An Introduction to the OpenPOWER Foundation

In a bid to challenge Intel's dominance of high performance computing environments, IBM have introduced OpenPOWER, a flexible open-source framework based on their POWER processor architecture. Robert Sheldon looks at the potential of this new platform.

Founded in 2013, the OpenPOWER Foundation represents a collaborative effort between IBM and over 100 organizations and individuals from 22 countries. The foundation seeks to challenge Intel’s stronghold in the low-end server market by offering an open ecosystem based on IBM’s POWER microprocessor architecture. The OpenPOWER platform targets the type of hyperscale data centers that support cloud services and other high performance computing (HPC) environments, a market overwhelmingly dominated by the x86 processor.

IBM, along with such companies as Google, Tyan, Mellanox, and Nvidia, originally launched the OpenPOWER Foundation with an ARM-like strategy in mind: IBM would open up its POWER technology stack, including CPUs, motherboards, APIs, and applications, so those participating in the foundation could build systems based on POWER chip technologies. In this way, they could customize the open chip architecture for specific workloads and solutions, with the goal of producing servers that will out-perform Intel systems, while offering a cost-performance advantage.

The OpenPOWER Foundation

The OpenPOWER Foundation is encouraging its members to adopt the OpenPOWER architecture as a way to innovate and evangelize the POWER technologies. To facilitate this process, IBM is licensing the chip specifications to foundation members that pay the necessary fees. However, IBM is not opening up any files or specifications that would allow an organization to make its own POWER clone or derivatives.

The OpenPOWER platform includes bus specifications as well as reference and system-on-chip (SoC) designs. In addition, IBM has open-sourced the firmware, hypervisor, and operating system (OS) so foundation members can customize the platform to best suit their needs. Last year, for example, IBM posted the OpenPOWER firmware to GitHub, where participants can contribute to the many thousands of lines of code.

Foundation members are also expected to contribute their expertise and products into the platform, providing access to other members and, in the process, create an environment of collaboration that benefits all participants.

What started as a handful of organizations has grown to 110 members. Since the foundation’s early inception, organizations such as Rackspace, Samsung, Canonical, Micron, Ubuntu, and Suzhou PowerCore have jumped onboard. More recently, the foundation has formed an advisory board that provides a forum for engaging with other open development organizations. And this past March, foundation members gathered in San Jose for the first official OpenPOWER summit, where they showed off over 10 POWER-based hardware innovations, demonstrating their commitment to provide an Intel alternative for cloud-scale workloads.

The IBM POWER Technology

IBM came into the OpenPOWER game with a systems and processor architecture well suited to cloud and hyperscale data centers. IBM’s latest version of the POWER microprocessor, POWER8, was designed from the ground up to deliver high-performing analytics and big data workloads, with the Intel Xeon systems squarely in the crosshairs.

The POWER8 architecture includes a number of performance-enhancing features such as supporting eight threads per processor and streamlined integration with external processors or I/O cards. In addition, the POWER8 processors are theoretically capable of reaching clock speeds up to 5 GHz, while providing for hefty memory bandwidth through on-chip memory controllers as well as on-chip and off-chip eDRAM caches. The POWER platform also supports a complete little-endian mode for porting software built for Intel systems to the POWER8 platform.

What is perhaps one of the most talked about POWER8 features is the Coherence Accelerated Processor Interface (CAPI), which provides a mechanism for bettering linking the CPU with specialty GPU accelerators. CAPI can help bypass many of the OS and driver overheads often associated with accelerators, providing the same access to the memory and caches as the CPU. As a result, organizations such as Nvidia and Altera can now build accelerators optimized for specific workloads, while delivering greater throughput and lower latency. This feature alone could prove one of the greatest driving forces behind early OpenPOWER adoption.

Although initial benchmarks suggest that the POWER8 system can indeed outperform its Intel counterpart, we’re still too early in the game to draw any conclusions about the OpenPOWER platform. More evidence has to roll in and we need to have a better sense of how the systems will prove out over the long haul and what they will actually cost. We also need to better understand how migrating to a new platform could impact integration with other systems and services.

That said, IBM-and by extension the OpenPOWER Foundation-is already starting to see its efforts pay off. This past November, the US Department of Energy signed a $325 million deal with IBM for supercomputers based on the OpenPOWER platform. The computers, which will incorporate Nvidia and Mellanox technologies, are slated to deliver speeds up to 5 to 10 times faster than today’s leading supercomputers. The new systems will run in the Lawrence Livermore and Oak Ridge National Laboratories, where they will be used to process massive amounts of data in the areas of science, engineering, and national security.

The US is not the only government to get into the act. The British government plans to partner with IBM to facilitate data-centric and cognitive computing research. In this case, the concerned party is the Science & Technology Facilities Council (STFC), one of Europe’s largest multidisciplinary research organizations.

OpenPOWER in Action

If the OpenPOWER Foundation is to make any headway, they must demonstrate that the OpenPOWER technology can be made to work, which is exactly what the organization hoped to achieve at the March summit. There, foundation members showed off an assortment of systems, boards, and cards built on POWER8 technology.

Not surprisingly, IBM was quick to present its Power Systems S824L server prototype, codenamed Firestone, which incorporates Nvidia GPU technology and interconnects from Mellanox. This is the system that will serve as the foundation for the Department of Energy servers. IBM also presented its Data Engine for NoSQL system, a collaborative effort with Altera, Canonical, and Redis Labs for deploying NoSQL data stores.

Rackspace brought its own contributions-an open server specification and motherboard mock-up. The Rackspace offering combines the OpenPOWER platform with OpenStack and Open Computer technologies. At the same time, Tyan introduced the first OpenPOWER server to be made commercially available. The server integrates Tyan’s OpenPOWER customer reference system and targets large-scale cloud deployments.

Several of the prototypes shown at the summit leverage the POWER8 CAPI features, such as the CAPI developer kit from Convey, the ConnectX-4 adapter card from Mellanox, and the shared virtual memory solution from IBM and Altera. The summit also introduced the first OpenPOWER development platform, the Cirrascale RM4950, a collaborative effort between Nvidia, Tyan, and Cirrascale. The solution provides a GPU-accelerated platform for analytics, deep learning, and scientific computing applications.

The summit also presented plenty of other prototypes from its members. There was even reference to a POWER-based server board that Google had built the previous year. But it was the summit demos themselves that caught the participants’ attention and allowed the OpenPOWER Foundation to demonstrate what it can do and what it hopes to do in the future.

The China Connection

Chinese companies were well represented at the March summit and could prove to play a pivotal role in how the OpenPOWER Foundation moves forward. With pressure from their government to favor local suppliers, Chinese companies are looking for ways to advance their technologies while adhering to national trade policies. The OpenPOWER platform would allow these companies to build their own systems with all-Chinese resources, but based on IBM’s POWER architecture.

Indeed, the OpenPOWER model must have clicked with many Chinese companies, given how well represented they were at the summit and how many prototypes they collaborated on. Yet it’s not just the OpenPOWER platform that had tempted the companies into the organization.

In the fall of 2014, the Chinese government formed the China POWER Technology Alliance (CPTA), a public and private alliance that served as a formal endorsement of the OpenPOWER Foundation and its efforts. The alliance’s goal is to bring the POWER architecture to China, making it possible to build complete systems within the country as an alternative to Intel.

Not surprisingly, the CPTA nudged a number of Chinese companies into the fold, including Inspur, ChuangHe, and PowerCore. PowerCore has already agreed to license the POWER8 chip and plans to release the first non-IBM POWER server chip, dubbed CP1. Zoom Netcom will utilize the chip in its new RedPower line of servers, all of which will be built in China.

The extent to which the Chinese companies will impact the OpenPOWER movement is yet to be seen, but the fact that the government has thrown its weight behind OpenPOWER is no small matter. Not only does this endorsement help to shore up OpenPOWER forces, but it could also lead to a market shift with far reaching consequences-to Intel and to other non-Chinese companies-as China’s dependency on foreign technology suppliers continues to wane.

The OpenPOWER Advantage

It’s not just the Chinese who stand to benefit from the OpenPOWER tide. Today’s workloads demand that we rethink the current processor-centric model and look to innovative approaches that can better scale to meet the high-throughput, low-latency demands of massive data workloads. An open platform such as OpenPOWER offers an opportunity to bring new perspectives and innovations to our current approaches for handling data and supporting large-scale operations.

The Linux OS demonstrates what is possible through open collaboration. Although Linux has been slow to catch on for desktop deployment, the story at the server level is much different, having made significant inroads into private and public networks. According to the Linux “2014 Enterprise End User Trends Report,” 75 percent of enterprises participating in the study said they use Linux as their primary cloud platform. Android is another open source project that has seen worldwide adoption.

Although the OpenPOWER Foundation is striking out into new territory by opening up the technology stack, the organization still hopes to benefit from the innovation that open collaboration can bring. Working together, members can share their expertise and experiences in order to achieve a new level of data-centric computing that might not be possible for the company that tries to go it alone. Perhaps IBM understood this when embarking on the OpenPOWER strategy. The world is looking for systems that can address the needs of the growing number of hyperscale data centers. By working together under the OpenPOWER umbrella, members can contribute to the discussion and benefit from the innovation and technical advancements, while being able to build custom solutions specific to the needs of their customers. If this trend takes hold, a broad range of organizations will be exposed to the OpenPOWER technology and get to see it in action for themselves.

Certainly it could be argued that IBM chose the OpenPOWER route, at least in part, because of lagging server sales and the diminished demand of RISC-based machines, leaving the company no choice but to seek help elsewhere to help stem the Intel tide. Even if that is the case, the OpenPOWER Foundation still represents a larger trend in which opening up technology and collaborating at all levels might prove the only way to achieve the types of technological breakthroughs that the new era of big data demands.

That’s not to say IBM and many of the other foundation members would not like to strike a definitive blow against Intel, if not take them out altogether. Intel reportedly commands 98 percent of the server chip market. But the company is no easy target. It has a reputation for being a well-run organization, with plans to keep investing in its own development efforts. Anyone who considers going up against Intel must be ready to take on a formidable opponent.

Not So Fast, OpenPOWER

OpenPOWER advocates hope to win converts with what they perceive as the platform’s superior architecture, pointing to such POWER8 features as CAPI, SMP scalability, memory bandwidth, and expanded multithreading capabilities, which can lead to significant cost-performance advantages over Xeon. By its very nature, the POWER8 platform already seems prepared to take on the challenges of GPU-accelerated deep learning and scientific analytics. And because OpenPOWER is based on an open platform, members will be able to innovate to a degree that will be difficult for Intel to achieve as it continues to work behind closed doors. Even so, there are migration issues and legacy systems to consider when introducing a new technology. Intel’s 98 percent share of the market will likely keep the reigns in Intel’s hands for some time to come.

The OpenPOWER summit demonstrated the potential of an open platform and the commitment by the foundation members to come up with ways to meet the future challenges of massive data processing. One of the unanswered questions, however, is whether the platform will be compelling enough to convince decision makers to migrate to such a new system. There have been plenty of suggestions that OpenPOWER will offer big price-performance advantages, but the industry will need to see more concrete details to be able to make any sort of informed decisions.

The Intel processor is a tried and true architecture. Although it has limitations, just like any technology, Intel’s command of the market puts the onus on OpenPOWER to convince buyers to disrupt their standardized environments. At the same time, Intel won’t be standing idly by while OpenPOWER tries to catch up. The company continues to invest heavily in its data center business and advancing its own technologies, such as incorporating the Haswell microarchitecture into the Xeon E5 series. Intel has a proven track record. Trying to grab pieces of that massive pie will be a challenge for any competitor.

Even so, the OpenPOWER Foundation has one important advantage. Members can collaborate on a global scale, leading to greater innovation and faster implementation of new technologies. Then there’s IBM, which has much to gain by taking Intel down a notch or two. Many companies, in fact, would like to break the Intel stronghold, and the Chinese government would like to remove Intel from the equation altogether. This next year will be an important one for OpenPOWER, as products built on the new platform start to emerge. Whether it will be enough to significantly change the server landscape is yet to be seen, but with 98 percent of the server market share, Intel has little place to go but down.