AMD’s broader plans to develop chips based on ARM as well as x86--what it calls an “ambidextrous” strategy--is an ambitious attempt to steer clear of Intel and navigate a post-PC world. But the latest announcements leave lots of questions, not only about AMD’s plans for servers, but also about the near-term future of ARM-based micro-servers in general.
Earlier this year at Facebook’s Open Compute conference, AMD announced that it was about to start providing key customers with early samples of its first ARM-based server processor, the Opteron A1100 Series. The chip, code-named Seattle, is one of the industry’s first 64-bit ARM server processors (Applied Micro claims to be first with its X-Gene) and the first manufactured on an advanced 28nm process. The A1100 comes in versions with four or eight ARM Cortex-A57 CPU cores.
AMD said that the A1100 is designed for servers for Web hosting, big data and storage applications. At its press event yesterday, the company publicly demonstrated Seattle for the first time performing basic hosting tasks such as serving up dynamically-generated Web pages, a Wordpress liveblog and streaming video. However AMD did not provide any details on product plans, performance or power consumption.
The A1100 will be commercially available by the end of the year and, in response to a question, Dr. Lisa Su, a senior vice president, said it will appear in a SeaMicro dense server (the current SeaMicro SM15000 uses x86 Opterons, or Intel Xeon E3s or Atoms), but there was no indication that Facebook will be deploying thousands of these servers in its massive data centers. Instead the Opteron A1100 is primarily aimed at developers, in the hope that they will build the software ecosystem that will be necessary to challenge x86 servers down the road. AMD has already announced a developer kit that includes the A1100 processor, a developer board and software.
AMD also announced a new product family, code-named Project SkyBridge, which will be manufactured on a next-generation 20nm process and ship sometime in 2015. The SkyBridge family will include both an x86 APU, based on the same Puma+ CPU core in the recently-announced Beema and Mullins processors for ultrathin laptops and tablets, and an ARM version, based on an optimized version of the Cortex-A57. The twist is that the chips are “pin-compatible,” meaning they can both use the same circuit board. This simplifies things for device makers, but since they will generally be soldered onto the motherboard, the benefit to end-users is less clear.
SkyBridge is not the successor to Seattle and it isn’t meant for servers. Rather AMD is targeting embedded applications, in particular, industrial equipment and networking gear, which today use a hodgepodge of incompatible hardware and software including x86, ARM, MIPS and PowerPC. According to AMD’s roadmap, it plans to ship several 28nm embedded chips this year including an x86-based low-power APU (Steppe Eagle) and an A57-based SoC (Hierofalcon), so presumably SkyBridge will replace some of these products.
Interestingly AMD said that the ARM-based version of SkyBridge will run Android and appear in “some clients.” That sounds a lot like a tablet chip, though AMD has repeatedly said that it isn’t interested in chasing the low-end tablet market. AMD is also one of the key drivers behind the HSA (Heterogeneous System Architecture), which promises to make it easier for developers to build mobile apps that leverage the graphics processors, DSPs and other accelerators in SoCs to boost performance and reduce power.
AMD’s final announcement was that it had secured a so-called architectural license, enabling it to design a custom CPU core based on the ARM instruction set. By designing their own cores from the ground up, chipmakers can fine-tune the performance and power for specific applications, but it can take years, which is why many companies including Qualcomm have adopted a strategy of using ARM’s off-the-shelf Cortex cores to get to market quickly while working on their own implementations in the background. AMD’s first processors based on its custom “K12" core are slated to ship in 2016, and based on that timing they will presumably use a 14nm process with 3D or FinFET transistors, which would provide better performance and power. The company did not say which markets these first custom chips will target.
AMD is the only company pursuing both x86 and ARM on an equal footing, and given its limited resources, it seems to be doing a nice job pivoting from PCs to what it calls “high-growth” markets where it can leverage its graphics expertise. So far, that primarily consists of the chips that power the Sony PlayStation 4 and Xbox One, but AMD has its sights on other “semi-custom” and embedded markets such as gaming machines, set-top boxes, digital signage, medical imaging, industrial and networking. CEO Rory Read noted that when he joined AMD in August 2011 the company generated 95 percent of its sales from PCs, but by the end of 2015, it will get half its revenue from other areas.
The question is where does that leave its traditional PC and server markets?
Beema and Mullins compete with Intel’s Bay Trail Atom processors and the Kaveri mainstream APU, announced earlier this year, competes with the Core i3, but AMD has no real answer for the Core i5 or i7. At the high-end, AMD continues to deliver highly-competitive graphics solutions, including the Radeon R9 Series, but the high-end FX Series CPU hasn’t received a major overhaul in years and there is no replacement on the roadmap.
In servers, Seattle is for now a development platform and SkyBridge is for other markets, suggesting it will be a couple years before AMD makes a serious assault on the dense server market with a custom SoC. Even with the demise of Calxeda, there are plenty of other chip companies racing to develop ARM-based server SoCs including Applied Micro, Cavium, Marvell and Texas Instruments, but AMD’s plans indicate that it will take another one to two years to get the hardware and software in place, and only then will we get a real idea of the potential for ARM servers. In terms of x86, the low-power Opteron 3300 will be replaced by a 28nm quad-core processor, with a newer Steamroller core, code-named Berlin, but AMD’s mainstream Opteron 4300 and 6300 will stay on 32nm and remain largely unchanged.
It could be that AMD plans to make other announcements later this year regarding processors for mainstream PCs and servers. But it seems more likely that the company will move increasingly further away from its traditional markets toward embedded, dense micro-servers and low-power clients. That may be a sound strategy for AMD, but it leaves Intel with an open field. In an article in The Wall Street Journal, Read said that the “unhealthy duopoly” in x86 servers has left Intel free to extract higher prices. Based on yesterday’s announcements, that isn’t likely to change in the next couple of years.