Data center server buyers will soon have a new choice for software-defined and high-performance data analysis workloads, as AMD plans a return to x86 server prominence.
Advanced Micro Devices (AMD) is betting on enterprise and hyperscale IT demand for a choice beyond Intel CPUs. By 2017-2018, AMD's roadmap is to gain customers in enterprise and customized large-scale -- such as cloud service provider -- data centers.
"Server users and OEMs want an alternative to Intel, if only to put competitive pressure on Intel for technology innovation," said Nathan Brookwood, research fellow at Insight 64 in Saratoga, Calif. AMD is the only company other than Intel and Via Technologies that can legally build x86 processors.
The share of data center servers based on AMD chips has shrunken considerably since the company once held about a quarter of that space. In 2006, AMD estimated its server processor market share in the high 20% range. By 2013, analyst firm Gartner put it in single digits.
Analysts say AMD took some wrong turns with data center offerings, including a foray into the server business with SeaMicro that it will shutter in 2015, and a socket-compatible ARM and x86 design -- Skybridge -- that lacked enough end-user support to get off the ground. The company also failed to develop a roadmap that was competitive with Intel, which contributed to the Wintel (Intel-powered, Windows OS servers) domination in today's enterprises."I don't see users abandoning Intel very quickly," said Charles King, principal analyst of Pund-IT in Hayward, Calif. "Intel has delivered on product innovation and has a clear roadmap ahead at least down to 10nm and probably below that."
AMD representatives also stressed that software and hardware partnerships will further develop its chip architectures. This is a call for openness that IBM also has pursued, by open-sourcing its Power chip architecture, King said.
AMD's architectural roadmap
Software-defined everything means that "the server has won," said Forrest Norrod, senior vice president and general manager of AMD's Enterprise, Embedded and Semi-Custom group. AMD will position its x86 and ARM architectures to replace proprietary hardware designs from storage and networking companies, to execute all the commands that were previously hard-wired into hardware.
Brookwood says AMD is in tune with the data center industry, tailoring its new architectures to support the software-defined era. In 2015, AMD will offer the Seattle ARM chip, Opteron A1100, for storage servers and a Hierofalcon ARM chip for embedded designs in networking. The ecosystem for A1100 includes software for Linux OSes, drivers, tool chains and software development kits.
AMD's Zen core microarchitecture for x86 servers will suit data center workloads, such as data analysis and visualization, as well as scalable apps, Norrod said. It targets high-performance computing with a secondary focus on energy efficiency. With simultaneous multi-threading and high-bandwidth memory to use a new cache system, the Zen architecture boasts a 40% improvement in instruction per clock over AMD's existing architecture, Excavator, and high native I/O capacity. The architecture releases in 2016 with Opteron on Zen. It is designed to solve the problems that virtualization creates on standard hardware, like memory starvation, memory overcommit and so on.
"AMD needs to be in the same performance class [as Intel Xeon processors] to handle software-defined data centers," Brookwood said. He added that it is too soon to tell if AMD can create an architecture to match what's on Intel's roadmap for 2016. AMD declined to share technical details around Zen, but called Intel "a ruthless and strong competitor in x86," a lesson AMD learned all too well in the late 2000s.
The architecture may not be AMD's biggest challenge in the mature x86 server space, according to King, but rather getting server buyers to care about their architecture.
"AMD does have some very interesting technologies, and can blend CPU and GPU capabilities in ways that are very intriguing for higher-end, higher-performance workloads," King said, "But they're almost starting all over again [in x86 servers]."
Skybridge proves pie-in-the-sky
AMD is delaying the Skybridge concept of a single motherboard for an ARM or x86 CPU for the foreseeable future. An AMD representative was hesitant to say it's canceled, but that AMD is waiting for the timing around ARM chips to work on it.
"It was an interesting idea, and AMD was the only one that could do it," said Brookwood. "But motherboard companies and server suppliers probably through that was a relatively small market and didn't buy in."
Hyperconverged data centers may have been ARM's driver, but Intel and IBM released impressive power efficiency and virtualization capabilities on their own silicon for this market, and ARM momentum seriously fell off in the past six to 12 months, King said.
Widespread ARM adoption in the data center isn't happening as quickly as once projected. If ARM form factors of today aren't traditional servers, and are still niches, then Skybridge isn't necessary in the immediate timeframe, according to AMD.
K12, AMD's ARM high-performance core, will debut in 2017. The logic for Zen x86 chips will carry over into the K12 ARM design, designed to make AMD competitive in server designs without a lot of roadmap juggling when ARM interest picks up, Brookwood said. The K12 uses the 2014 ARM v8 chip architecture license and will go into new servers and embedded systems designs. AMD's representatives admitted ARM hype cycles are nothing new to data centers, but expect 2017 to be a year of real adoption, which will trickle down from leading-edge IT companies into enterprises and small businesses as workloads are designed to use the alternative architecture.
"ARM will be important over time to the general purpose x86 server market," Norrod said. In the intermediate future, AMD will work to make x86 processors with the capacity to run software-defined data centers.
AMD also sees opportunity for graphic processing units (GPUs) and accelerated processing units (combining CPU and GPU) supporting workstation graphics from the data center and running high-performance computing applications. AMD's GPU developments include energy efficiency improvements, memory compression, graphics virtualization and high-bandwidth memory. High-bandwidth memory -- which stacks the memory on the chip -- uses less power and space on the motherboard than other GPU designs, according to company CTO Mark Papermaster.
These designs will rely on ARM Secure Processor and TrustZone for security, isolating and protecting applications and interfaces.
AMD also envisions its flexible chip design methodology -- which makes some processors more CPU heavy, adding GPU accelerators for other chips -- used in semi-custom hyperscale server designs tailored to the needs of the particular end user. For example, large cloud service providers may work with AMD to create chips that get the most processing per Watt efficiency based on the workloads it runs. Cloud suppliers have very different needs than the traditional enterprise, Brookwood said, and AMD's semi-custom business has a track record in this area.
Why micro servers failed
Lisa Su, president and CEO of AMD, said that the SeaMicro business unit didn't allow AMD to add differentiation for data center users and that the micro server segment hasn't taken off as quickly as expected.
But the main problem with AMD systems may have been AMD's primary data center customers -- systems OEMs. By offering full servers, AMD competed against its customer base. It was "very tenuous" from the start, Brookwood said, with inherent problems.
Another issue with micro servers was that the processors communicated via a proprietary fabric rather than a standards-based technology like Ethernet, he added. As that held back user interest in micro servers, the SeaMicro experiment became more of a problem.
"We can't cover every space in [the data center]," AMD's Norrod said, "Our strategy isn't centered on broad design points with a general purpose." Instead, he said, AMD will target workloads that Intel will not or cannot effectively support, with innovations and customizations at the chip level.
Meredith Courtemanche is the senior site editor for SearchDataCenter. Follow@DataCenterTT for news and tips on data center IT and facilities.
Compare these processor roadmaps with with Intel's E7 v3