the industry is buzzing with talk about so-called edge-cloud networks that carriers and cloud-service providers may build over the next few years as a new layer of the internet.

Arm committed to roll out Ares, its first Neoverse-branded products in 2019, optimized for a 7-nm node. It will follow up in 2020 and 2021 with Zeus and Poseidon cores targeting 7-plus-nm and 5-nm nodes, respectively.

Versions of the products will support SoCs packing up to 256 cores and up to eight DDR memory channels, eight HBM stacks, 128-Mbit caches, and fast interfaces such as CCIX. Arm did not detail any specific products ahead of keynotes on the initiative at its event.

“Neoverse is different from what we have been developing for client devices [with the Cortex brand],” said Drew Henry, who joined Arm a year ago to lead the initiative. “We’re targeting high-performance secure products that will scale from hyperscale data centers to base stations and gateways at the network’s edge.”

The Neoverse plan is in lock step with the roadmap of TSMC, increasingly seen as the industry’s process tech leader. However, Arm said that Neoverse will also support Samsung and UMC nodes.

So far, the Neoverse roadmap lacks any specific product details. (Images: Arm)
So far, the Neoverse roadmap lacks any specific product details. (Images: Arm)

Long term, the brand is expected to include a variety of CPU, GPU, and other cores, as well as SoCs from customers. Arm will also offer architecture licensees to customers who want to customize Neoverse cores.

Across all of the products, “software compatibility is an aggressive focus for us,” said Brian Jeff, a Neoverse product manager.

Analysts gave the move a thumbs up. “They’ve been trying to compete in the enterprise with the same cores that they are using in smartphones, and there are a lot of differences in performance and reliability, so having a separate set of cores should help them,” said Linley Gwennap of the Linley Group.

In general, the cores should focus on instruction level parallelism with out-of-order pipelines and ECC and extensive parity checks. They could also adopt scalable vector extensions like Intel’s AVX, said Gwennap.

Neoverse and Cortex will be like Intel’s Xeon and Core brands, according to Nathan Brookwood of Insight64. “Arm’s thrust into general-purpose servers has been difficult for them, but the server space is a lot more than general-purpose systems,” he said, citing smart city and connected car sectors.

Arm claims that its processors already dominate infrastructure markets with a 27% market share in unit terms. Analysts note that Intel leads in processor revenues for the heavily fragmented sector that also uses Mips, Power, and many legacy architectures.

Much of Arm’s success to date has been in data center appliances and storage controllers. It has put significant energy behind server-class designs that, to date, have gained little traction. Henry promised news of progress on that front at Arm Tech Con this week.

The sector is a strategic target for Arm given that SoCs here can sell for $100 to $1,000 compared to under $50 for mobile SoCs, according to Gwennap. Server processors are the cream of the market, but Arm server SoCs have “gone into another period of retrenchment,” he said.

“There was a lot of excitement a year or so ago, then Qualcomm pulled the plug on its Centriq and AMCC’s XGene was sold to a startup and Cavium’s ThunderX was involved in the merger with Marvell.”

Meanwhile, AMD has rolled out a much more competitive line of x86 processors based on its Zen core. “People were excited about Arm when AMD was falling flat, but now that AMD is doing much better, the data center users have a second source and don’t need Arm as much,” said Gwennap.

Qualcomm’s decision to cancel the Centriq initiative was “a very sad story,” said Brookwood.

“Qualcomm got caught in the crossfire of the China trade war [preventing its merger with NXP] and its patent dispute with Apple, so it had to find a way to take a billion dollars out of its expenses and apparently didn’t even have time to shop the Centriq group around properly, so a quality design may die on the vine.”

The result is that “server farms remain an x86 world, but I would not count Arm out — it is still well-positioned and has the relationships in place,” he added.

Arm targets up to server-class processors with Neoverse.
Arm targets up to server-class processors with Neoverse.

Arm hopes to leverage plans for edge-cloud networks that could be a disruptive opportunity.

Cloud and telecom providers are expected to build out the many small data centers to improve the quality of their services. AT&T described plans that range from small closets of a few servers next to a base station to medium-sized central offices. Startups are already forming plans to build edge-cloud networks to host services for providers that don’t want to own the gear themselves.

The trend amounts to “a redistribution of how [internet] processing is being done,” said Henry of Arm.

“There will be systems optimized around local decision-making with high I/O performance and built-in accelerators. There will be some redrawing of silicon boundaries with more I/O in the SoC to fit in the thermal profiles of edge networks.”

Gwennap expressed some skepticism about the trend, especially as carriers position it as a way to deliver new low-latency services with 5G. “The latency issue is overblown — latency at 100 versus 1,000 miles is not that different. You don’t want to serve Europe from San Francisco, but you don’t need a server in every town.”

At its event, Arm is likely to position edge-cloud networks as a new layer of the internet, part of its corporate vision of a fifth generation of computing along with the cloud and the internet of things.

Beyond the infrastructure debates, Arm has two other issues that it needs to address at its annual Silicon Valley gathering — AI and RISC-V.

“RISC-V is the hot thing now with startups, such as Esperanto talking about getting it into the data center,” said Gwennap. “People have gone beyond Arm as the alternative in servers and elsewhere.”

An early marketing effort to puncture holes in the RISC-V story backfired for Arm. More recently, it opted for the carrot, offering some low-end cores for free when used in FPGAs.

In AI, Arm has described a well-designed accelerator core. However, it has yet to appear in shipping SoCs, and Arm has yet to articulate its roadmap or its performance relative to cores from Cadence, Synopsys, and others.

“There’s a lot of competition for AI IP, and they need to show more details and some momentum,” said Gwennap.

— Rick Merritt, Silicon Valley Bureau Chief, EE Times