Trends that emerged a few years ago continue to transform the server-processor market. Although Intel still dominates with about 99% market share, glimmers of new competition are appearing. Cloud computing and hyperscale data centers operated by companies such as Amazon, Apple, Facebook, Google, and Microsoft are creating opportunities for ARM-based vendors to challenge the x86 status quo. Even Intel’s old nemesis, AMD, is trying to make a comeback. Rumblings in China and Russia hint at fresh competition from afar, motivated by nationalist policies that encourage independence from U.S. technology. In the near term, these trends will fail to challenge Intel’s dominance, but they are a bellwether of things to come.
Until recently, processor performance and capital expenses were the vital metrics for server customers. 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 that pack more nodes into precious rack space.
Although software compatibility remains important, data-center customers are willing to adopt new platforms if the cost savings and performance are compelling. But ARM’s challenge to the incumbent x86 architecture is unfolding slower than stakeholders expected. By mid-2016, only three companies were shipping 64-bit ARM-compatible server processors: AMD, AppliedMicro, and Cavium. None of their initial products comes close to Xeon in performance or performance-per-watt, except on a few niche applications. Although more competition is coming soon, so far Intel’s newest Xeon and Xeon D products are more than holding their own.
Another new challenger, IBM, formed the OpenPower Foundation to promote its Power Architecture. The company’s Power8 server processors are now available on the merchant market. So far, however, they suffer from power-efficiency handicaps that won’t be addressed until the next-generation Power9 debuts in 2017.
Server processors are offering more CPU cores and more threads, which are valuable for parallel processing—running multiple tasks at the same time. Intel’s Xeon now offers as many as 24 cores and 48 threads per chip. Cavium’s ThunderX CN8890 has 48 single-threaded cores. These high-thread-count processors are valuable for big-data analytics, web servers, scientific computing, transactional databases, and other applications.
For low-cost processors targeting high-density servers, 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 are SoC designs. Intel’s new Xeon D combines the CPU and south-bridge chips in a single package, mimicking these SoCs, but it still lags in hardware acceleration and networking integration.
Intel maintains a shrinking lead in chip-fabrication technology. The newest Xeon processors are built in 14nm FinFET technology, which reduces cost and power while improving transistor performance. By contrast, all non-Intel server processors are currently manufactured in older technologies without FinFETs. But several server processors built in 14nm or 16nm FinFET processes will ship in 2017.
The other x86 vendor, AMD, is struggling financially and has allowed its Opteron x86 server processors to fall far behind. Its initial ARM-based server chips are also unexciting. The company is betting almost everything on its next-generation Zen microarchitecture. The first Zen-based server processors are due to ship in mid-2017, however, so AMD’s fortunes will change slowly. To generate new revenue and open a new market, the company has licensed its processor technology to a Chinese joint venture (Thatic).
AppliedMicro’s third-generation X-Gene 3 will have 32 cores — a big leap over the octa-core X-Gene 2. If the company can deliver this product by the end of 2017, it will have the first ARM-based processor that compares well against Xeon E5 in per-thread and multithread performance, thrusting the company squarely into the mainstream server market.
ThunderX has already shipped for millions of dollars in revenue. Cavium is already working on 54-core ThunderX2 with improved performance that is due for production in 2H17. The company badly missed its targets on the first ThunderX, however, and needs to focus on execution. Qualcomm recently confirmed that its first server processor, which will likely have 32 ARM cores, is on track for production in late 2017. Broadcom has also been developing a powerful ARM-based server processor, but the project may be sold or shut down after Avago’s recent acquisition.
More competition may come from Baikal (Russia), HiSilicon (China), and Phytium (China). All are developing ARM-compatible server processors. But their early products are unlikely to be competitive with those from established vendors. High-performance server processors are difficult to design, and the first chips will be built in older fabrication technology. The Chinese and Russian governments — increasingly resentful and suspicious of U.S. technology — are encouraging or actively backing these projects. These new companies won’t compete globally in the near term, but they could block foreign competitors in their local markets.