In the smartphone app development business, getting software to market means weighing whether a program will run as well for a farmer checking the weather in India as for a London banker hurtling to work in the Tube.

A French-American company, Genymobile, provides a cloud-based platform that emulates these different environments, letting big application developers test their Android apps for factors such as what network a phone is using, its location, or how fast it’s traveling. Using servers running on Oracle Cloud Infrastructure based on the Arm microprocessor design, Genymobile can pack up to 10 times as many test devices on each server with big performance gains, compared with servers using conventional chips, CEO Tim Danford estimates.

“It saves our customers time, and it saves them money,” he says. And since Arm chips also run the world’s phones and tablets, “I can trust that it’s bit-for-bit compatible.”

Arm, an alternative to the x86-compatible chips that power the vast majority of today’s servers, is working its way up from devices to data centers. The chip technology, whose power-saving design led to its use in billions of mobile devices and Apple’s newest desktop and notebook Macs, is gaining ground in cloud computing and high-performance computing centers, thanks to its low cost, flexibility, and speed. The architecture underpinned the smartphone rise of the 2000s; found its way into cars, fitness devices, and drones; and has been recently winning adherents for heavy-duty cloud computing jobs. Those include serving web pages to hundreds of millions of users; helping machine learning models with billions of data points make inferences about new data; and running high-performance computing software that simulates the effects of air flow, water, and biomolecular systems.

“I’ve seen interest across the board—in weather prediction, manufacturing, financial services, life sciences, and image recognition,” says Kevin Jorissen, a distinguished high-performance computing cloud architect at Oracle. “Anything that can push down costs 20% is a godsend, and Arm does this. Cloud providers have struggled for years with costs and still do. Running on Arm could make some of these services cheaper to operate.”

Britain’s Arm—the abbreviation stands for “advanced RISC machines”—doesn’t produce chips of its own. Instead, the 31-year-old company owned by Japan’s SoftBank (and target of a pending US$40 billion acquisition offer by NVIDIA) licenses its processor designs to semiconductor, computer, and device makers. This spring Arm introduced its first new architecture in a decade, which helps protect sensitive business data and code from the rest of a computer system, even while that data is being processed or transferred.

Oracle on May 25 became the latest cloud provider to introduce a cloud computing service based on Arm. Businesses that use Oracle Cloud Infrastructure to power their applications will be able to run servers based on chip designer Ampere Computing’s Altra CPU, which can pack up to 160 processing cores and 4 terabytes of memory in a machine (Oracle’s machines will include up to 1 TB per server). Oracle is touting the Arm service’s “penny-per-core-hour” pricing and Ampere-based machines’ ability to scale up predictably as users add more processing cores to their applications.

Oracle is also working to certify Oracle Database to run on Arm and has certified Oracle Linux, which would let businesses eke more performance out of their databases.

In high-performance computing, lowering the cost per computing cycle pushes the frontier of solvable problems. Business computing can also benefit from Arm’s lower costs.

Chipmakers and cloud providers are counting on programmers’ familiarity with the architecture from phones and other devices to spur a wave of development on data center software that’s mostly been written for x86-compatible chips. ”The companies leveraging Arm have been web and cloud native, not the Fortune 500,” says Patrick Moorhead, principal analyst at Moor Insights & Strategy. “We view this as Arm’s foray in the enterprise data center. Databases like big caches and a lot of cores, which this chip has. Oracle customers will likely be able to save money, at good performance.”

After Moore’s Law

In an era when the biennial doubling of semiconductor processing power the computer industry counted on for decades to fuel invention is waning, Arm is stepping into the breach. The chips house many single-threaded cores that execute tasks independently without having to wait for more code. That suits computing jobs whose components run in parallel to one another such as machine learning and computational fluid dynamics that increasingly run in cloud computing data centers. Such chips can offer better performance for the same price as server chips based on the x86 design. “Hardware diversification is where we’ll get the gains,” says Jorissen. “A cloud user can gain performance and cost efficiency by running each task on the processor that does it best.”

The rise of Arm in cloud computing shows how the computer industry is partially returning to an era when hardware platforms were paired with the applications they best supported, says Clay Magouyrk, Oracle’s executive vice president of cloud infrastructure engineering, in a recent blog post. “For the past decade or two, x86 has become completely dominant on the server, removing the need for most applications and libraries to work across multiple architectures,” he says. “As an industry we have to invest in the Arm development and operating environment, rediscovering our multiplatform skills of the past.” Oracle invests in Ampere and its CEO, Renée James, sits on Oracle’s board.

Arm’s price-performance edge has to do with the way its chips interact with the software they’re running. The x86 processors use a technique called multithreading, in which CPU cores juggle multiple tasks, pausing while they fetch instructions to run. The approach makes use of spare computing cycles, but can lead to variability in how long a task takes to complete. With Arm, each core handles a single thread, or set of scheduled instructions, and runs it until it’s done. The result can be a highly predictable processor with a very low number of errors. Cloud providers can pass on cost savings to customers.

 

“We view this as Arm’s foray in the enterprise data center. Databases like big caches and a lot of cores, which this chip has.”

Patrick Moorhead, Principal Analyst at Moor Insights & Strategy

Businesses still face hurdles in getting commonly used applications to take advantage of Arm. Many applications for managing financial data, human resources, or customer accounts weren’t written to run on more than about two dozen processing cores. “Very few are written for 160 cores in a machine,” says Sean Varley, senior solutions director at Ampere.

Enterprise software makers haven’t by and large rewritten their software to run on Arm, either. And Arm isn’t always the best choice for computing jobs that need the highest possible performance per core, such as applications whose licensing fees depend on the number of cores users run.

Arm’s piece of the pie

While still powering only a small portion of the world’s servers, Arm CPUs are gaining ground. Computers containing the chips are expected to account for 3.7% of server spending globally by 2024, up from less than 0.3% in 2019. According to market researcher IDC’s Worldwide Quarterly Server Tracker that’s an annual growth rate of 67%. The forecast doesn’t include supplemental accelerator processors that control storage or network devices. “The growth is being driven by the public cloud services and not enterprise data centers,” says Ashish Nadkarni, an analyst at IDC. “Web serving, silicon-independent apps that can be ported over and high-performance computing are where the cloud service providers want a piece of the pie.”

Meeting the needs of hyper-growth markets, such as artificial intelligence and cloud-based supercomputing, increasingly means turning to specialized hardware. Using Arm designs, cloud providers can cram more servers into data centers while staying within energy budgets. In the more than US$60 billion cloud computing infrastructure market, small gains in performance or power consumption can add up to big savings. Arm’s customizable blueprint also means cloud providers can tailor chips to their needs rather than buying off-the-shelf products that contain functions they don’t require.

The adoption curve is setting up a battle among established chipmakers and a raft of well-funded startups developing Arm-based chips and computers to supply the silicon for tomorrow’s applications.

NVIDIA plans to release an Arm CPU called Grace aimed at high-performance computing and AI by 2023; the Swiss National Supercomputing Centre and the United States Energy Department’s Los Alamos National Laboratory will be its first customers. The processor promises speed boosts of 10 times those of today’s CPU-GPU combinations for understanding language, running AI jobs on supercomputers, or making neural network recommendations. Grace will be able to train future large language models with up to 1 trillion variables in three days, versus a projected month today, NVIDIA has said.

Fugaku, the world’s fastest supercomputer, housed at the RIKEN Center for Computational Science in Kobe, Japan, runs more than 7.6 million Arm cores. And Apple’s Arm-based M1 chips, which it’s using in its newest desktops, laptops, and tablets, can get double the performance of the latest PC chip, and consume one quarter of the power at the same peak performance. That could lure even more developers to Arm.

Intel, too, is addressing the industry changes under new CEO Pat Gelsinger. The company is building fabrication facilities to manufacture chip designs of other companies, and Gelsinger has said customers can combine Intel’s x86 designs with those of Arm and the open source instruction set RISC-V.

The moves come amid a chip supply shortage that’s affected the automotive, computer, and other industries. President Biden’s broad-reaching American infrastructure improvement plan includes US$50 billion for production incentives and research investments in the US semiconductor industry.

Entrepreneurs eye Arm

Arm’s momentum is attracting entrepreneurs and venture investors. Silicon Valley startup Groq, which raised US$300 million in April, has designed a chip aimed at AI decision-making capable of a quadrillion operations per second. Graphcore, whose specialized AI chips emphasize memory capacity, in December raised US$222 million, on top of a US$150 million round earlier in the year. Canada’s Tenstorrent this year hired a former AMD, Intel, and Tesla executive and chip designer as its president and chief technology officer, and announced a CA$200 million funding round May 20. AI computer system developer SambaNova Systems raised US$676 million in April.

In March, Qualcomm announced that it completed its US$1.4 billion acquisition of NUVIA, an Arm CPU startup. Qualcomm and Microsoft said on May 24 they built a compact Windows 10 PC for developers running Arm.

Biggest hurdle When it comes to shifting big enterprise workloads to the cloud, the database is paramount. That’s why Oracle certified its Oracle Instant Client for writing and running applications that connect to Oracle Database and for Oracle Linux running on Arm. Further out, Oracle plans to certify Oracle Database server code running on Linux for Arm as well, according to Vice President of Database Product Management Jenny Tsai-Smith.

IDC analyst Nadkarni says the moves may provide businesses more reasons to add Arm to their mix. “If you’re talking about enterprise applications, you can’t have that conversation without talking about the Oracle Database,” he says “Your biggest hurdle to porting an enterprise workload is the complex database. It’s the start of a big journey.”

• Arm-based cloud computing is the next big thing: Introducing Arm on Oracle Cloud Infrastructure
• Learn more about Oracle’s next-generation cloud infrastructure
• Get the guide to Cloud Essentials (PDF)
• Get started with Oracle Cloud for free