OpenCAPI Consortium – How Open Standards and Development Fuel System Innovation

By Charles King, Pund-IT, Inc.  October 19, 2016

As my colleagues Rob Enderle and Roger Kay discuss in this Pund-IT Review, the new OpenCAPI Consortium announced last week aims to develop new interconnect solutions that will improve server performance by as much as 10X over currently available systems.

That’s impressive in eye-opening ways but even more so is OpenCAPI’s roster of founding members – AMD, Dell EMC, Google, HPE, IBM, Mellanox, Micron, NVIDIA and Xilinx – many of which are forcefully direct competitors. The fact that these companies have agreed to lay down their swords, at least for a while, and contribute their plowshare energies and imaginations to the Consortium makes OpenCAPI one of the more unusual collaborations to come down the pike in some time.

Rob and Roger did a great job covering the technological elements of OpenCAPI and its potential marketplace impacts, especially on Intel, which was noticeably absent from the Consortium (as was Oracle). So I’d like to look a bit further afield and consider how/where OpenCAPI fits into the larger scheme of IT industry progress, and whether it and other open standards and open development efforts offer viable alternatives to traditional methodologies.

Pluses and minuses of industry standards

Within IT, “standards” is a slippery term that often refers to industry standards, but those come in two basic forms:

  1. Standards created, developed and managed by industry-aligned and approved organizations. SAE International, the engineering organization founded in 1902 to oversee standards for components like the nuts, bolts, screws and other fasteners used in industrial and commercial products, is a good example.
  2. Those which become industry standards by de facto – that is, through widespread commercial adoption and use. Intel’s x86 server microprocessors and architectures exemplify this approach since they are used in 90%+ of the PCs and servers sold annually.

Numerous significant benefits can accrue from traditional industry standard technologies and products. Chief among those are highly predictable quality and performance characteristics, points that are especially important in complex computing platforms and systems. Vendors can also drive innovative changes in industry standard platforms, like Intel did over a decade ago with its Centrino wireless adapter and interface.

But certain challenges can also sometimes arise in de facto standards created and controlled by individual vendors. These can include product development efforts that are out of sync with customers’ needs, or failing to anticipate and adapt to critical marketplace changes.

In fact, these are focal points in the OpenCAPI launch announcement that contrasts the Consortium’s plans and goals with the PCIe interconnect technologies long supported by Intel and other system and component vendors. The central concern is that the architectural limitations of PCIe results in system bottlenecks that throttle emerging data-intensive workloads, including advanced analytics, big data, machine learning and deep learning.

Next gen advancements will double the 16 Gbps data rate supported by existing PCIe 3.0 solutions but the PCIe 4.0 spec isn’t expected until sometime in 2017, with the availability of commercial solutions slated for 12-36 months later. In sharp contrast, the initial systems leveraging OpenCAPI (supporting 25 Gbps data rates) should become available in 2H 2017.

In data center terms, that qualifies as instant gratification.

The viability of open standards development

OpenCAPI is impressive by most any technical measure, but let’s consider another point – whether open standards and development efforts are better at fueling innovative, next gen advancements than traditional or de facto industry standard efforts and offerings.

I’d argue that’s a gross over-simplification of a complex market dynamic. If OpenCAPI and associated open technology organizations, like the recently launched CCIX and Gen-z consortiums reach their full, intended potential, server vendors and customers hoping to leverage data-intensive workloads will likely reap substantial benefits. But that success won’t sink Intel or other players with sizable industry standard portfolios and investments.

The larger issue lies in the value that open technologies and highly flexible collaboration models offer development efforts. It isn’t that PCIe has failed. In fact, its decade+ longevity is testimony to the value it offers countless business applications and processes. But PCIe 3.0 is falling behind the curve in regards to complex, data-intensive workloads and catching-up, as noted above, is not a near-term proposition.

Open standard development can certainly boast numerous success stories, but it is anything but perfect. In fact, the nature of these efforts means that some groups and some projects tend to be more successful than others. Lack of centralized control, failed execution and faltering interest have plagued certain open projects or doomed others outright.

However, the collaborative energy behind open technologies can also result in notable achievements. Moreover, the diversity of opinion and engagements at the heart of open standards and development efforts often results in those technologies advancing and evolving at far faster rates than institutionalized industry standards or those controlled by individual vendors.

Quite often, the most successful open technology efforts have the active support of deeply experienced and committed companies and individuals. Looking at the Consortium’s founding membership roster suggests that OpenCAPI will be one of those successful efforts.

IBM’s “open” commitment

IBM’s central involvement and role in the OpenCAPI offers another good reason to take the Consortium seriously. The Coherent Accelerator Processor Interface (CAPI) technologies and tools were, after all, first developed by IBM and offered as part of its POWER8-based Power Systems family. CAPI is also central to upcoming POWER9-based IBM solutions, and play a central role in the bleeding edge, next-gen Summit and Sierra supercomputers that IBM and NVIDIA are building for the U.S. Department of Energy (DOE).

But taking a broader view, OpenCAPI is simply the latest of a long string of open standards/source/development projects that IBM has spearheaded for 15+ years. That began in the mid-to-late 1990s with the company’s support for and investments in Linux development, and continued with efforts like the Eclipse Project (launched in 2001 and open-sourced in 2004).

IBM has continued to support and fund a wide array of open technologies, leading up to August 2013 when the company announced that it was donating its POWER8 processor architecture and IP to the OpenPOWER Foundation it launched with Google‎ ‎, ‎Mellanox‎, ‎NVIDIA‎ and ‎Tyan (and which currently has 270 members). Then last July, IBM launched developerWorks Open, a cloud-based environment for developers to access emerging company technologies, including fifty IBM projects it released to the community.

The point of all this is to underscore the depth and breadth of IBM’s longstanding commitment to open source, open technologies and open standards. In other words, if competitors or vendors committed to PCIe and other interconnect technologies assume that IBM and the other OpenCAPI Consortium members will quickly lose interest and withdraw, they are dreaming.

Final analysis

For years now, high-end computing systems, including supercomputers and high performance computing (HPC) installations have successfully leveraged industry standard technologies and homegrown innovations. But the sheer volume of data fueling new systems and projects is moving beyond the abilities of some industry standard components, resulting in bottlenecks that can throttle system and workload performance.

The new OpenCAPI Consortium aims to address those challenges by leveraging and evolving Coherent Accelerator Processor Interface technologies, resulting in up to 10X better system performance. That’s a worthy goal in itself but the fact that solutions supporting OpenCAPI technologies are likely to begin shipping next year before the ink has dried on the next gen PCIe 4.0 spec should substantially benefit advanced analytics, big data, machine learning, deep learning and other emerging data-intensive workloads.

If the OpenCAPI succeeds as its supporters hope, Consortium membership should grow significantly. If the project’s open development and standards model works as intended, it seems likely that next generation OpenCAPI solutions will eclipse the capabilities of PCIe 4.0. That may not have a noticeable impact on consumer and general purpose business technologies, but it should cement the demonstrable value of open standards and development practices for years to come.

© 2016 Pund-IT, Inc. All rights reserved.