Why Microsoft Says ARM Chips Can Replace Half of Its Data Center Muscle
There are technological advantages, but it’s mostly about creating more competition for Intel.
![Microsoft's Open Compute-inspired servers on display at the Open Compute Summit in Santa Clara in March 2017 Microsoft's Open Compute-inspired servers on display at the Open Compute Summit in Santa Clara in March 2017](https://eu-images.contentstack.com/v3/assets/blt8eb3cdfc1fce5194/blt2d65317d78cc5116/66210ce31523d5e5d8b1ecb1/microsoft-ocp-servers-ocs17-e1489433720994.jpg?width=1280&auto=webp&quality=95&format=jpg&disable=upscale)
Recent years’ relative quiet on the ARM server front, the scarcity of large-scale IBM Power deployment news, and the consistent growth reported by Intel’s Data Center Group may in totality create an impression that the data center chip wars have subsided, with Intel enjoying a massive – and secure – lead. A few announcements that came out of last week’s Open Compute Summit in Santa Clara and the Google Cloud Next conference in San Francisco however showed that while Intel’s lead may be massive, it’s under bigger threat than may have appeared.
The biggest news out of the Open Compute event was that Microsoft had been working with ARM chip makers Cavium and Qualcomm on ARM server designs to run many of its cloud services (Qualcomm’s processors are in a big but reportedly waning share of the world’s smartphones). This was not an announcement like the ones we’ve heard in the past from Facebook and Google –non-committal revelations that they had been testing ARM servers here and there, and that they were really always testing everything on the market, just in case.
Microsoft thinks there’s real potential for ARM servers to eventually provide more than half of its cloud data center capacity. Coming from the world’s second-largest cloud provider, that kind of announcement should give Intel a lot to think about. As more and more corporate applications are headed for the cloud, the number of servers traditional hardware vendors sell to enterprises is on a gradual decline, while cloud providers are buying more and more processors to support those migrating workloads. The prospect of ARM servers working out the way Microsoft is picturing they may for its Azure cloud threatens Intel’s biggest source of revenue growth.
Learn about Project Olympus, Microsoft’s pioneering effort to design data center hardware the same way open source software is developed, at Data Center World, which is taking place next month in Los Angeles. Kushagra Vaid, Microsoft’s general manager for Azure Cloud Hardware Infrastructure, will be giving a keynote titled Open Source Hardware Development at Cloud Speed.
The same week, speaking on stage at Google’s big cloud event, Raejeanne Skillern, leader of Intel’s Cloud Service Provider Group, confirmed what Google had announced earlier, that Intel had sold to Google its latest-generation Skylake server chips before it would let any other company have them, giving Google’s cloud platform a temporary performance advantage over its competitors. Timing of both announcements may be a coincidence, but one reason Microsoft has been working with ARM server vendors is to avoid putting all its eggs in one basket, especially if that basket is a supplier that’s not required to provide all its customers with a level playing field.
Read more: Google Expands Cloud Data Center Plans, Asserts Hardware, Connectivity Leadership
In order to avoid having to rewrite much of the software that powers its cloud services, Microsoft has ported Windows Server 2016 to ARM, Leendert van Doorn, distinguished engineer for Azure, announced from stage at the summit. The company envisions using ARM servers to power its cloud storage, Big Data, Machine Learning, search, index, and other workloads. “Those properties together actually represent over half of our data center capacity, so there’s quite a lot of potential for different kinds of servers there,” he said, adding that the workload that for now is safely in the x86 corner is running customers’ cloud VMs, also known as infrastructure-as-a-service.
“We work closely with our x86 partners too, so one of the key things here for us is choice,” van Doorn said in an interview with Data Center Knowledge. Besides Intel, those x86 partners also include AMD, which is staging a “comeback” to the data center market, leading with its upcoming high-performance Naples chip. Project Olympus, Microsoft’s effort to leverage the open source hardware design community of the Open Compute Project to create its next cloud server, includes motherboards for Intel, AMD chips, as well as the ARM variants by Cavium and Qualcomm.
The ARM server ecosystem has to a great extent benefited from the massive scale of the high-end smartphone market, van Doorn wrote in a blog post. The developer ecosystem that has grown around ARM-powered smartphones has “significantly helped Microsoft in porting its cloud software to ARM servers.”
Van Doorn cited throughput-oriented ARM designs as a key reason Microsoft is getting so heavily involved with the architecture. “Those are things with high-performance IO, high IPCs (instructions per cycle), lots of cores and threads, large numbers of them, and lots of interesting connectivity options, especially with some of the newer bus standards, which are very interesting from an accelerator perspective,” he said.
Simpler hardware compatibility is another factor. Because ARM chipsets and motherboards are built on open standards, there’s absolutely no difference between Cavium and Qualcomm versions of Windows Server 2016 for ARM, van Doorn said. A single generic ARM ACPI driver will enable the OS to discover and onboard peripherals and such instead of different drivers for different chipsets, as explained by The Register.
Another trend that’s playing out in Microsoft’s ARM and AMD announcements is the new opportunity to optimize cloud hardware for specific workloads. That opportunity is in the economies of scale the cloud has made possible. From van Doorn’s blog post:
“Due to the scale required for certain cloud services, i.e. the number of machines allocated to them, it becomes more economically feasible to optimize the hardware to the workload instead of the other way around, even if that means changing the Instruction Set Architecture.”
A bigger variety of chipmakers in theory makes that kind of optimization easier.
About the Author
You May Also Like