Cloud computing and hyperscale data centers operated by companies like Amazon, Apple, Facebook, Google, and Microsoft are changing the market for server processors, creating opportunities for new vendors. In the past, raw processor performance and capital expenses were the vital metrics. Today, long-term operating costs are equally important. 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 systems such as microservers that pack more nodes into precious rack space.
Although software compatibility remains important, data-center customers are more willing than before to adopt new platforms if the cost savings and performance are compelling. This opening has created an opportunity for ARM and its numerous licensees to challenge the incumbent x86 architecture. But this drama is unfolding slower than expected. As 2015 dawned, only one 64-bit ARM server processor — AppliedMicro’s X Gene — was actually shipping in systems. Although more are coming soon, Intel is responding vigorously with new versions of its Xeon and Atom-based server processors.
Meanwhile, a new challenger has emerged: IBM. In partnership with its rapidly growing OpenPower Foundation, the company is offering its Power server processor on the merchant market for the first time. Because the server-processor market exceeds $9 billion annually, these newcomers can build a profitable business by taking only a few percentage points of share from Intel.
Server processors are raising performance using more CPU cores and more threads, which are extremely useful for parallel processing — running multiple tasks simultaneously. Intel’s Xeon products now offer as many as 18 cores per chip. Cavium’s 48-core ThunderX CN88xx is the largest ARM-compatible processor announced to date. Intel’s Hyper-Threading technology enables a CPU core to execute two threads at once, and Broadcom is developing ARM-compatible server processors that can run four threads per core. Together with advances in parallel-programming tools, these highly threaded processors are valuable for big-data analytics, web servers, scientific computing, transactional databases, and many other applications.
For low-cost processors targeting microservers, system-on-a-chip (SoC) design is now common. By integrating several functions that previously required two or more separate chips, SoCs can cut costs, reduce power consumption, and save board space. Those functions typically include memory interfaces and I/O interfaces. More often, they also include hardware accelerators for specialized tasks. All ARM server processors announced to date use SoC designs. Intel’s new Xeon D combines the CPU and south-bridge chips in a single package, mimicking these SoC products.
In the past year, Intel has refreshed its entire Xeon line, moving to 22nm FinFET technology to reduce cost and power while improving transistor performance. By contrast, all non-Intel server processors are still manufactured in 28nm or older technologies and lack FinFETs. The sole exception is IBM’s Power8, which is manufactured in a 22nm silicon-on-insulator (SOI) process. IBM has always owned and operated its own fabs but recently transferred them to GlobalFoundries, leaving Intel as the only vendor still manufacturing its own server processors for the merchant market. Intel’s 14nm process is already in volume production, but the low-end Xeon D is the only server processor using that node.
The other x86 vendor, AMD, is more static. Struggling financially, the company has managed only minor updates to its Opteron x86 server processors in recent years, falling further behind Xeon in performance and performance per watt. As a result, Opteron’s market share has dropped to 3%. The company’s server hopes rest on a new CPU micro-architecture, code-named Zen, that should reduce the performance gap. New CEO Lisa Su recently prioritized Zen over a new ARM project (K12) in hopes of keeping Zen-based products on track for 2016 production. In the meantime, AMD’s first ARM-compatible server processor, the Opteron A1100 (Seattle), was originally scheduled to ship in 2014, but with little customer interest, production has been delayed until 2015.
In late 2014, AppliedMicro’s X-Gene became the first 64-bit ARM server processor to reach the market. It is a highly integrated eight-core SoC that inherits several features from the company’s high-performance embedded processors. X-Gene 2 moves to newer 28nm technology and is scheduled for production in 3Q15.
Cavium is the next ARM server-processor vendor to reach the market. Its ThunderX CN88xx chips have been sampling since 4Q14 and are sched-uled for production in 3Q15. They have up to 48 CPU cores, allowing them to match the per-socket performance of low-end Xeon E5 products. ThunderX is well suited to scale-out applications (such as web servers) that can exploit its high parallelism.
Other ARM-compatible processors are in the pipeline. Broadcom has been developing Vulcan, a high-end ARM-compatible CPU core, for embedded and server applications; we expect the first Vulcan-based products to enter production in 2016, although the pending Avago deal may disrupt these plans. HiSilicon, a subsidiary of Huawei, has demonstrated a 16-core ARM server processor that could reach production this year. Qualcomm has confirmed its plans to develop ARM-compatible server processors but disclosed no timeframe for its entry.