Certainly, the timing is right from a technology perspective, but Arm's success in this market depends on more than its customers’ ability to meet hyperscalers’ stringent demands...
A recent high-profile product launch has pushed the concept of using Arm-based CPUs in the data center back into the limelight. Ampere, a Silicon Valley startup that focuses exclusively on Arm-based CPUs for the data center, launched its Altra CPU. Built on up to 80 Arm Neoverse N1 cores, Altra consumes 210 W at peak workload.
Ampere designed the Altra specifically for hyperscalers that need CPUs for their cloud offerings (an earlier Ampere CPU, eMAG, is optimized for enterprise data-center applications, which have slightly different requirements). Ampere says the Altra delivers on power efficiency (performance per watt) as well as watts per core, allowing for denser racks and maximizing performance per rack.
“Ampere’s Altra server processor news is more than just a milestone for Arm in the data center; it is a turning point for the industry in terms of what is possible in data-center computing,” said Chris Bergey, senior vice president and general manager for Arm’s infrastructure business, in a blog post. “Today’s announcement demonstrates the power of building the right compute for the right applications. Ampere took the N1 platform and integrated its own innovations to design an SoC uniquely built for applications across hyperscale cloud and edge markets while bringing it to market quickly to address evolving compute requirements.”
Let’s not forget that the road to Arm in the data center has been peppered with potholes thus far. An early startup in the space, Calxeda, was working on a 120 quad-core Cortex-A9 design (480 cores total) but closed its doors at the end of 2013 amid criticism of its 32-bit architecture, which some said was unsuitable for servers. A partnership with Hewlett-Packard for enterprise servers failed to save the company.
Fast-forward to 2016. Broadcom had its own project in the works: a 64-bit multicore Arm architecture, codenamed Vulcan, for networking, storage, communications, big data, and security applications. But the effort was a casualty of Broadcom’s acquisition that year by Avago. Then, at the end of 2018, Qualcomm abruptly shut down its Centriq division, which produced its line of Arm-based server chips.
Among those false starts, however, there is one notable success story: Amazon. The hyperscaler’s 2015 acquisition of system-on-chip maker Annapurna Labs resulted in the Graviton chip — the most famous success story thus far for Arm in the data center. Given that Amazon is its own customer for this chip, it may not be the best indicator of the market for Arm in the data center. But the company has said that developing its own chips optimized its precarious price/performance balance. Specifically, according to Amazon, Graviton A1 instances reduced customer costs by 45% for scale-out workloads such as micro-services and web servers compared with general-purpose Intel Xeon EC2 instances.
Amazon renewed its commitment to Arm architectures with the launch of the Graviton 2 in December 2019. This chip boosted performance to 7× that of its predecessor, with 4× the number of cores. Significantly, Amazon put the Graviton 2 into its Elastic Compute Cloud (EC2) instances, which are intended for high-intensity workloads, such as high-performance computing and application servers; previously, those instances were based on Intel Xeon. The shift signaled Amazon’s belief in Arm’s vision of diverse CPU architectures (heterogeneous computing) for even the most intense workloads in the cloud.
Amazon’s vote of confidence with the launch of Graviton 2 in December, plus the Ampere launch this month, has given Arm’s cause a bit of momentum. So will 2020 be the year that Arm finally makes serious inroads into the data center? Certainly, the timing is right from a technology perspective; power efficiency demands have never been stronger, for example, and may dictate a different approach than scaling up existing CPUs. The need for heterogeneous compute to process workloads such as AI also supports Arm’s case; we are seeing specialized accelerators such as GPUs, FPGAs, and even some ASICs getting more traction.
Arm’s success in this market, however, depends on more than its customers’ ability to meet hyperscalers’ stringent demands for performance, power efficiency, and price. It also depends on whether those customers can grab any market share from Intel’s dominant Xeon platform. And Arm isn’t the only other option here: AMD and Nvidia are also in the fray. So while we may be seeing a few rays of hope, it’s not quite the light at the end of the tunnel for Arm’s infrastructure business — at least not yet.