IBM POWER9 and the Value of Innovative Collaboration

By Charles King, Pund-IT, Inc.  December 6, 2017

An interesting dichotomy is common among technology vendors and customers; while most profess love or adulation for cutting edge innovations, many also assume that some leading vendors or platforms are so essentially entrenched that upsetting the order of things is impossible or not worth the trouble.

Call it an odd quirk in supposedly revolutionary mindsets. It’s as if the framers of the U.S. Constitution gave some thought to 18th century Britain’s legendary military and naval might and decided that, what the hell, maybe the Stamp and Tea Acts weren’t really such big deals after all.

What’s doubly weird is how willingly this attitude flies in the face of contrary evidence. Leading and dominant vendors can and do fail enough that examples are easy to find. Silicon Graphics’ (SGI’s) early leadership in computer-generated imagery (CGI) made it a Hollywood mainstay until lame executives’ nearsighted strategies sent the company over a cliff. Sun Microsystems long dominated the scale-up RISC workstation and server markets but was unable to adapt to sea change trends, including the rise of Linux and x86 server silicon.

This isn’t to say that leading vendors lack viable challengers. For example, though Intel currently owns 90%+ of the market for data center processors, the company faces a range of threats, including entrepreneurs focused on developing servers running ARM-based chips. However, IBM’s new next generation POWER9 processors and AC922 Power System servers show that Intel continues to have reasons for concern at the high end of the supercomputing and HPC markets, as well as in promising new areas, like Artificial Intelligence (AI).

Let’s consider that.

The road to POWER9

IBM’s first Power-based systems were the RS/6000 workstations and servers it released in 1990, and new generation chips arrived every 3-5 years afterward. There were some notable diversions along the way. For example, the company used the Power Architecture in PowerPC processors for Apple’s Macintosh systems. Later generation PowerPC chips found their way into embedded applications and gaming systems, including Microsoft’s Xbox, Nintendo’s Wii and Sony’s PlayStation 3.

But the data center was where IBM POWER really shined. POWER3 was the first server processor to utilize copper interconnects. POWER4 was the first commercially available multi-core processor. POWER5 incorporated simultaneous multi-threading, partly to support massively scalable multiprocessing. POWER7’s symmetric multiprocessor design blended performance improvements with power efficiency. POWER8 introduced IBM’s CAPI (now OpenCAPI) extension bus that runs on top of PCIe and can be used to attach dedicated off-chip accelerators, like ASICs, FPGAs and GPUs.

OpenCAPI plays a crucial role in the new POWER9 processors. IBM notes that it designed the new chips from a “blank sheet” with the aim of creating a platform and solutions that delivered demonstrable performance improvements in demanding AI-related applications, like deep learning frameworks, including Caffe, Chainer and TensorFlow, as well as in accelerated databases, such as Kinetica.

IBM’s collaborative innovations

What is OpenCAPI’s role? By supporting embedded PCI-E 4.0 and next gen NVIDIA NVLink 2.0 technologies, OpenCAPI enables POWER9-based solutions, like the new AC922, to move data up to 9.5X faster than PCI-E 3.0-based x86 systems. That also underlies the 4X performance improvements in deep learning training times that POWER9-based systems can deliver. In other words, POWER9 is an absolute beast when it comes to moving data.

That’s a critical point to remember, especially in a performance-dependent, commercially promising, highly competitive area, like AI. The value of systems supporting AI workloads depends heavily on their ability to efficiently perform and dependably speed processes, like the deep learning training exercises AI applications need to repeat thousands or tens of thousands of times.

Enabling customers to finish that work 4X faster than competing systems isn’t just net positive—it’s a potential game-changer. That’s especially true if IBM delivers AC922 solutions that VP Sumit Gupta, (who leads the company’s AI, Machine learning, and HPC products and business line, said will be cost competitive with comparable x86-based systems.

Those improvements also highlight the importance of the collaborative partnerships IBM has developed through the OpenPOWER Foundation it launched in 2013 with founding members, Google, NVIDIA, Mellanox and Tyan. Now boasting over 300-member companies, the Foundation is designed to support a range of new efforts and technologies leveraging IBM’s open source Power Architecture.

One of the highest profile of these efforts are the new “Summit” and “Sierra” 100+ petaflop supercomputers built through the “Coral” (Collaboration of Oak Ridge, Argonne, and Livermore national labs) program. Both systems are based on IBM POWER9 systems, NVIDIA Volta GPUs and related technologies, and Mellanox interconnect solutions. The new installations utterly eclipse the performance of Oak Ridge’s existing Titan system (which utilizes AMD Opteron and NVIDIA Tesla technologies) but do so in a footprint about a quarter of the size of Titan.

Final analysis

Now that POWER9 is commercially available, additional new offerings and future achievements based on IBM’s innovations are sure to follow. As an example, it’s interesting that Bart Sano, VP of Google Platforms was quoted in the POWER9 press release, saying, “The POWER9 OpenCAPI Bus and large memory capabilities allow for further opportunities for innovation in Google data centers.” Google’s usual reticence in talking about strategic efforts makes Sano’s comment doubly intriguing.

But that can’t and won’t overshadow what IBM has achieved here. Vendors tend to win when they deliver technologies that are faster and/or better than other options, or offer customers unique capabilities. At least initially, IBM’s new generation POWER9 offerings lean toward the former part of that equation by notably enhancing existing AI workloads and capabilities. But over time, and with a little help from its OpenPOWER friends, IBM and POWER9 innovations are likely to inspire and inhabit a host of revolutionary new data center solutions.

© 2017 Pund-IT, Inc. All rights reserved.