Pressure rises for Intel as Microsoft, AMD, Nvidia, ARM move on data centers

Microsoft has been showing off its Project Olympus server platform at the Open Compute Summit, and giving Intel headaches in the process. On show were ARM-based datacenter servers using Qualcomm and Cavium systems, and AMD timed the unveil of its new Naples server chip accordingly. Shut out of the GPU-leaning machine-learning and AI markets, the screw is turning on Intel.

Microsoft’s Azure cloud platform is a huge customer for Intel chips, and to see it exploring alternative processing options so openly will have caused some consternation at Intel. While the ARM version of Windows Server is still an internal build, the Project Olympus’ Open Compute hardware is an open-source reference design that Microsoft is hoping will shake-up the market.

The two ARM servers on show were a Qualcomm system based on its Centriq 2400 CPU, with 48-cores at a 10nm die-size, and a Cavium motherboard housing its ThunderX2 ARMv8-A, a 54-core CPU. Currently, Microsoft says it is using the systems in its own data centers, running the aforementioned internal applications that it says suit the ARM systems best.

Microsoft says these internal cloud applications include ‘search and indexing, storage, databases, big data, and machine-learning,” which are all areas that will set alarm bells ringing at Intel. The latter two apps are major growth areas that Intel has been eying up for years, and while this certainly isn’t a sign that Intel is being dethroned, the move does show that competition is threatening both Intel’s core markets and its data center expansion plans.

Intel has been looking to diversify from its core PC and server markets for a fair few years now. Its attempt to enter the mobile market was a disaster, although it has now found itself providing LTE modems to Apple. Its IoT projects appear to be going well, even if their revenue remains a fraction of its total quarterly results, and Intel is also now turning its attention to drones, VR/AR, and automotive applications – to ensure that losses in the data center don’t damage its bottom line – which saw it post $7.5bn profits on its data center revenue of $17.2bn, last year.

As the demand for compute resources increases over time, as people and businesses consume and generate more data, Intel had always seemed pretty secure – able to depend on growing data center demand for its Xeon CPUs, as long as it can keep its x86-rival AMD at bay, which has just unveiled its new Naples server chips.

The emergence of machine-learning and AI-enabled applications has dented Intel’s growth curve expectations, as these applications currently favor GPU-heavy processing power. Instead of a rack of Xeons, these systems might only require two or four CPUs, used to power a suite of GPUs running in a PCI card. Nvidia has been leading the charge here, with its Tesla P100 card, as well as its new HGX-1 boxes that were unveiled in partnership with Microsoft, but AMD has also charged onto the scene with its new Instinct GPUs.

The GPU accelerator cards appear to be enjoying success in these new applications, and because they require far less Intel components, Intel has a pretty limited upside here. What’s more, these are applications that ARM might find itself competitive in, because the applications aren’t bound to the same kinds of legacy infrastructure and systems that x86 is entwined in, in the more common data center functions – functions that Intel still has good short-term potential in, but will likely face increasing competition from ARM as time goes on. Intel is also looking to FPGAs as a way to expand, announcing a deal with Alibaba Cloud (Aliyun) this week too.

But the amount of corporate inertia that prevents IT decision makers from simply moving to an ARM-based architecture is enormous in traditional data center workloads. It’s one thing to get a version of Windows Server that supports the ARM architecture, but there is a much wider software ecosystem to consider, as well as all the peripheral hardware that supports the servers themselves.

While it might simply come down the potential operational cost savings that an ARM architecture provides, it is far from simple to calculate the performance/cost per watt for a data center migration. As it stands, the cost of a complete overhaul mean that adoption is going to be slow – perhaps on just a rack-by-rack basis. It’s rather unlikely that data centers are going to start ditching x86 servers en masse, but if Intel is weakened, it’s going to be more like coastal erosion rather than pyroclastic flows.