The market for server processors is changing, creating openings for new vendors. With the emergence of mega data centers and cloud computing, server economics no longer focuses on capital expenses alone. Demand for ultimate performance from a single processor has been replaced by a balanced view of capital and operating costs. Performance per watt and performance per watt per dollar are the new metrics driving purchasing decisions in large data centers. Physical density is also growing in importance, driving greater scalability and new form factors such as microservers that pack more nodes into precious rack space.
The market is moving to a new era where backward compatibility is less important than before and innovation takes the front seat. Intel and AMD—the incumbent vendors—continue to innovate and advance their Xeon and Opteron designs, respectively. Integration, microarchitecture advances, and process technology are the primary factors when evolving these x86 processors. But new entrants are eyeing cloud-computing environments as an opening for radically different architectures and more-power-efficient CPU designs. Because the merchant server-processor market exceeds $8 billion, success requires taking only a few percentage points of share from Intel.
Having reached practical power limits, server-processor designers are increasing performance primarily by adding cores rather than increasing clock speeds. Mainstream x86 processors now offer 16 cores per chip, while ARM vendors are planning 32-core processors. Intel has recently refreshed its entire lineup, moving to 22nm FinFET technology to reduce cost and power. By moving to a finer-geometry process, vendors get more transistors in the same die area and power envelope. This additional transistor count can be used to add CPUs or to increase cache sizes. Larger caches increase performance by absorbing DRAM latency, which is not decreasing as rapidly as processor performance is growing.
For low-cost processors targeting microservers, a single-chip design is now common. ARM vendors such as AppliedMicro bring this SoC model from their embedded products, and both AMD and Intel have responded with SoC products of their own. Intel’s Avoton processor, for example, integrates eight lightweight Atom CPUs along with the traditional north-bridge and south-bridge logic as well as four Gigabit Ethernet ports.
Intel also offers a broad line of higher-performance server processors. These devices include, for single-socket servers, the Xeon E3 v3 family based on the company’s newest Haswell CPU design. For the popular dual-socket configuration, Intel offers the new Xeon E5 v2, using the same 22nm process as the Xeon E3 but the slightly older Ivy Bridge CPU. This product, code-named Ivytown, scales to 12 cores. For top-end performance in systems with four or eight coherent processors, the Xeon E7 v2 delivers up to 15 cores at 2.8GHz in a 155W thermal envelope. By adding reliability, availability, and serviceability (RAS) features to the E7 line, Xeon now serves mission-critical designs that formerly required Itanium (IA-64) processors.
AMD’s Opteron chips have fallen far behind Xeon in both performance and market share. Using a 32nm SOI process, they generally deliver about half the performance per watt of comparable 22nm Xeon processors. The new Piledriver-based Opterons do little to improve on this short¬coming. For microservers, AMD also offers the Opteron X1150, based on its low-end Jaguar CPU, that falls well behind Intel’s Avoton. AMD’s products are most competitive for applications that don’t need maximum CPU performance because they are I/O or memory bound.
After initial enthusiasm over 32-bit processors from vendors such as Calxeda, Marvell, and Tilera, data-center operators have standardized on the new 64-bit ARMv8 instruction set, leaving out these RISC pioneers. The first ARMv8 server processors will enter production in 2014, led by AppliedMicro’s X-Gene. This eight-core product could outperform Intel’s eight-core Avoton, but the 40nm X-Gene is likely to use considerably more power. AppliedMicro is already sampling a 28nm X Gene that should reduce the power gap.
AMD, ironically, is the other leader in the ARMv8 race. As part of its “ambidextrous” strategy, the company is sampling a processor with eight Cortex-A57 CPUs that is due for production by the end of 2014. This SoC design should better match up against Avoton in power efficiency. Until these ARMv8 vendors release full specifications and performance scores for their products, however, a detailed comparison against Intel’s products is impossible.
Cavium is also developing an ARMv8 processor, code-named Thunder, scheduled for production in 2015. Broadcom has announced a high-end ARMv8 CPU known as Vulcan that could approach Xeon in performance, but it is even further out in time. Other ARMv8 vendors such as Qualcomm and Samsung could jump into the server market over time, creating additional competition for Intel.
Having 96% market share, Intel faces no effective competition in many segments of the server market and is able to name its own prices. This situation is good for Intel shareholders but bad for Intel customers. These customers are actively supporting new competitors, as well as AMD, in hopes of putting pressure on Intel to moderate its processor prices. ARM competition has already forced Intel to increase its pace of innovation, introducing Atom-based SoCs to the microserver market. It may take a few more years, but ARM vendors are likely to eat into Intel’s market share and return balance to the server-processor market.