IBM, OpenPOWER and Gaining a Competitive Edge in HPC

By Charles King, Pund-IT, Inc.  September 14, 2016

High performance computing (HPC) represents a pinnacle of computational excellence that’s also amazingly cool for those of an IT bent. It isn’t just the systems themselves, though there’s much to consider there. What’s more remarkable and often more impressive is how HPC can enable flights of the imagination that find landing places in the real world.

The effects of these excursions are often apparent to the narrowest of audiences. After all, the results of classified nuclear weapons research aren’t open to the general public but many other HPC-inspired advancements are. Plus, the same rules of commoditization that impact other IT markets hold true for HPC, too. So that HPC capabilities and applications that were unthinkable just a few years ago enable numerous commercial solutions and services today.

Those same commoditization rules make HPC every bit as dynamic as any other commercial IT market, if not more so. That’s because HPC vendors and their customers are always looking for an advantage, a way to get ahead and achieve consistently leading edge results. At the same time, HPC is one of IT’s least sentimental practice areas.

“What have you done for me today?” isn’t a cliché in HPC. It’s a mantra repeated daily, weekly, monthly, annually by scientists and research professionals on the hunt for the next big thing, and the bigger things to follow.

IBM’s new Linux server line additions

Why consider these points? Because the new systems IBM added to its Power-based Linux Server portfolio last week are designed to pique the imaginations and engage the attentions of HPC professionals. If you’re unfamiliar with these servers, the Power LC line is being developed by IBM and fellow members of the OpenPOWER Foundation, a coalition actively developing and promoting data center solutions based on IBM’s POWER processor architecture.

The three new Power LC offerings take particular advantage of new interconnect innovations known collectively as POWERAccel developed by a subset of OpenPOWER members creating solutions optimized for applications requiring robust, accelerated performance. Those include artificial intelligence (AI), deep learning, high performance data analytics and other compute-heavy workloads.

The new IBM Power System S822LC for HPC leverages a newly designed processor; the IBM POWER8 with NVIDIA NVLink. NVLink, a high speed, bi-directional, energy efficient interconnect is embedded at the silicon level and integrated into the larger system design. The Power System S822LC for HPC is an industry first with NVLink directly connecting a single POWER8 CPU with up to 2 NVIDIA Tesla P100 Pascal GPUs. IBM says that as a result, data flows through the new solution nearly 3X faster than on analogous Intel x86-based systems.

Two other new servers that were announced – the IBM Power System S821LC and the IBM Power System S822LC for Big Data – can leverage NVIDIA Telsa K80 GPU accelerators via PCIe connections. Clients can also use the POWERAccel Coherent Acceleration Processor Interface (CAPI) available on the S821LC and S822LC for Big Data models to leverage FPGA accelerators with hardware managed memory coherency.

Power performance/efficiency and real world benefits

As noted in the Power LC announcement, the new servers are optimized for data rich applications and are fully compatible in Linux-based cloud environments. But how will their advances benefit HPC customers?

Let’s consider the Power System S822LC for HPC first. According to IBM and NVIDIA, customers who are first in line to receive shipments of the new systems include the U.S. Department of Energy’s Oak Ridge National Laboratory (ORNL) and Lawrence Livermore National Laboratory (LLNL).

Both plan to use the new servers for early-generation test beds for developing applications for Summit and Sierra, the next generation supercomputers that IBM expects to deliver respectively to ORNL and LLNL in 2017. Lab representatives cited significant reductions in programming complexities and improved application performance as reasons for their interest in the new IBM systems.

In addition, IBM provided a cluster of the new systems to China-based ISP Tencent for testing data-intensive workloads. The result? The Power LC systems cluster, which had 1/3 the number of servers of Tencent’s existing x86-based infrastructure, ran the workload 3X faster than what the company had previously been able to achieve. Significantly boosting compute performance and reducing data center footprint and related facilities costs is the kind of value proposition HPC customers dream about.

The new IBM systems also offer distinct price advantages over Intel-based servers, saying that some Power LC configurations cost up to 30% less than comparative x86-based offerings. Additionally, the company claimed the new servers can deliver up to 80% more performance per dollar spent over x86-based systems. Online pricing for IBM’s Power LC line begins at $5999. Models with smaller configurations and lower pricing are available through IBM Business Partners.

Final analysis

What are we to make of all this? A number of points are worth considering. First, the new offerings should be considered in the context of IBM’s longstanding leadership position in HPC and supercomputing. In fact, the Summit and Sierra supercomputers the company is developing with NVIDIA for ORNL and LLNL are expected to be the world’s first 200 petaflop systems, delivering over twice the performance of China’s Sunway TaihuLight, currently the world’s fastest supercomputer.

The point is that IBM understands the technical and practical needs of HPC environments and users, and knows how to satisfy them. The fact that the new Power System S822LC for HPC will help researchers at ORNL and LLNL get a running start on programming and simulations prior to the delivery of Summit and Sierra next year is a case in point. But IBM is also leveraging its own POWER8 and NVIDIA’s NVLInk innovations to deliver those same advances to a wide variety of demanding industrial and business applications and customers.

At the same time, the new Power System S821LC and S822LC for Big Data, which leverage NVIDIA GPU accelerators via PCIe and CAPI interconnects, demonstrate that IBM is not shortchanging other commercial use cases. The combination of aggressive pricing and enhanced performance will make the new systems attractive to numerous public sector, university lab and enterprise customers. The same goes for cloud service providers (CSPs) looking for powerful, cost-effective engines to support compute-heavy workloads and services.

Overall, the new additions to IBM’s Power LC line show the company doing what it has always done well – significantly extending the boundaries of computing technology and then bringing those innovations to commercial markets. In addition, the involvement of NVIDIA and other POWERAccel participants from the OpenPOWER Foundation shows that IBM clearly understands how innovative partners can enhance and advance new endeavors.

These are the reasons that the company remains at the leading edge of supercomputing and HPC. It’s also why IBM solutions have enabled so many flights of the imagination to find places and make a difference in the real world.

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