A Sobering Assessment of the Microserver Chip Market
Analyst David Kanter looks at the emergence of microservers and how the upcoming ARM chips will figure into that market. If you're an ARM licensee looking to make it big in the server space in the next few years, his analysis may be a bit discouraging.
Kanter begins by noting that the server market may be on heading toward an inflection point, since energy efficiency is becoming the driving design across many application domains. This is the result of a new server landscape, which is being driven by cloud computing workloads and which require a scale of computing and efficiency that is far greater from those of the pre-Google era of IT.
The emphasis on energy efficiency and scale means that processors with more numerous, but less performant cores have the edge now. This is the opening that Tilera, ARM and its backers see as a way to challenge Intel's dominance in the server arena. Both ARM and Tilera are based on cleaner instruction architectures than the x86, and this affords them a certain efficiency advantage compared to its the more complex legacy ISA.
But Kanter estimates that a more efficient instruction set only buys between 5 to 15 percent better power efficiency compared to its older competition. And since those numbers only apply to the cores, not the caches or other system resources, the best one can hope for is perhaps a 5 percent advantage for the whole chip. While significant, that 5 percent advantage doesn't amount to much when you consider that Intel's semiconductor manufacturing prowess gives it a 20 to 30 percent performance/watt advantage.
Nevertheless, in a couple of years there should be a plethora of 64-bit ARM server chipmakers, including Samsung, Qualcomm AMD, Applied Micro, Calxeda, Cavium, Marvell, and probably others. Intel will offer its own microserver SoCs, based on its Atom x86 processor architecture.
Kanter believes microserver chipmakers much specialize to survive in this environment, where energy efficiency at scale seems to be driving all other considerations. The trick is capture a market segment that is large enough to sustain SoC volumes and ongoing R&D, but specific enough that it lends itself to a customized approach. Apparently though, Kanter doesn't think there is a whole lot of room for such specialization, since he is predicting that only one or two of the new microserver entrants will survive in the long run.
That kind of pessimism is certainly understandable, but Kanter doesn't consider a few alternate scenarios. One is that Intel's chip manufacturing dominance could fade or even disappear, a possibility that seems more likely the closer we get to the end of Moore's Law. In fact, after the computer industry gets beyond silicon-based computing, there could be a whole new set of players that come to the fore.
Another possibility is that specialization will indeed become a viable strategy for chipmakers. There is indeed a general trend towards more application diversity as the variety of computing devices increases and as more industries take advantage of digital technology. The applications underlying Google, Facebook, and eBay didn't supplant database warehousing, they got added to the mix. It's not to difficult to imagine other equally large application domains becoming commoditized over the next several years.
If either or both of these scenarios come to pass, then not only could there be a diverse set of viable microserver offerings, there could be many more companies involved than we have today. Intel, ARM chipmakers and Tilera might end up having lots of company.