Microservers on the Horizon
The last dramatic change in the world of server microprocessors came in 1995, when Intel released the P6, or Pentium Pro. At the time, the server market was fragmented between half a dozen proprietary RISC architectures and operating systems. Each major vendor, had its own subtly different flavor: SPARC/Solaris, PA-RISC/HP-UX, MIPS/IRIX, POWER/AIX, Alpha/Tru64, plus a handful of other solutions such as IBM’s mainframes or DEC’s VMS. While some servers used Intel’s x86 microprocessors, this was a rarity.
The P6 was the first Intel product to offer competitive performance, and it did so at a fraction of the cost. Moreover, the P6 used a shared front-side bus that allowed system vendors to easily build 1-4 socket servers. The volume economics behind Intel’s business model and the emergence of Windows and Linux as viable server OSes heralded the inevitable decline of proprietary servers. Over the next 5 years, proprietary system would slowly lose ground and this trend significantly accelerated in the aftermath of the dot-com bubble. Now, 17 years after the Pentium Pro, most of the proprietary server solutions have died out and Intel’s server microprocessors are well over 95% of the volume.
Intel’s x86 server microprocessors are far less expensive than the proprietary RISC alternatives they have supplanted. But even still, they are incredibly lucrative and profitable. Tablet or smartphone systems-on-chip (SoCs) are typically 100mm2 or less and sell for around $10-$20. In contrast, server designs might range from 200-600mm2 and sell as high as $4000. For servers, performance and power efficiency is incredibly valuable, which is reflected in the pricing.
Today the server market is on the cusp of a potential inflection point. Over the last 2-3 years, GPUs have emerged as an alternative computational device, specifically for high performance computing. More significantly, at least half a dozen companies have or are slated to release alternative server architectures. Tilera uses a custom instruction set and has been shipping products for networking for quite some time. The ARMv8 64-bit extensions have reduced the barriers to creating server SoCs; with AMD, Applied Micro, Calxeda, Cavium, Marvell, Qualcomm, Samsung, and others following this path. Almost uniformly, these companies are pursuing a re-imagining of the server (so called ‘microservers’), tailored towards scale out workloads that are common in cloud computing. Even Intel is developing a line of processors for microservers, based on the Atom core.
Looking forward, these offerings may inject new competitive life into the market for high density servers. Some optimistic industry participants and observers have predicted that ARM’s entry into servers is as much a watershed moment as Intel’s P6 was in 1995. Before making predictions about the future though, it is good to step back and fully comprehend the past and present. The first step is looking at the anatomy of modern server processors. From there, it is easier to grasp how this new breed of microservers might alter the architectural balance and deliver compelling performance. Ultimately, our analysis suggests that microserver vendors must tailor designs to specific workloads and markets, in order to achieve a sustainable advantage relative to the status quo, and even so there is only room for 1-2 new entrants.
Our analysis of server processors focuses on area as a metric, because it is easily quantifiable. In truth though, power and energy efficiency are the principal design constraints. However, it is nearly impossible to engage in an intelligent discussion of power and energy without actually simulating the physical behavior of the chip and power management under a given workload. While area is not quite the same as power or energy, there is a real relationship because architects can spend area to improve power and energy efficiency. For instance, adding more cores linearly improves aggregate performance, enabling lower voltage operation and better energy consumption (since energy is proportionate to the square of voltage). While our area-centric analysis is imperfect, it is still a good approximation and starting point for discussion.
Discuss (345 comments)