Intel vs everyone in the data center
Everyone not named Intel wants more data center business, including AMD, the ARM ecosystem, and NVIDIA. But Intel has seen this coming from miles away, so I'm expecting less titanic shifts for the company and more gradual gains for would-be competitors. It's a great time to shop for new servers, regardless of what architecture you might favor.
Like an annual return of cicada, AMD is once again challenging Intel. Expect a lot of buzz and hype about how AMD will do "more" than the competition with its Naples processor. If past performance is indicative of future results, we'll see a slow fading of noise, followed by AMD going back into relative hibernation for another three to seven years before emerging again.
AMD's Naples silicon provides up to 32 cores of x86 processing and also includes support from Microsoft to incorporate the processor into Project Olympus—the company's latest model for open source hardware development. First hardware will be available in the second quarter of this year with volume building in the second half of the year, with pricing yet to be announced.
The AMD/Microsoft announcement was made at the Open Compute Project U.S. Summit last week, with Microsoft putting up enough resources for its Project Olympus to get a dedicated workshop track at the event. But Project Olympus (and ergo, Microsoft) is also supporting ARM through a partnership with Qualcomm that might be more significant if it gains traction.
Over the years, there's been a lot of talk about ARM -- yes, even me -- finding its way into data centers. Qualcomm is putting its 10 nanometer (nm) Centriq 2400 platform into OCP and demonstrated Windows Server running on the Centrica 2400 processor. Microsoft has flirted with ARM for low-end devices, such as the entry-level Surface tablet, but the Qualcomm announcement puts a 10nm, 48 core server processor along with supporting memory, network, and peripherals onto a single 1U server. It's serious hardware for the data center.
While stock servers are the bread and butter of the data center, the need for specialization and speed to support AI and other compute-intensive applications continues to grow. Microsoft, teaming up with NVIDIA, is Intel's third warning shot. The HGX-1 hyperscale GPU accelerator is a Project Olympus open source design to add more compute power into the data center. Depending on the configuration, the HGX-1 provides up to 100 times faster deep learning performance compared to "legacy" CPU-based systems, with estimates of one-fifth the AI training costs and one-tenth the cost for an AI inference run.
Intel sales reps to Microsoft might feel more than a bit uncomfortable right now since Microsoft has announced three major partnerships with non-Intel processor manufacturers. AMD represents a simple plug-and-play threat, with current code relatively portable off Intel servers, just optimized and tweaked to work best on a Naples server.
A team up of Qualcomm and Microsoft should be more concerning. Qualcomm is starting at 10nm and 48 ARM cores this year, putting it on the market at the same time and at the same circuit density as Intel's Cannon Lake architecture. Intel can no longer claim the high-ground on having the latest and greatest in Moore's Law since Qualcomm is effectively tied in the race to ship commercial 10nm products. And any market share growth of ARM means fewer Intel chips moving into the data center, along with a longer term play of ARM serving as a baseline processor for consumer and business mass market Windows devices ranging from smartphones to desktop solutions.
NVIDIA just continues to kill when it comes to specialized high-end applications. It would have been more shocking if it hadn't put in work with Microsoft's Project Olympus. Intel would like to find a way into the market and/or a good AI story to tell, but it's an uphill battle compared to the record NVIDIA has and continues to build.
As I noted last month, more competition for Intel means more options and the ability to push for lower pricing on Intel-based solutions. It also enables the use of non-Intel solutions if the price and applications are right for your data center.
Image by: blogs-images.forbes.com