HPC Market Showdown: GPU Momentum Growing
While still generally experimental, GPUs have tripled their worldwide footprint in the past two years
Steve Conway
What a difference two years can make in the fast-paced world of HPC technology adoption — a market considerably less risk-averse than its mainstream IT counterpart. IDC’s 2008 worldwide study on HPC processors revealed that nine percent of HPC sites were using some form of accelerator technology in their installed systems. GPGPUs (henceforth to be called GPUs) shared the accelerator habitat back then with FPGAs, Cell processors and a few rarer species.
Fast forward to the 2010 version of the same IDC global study, and the scene has changed considerably. Accelerator technology has gone forth and multiplied. By this time, 28 percent of the HPC sites were using accelerator technology — a three-fold increase from two years earlier — and nearly all of these accelerators were GPUs. (FPGAs were the only other accelerators to exhibit growth, although at a much slower rate than GPUs.)
True, the GPU population explosion has been much wider than deep. While about one-quarter of the HPC sites employed GPUs in 2010, GPUs made up only five percent of the total processor count in the subset of installed hybrid systems, and a little more than one percent of the aggregate processor count in the superset of all installed HPC systems, both hybrid and non-hybrid.
GPUs today are still far from being able to tell x86 processors they have 24 hours to get out of town, but GPUs are clearly proliferating in the HPC market at a decent clip. An important sign is the spread of GPU-related academic offerings. At the September 2010 HPC User Forum meeting in Seattle, Rob Farber of the U.S. Department of Energy’s Pacific Northwest National Laboratory reported that GPU computing is part of the curriculum at more than 200 universities around the world, including marquee names, such as MIT, Harvard, Cambridge, Oxford, the Indian Institutes of Technology, National Taiwan University, and the Chinese Academy of Sciences.
What’s driving this technology dissemination?
The GPU appeal
The appeal of GPUs for high performance computing is manifold:
- First, the slowdown of Moore’s Law-governed generational advances in the single-threaded performance of x86 processors has left a wide opening in the market for alternatives. GPUs, like vector processors before them, promise attractive speedups on the important subset of application codes exhibiting data-level parallelism and thread-level parallelism.
- Second, GPUs provide a lot of peak and Linpack flops for the money. Especially for HPC sites seriously pursuing the upper ranges of the Top500 supercomputers list (www.top500.org), bolting on GPUs can provide a kind of flops warp drive, rocketing Linpack performance to where almost no one has gone before. Witness China’s Tianhe-1A supercomputer, which supplemented Intel Xeon x86 processors with NVIDIA GPUs to seize the number one spot on the November 2010 Top500 list. By June 2011, three of the top 10 systems on the list employed GPUs.
- Most important, GPUs are already enabling real-world advances in HPC domains, especially the life sciences, oil and gas, financial services, and digital content creation and distribution. GPUs are a particularly promising fit for molecular dynamics simulations, which extend across multiple applications domains.
Who’s on First?
In 2010, NVIDIA led the HPC charge in alternative processor adoption with the company’s Tesla GPU offerings — including a prominent role in China’s Tianhe-1A supercomputer. Going forward, NVIDIA promises to remain a hot ticket, but the GPGPU competition will heat up. Some large buyers have told IDC they prefer AMD’s ATI road map plans for their applications. Another new competitor will be Intel’s Many Integrated Core (MIC) technology, which incorporates GPGPU-like capabilities from Intel’s scaled-back Larrabee project in an architecture that shares the same standard Intel IA programming environment as the company’s CPUs. Next up on the MIC road map is the 50-core processor code named Knights Corner that is due out in 2012.
GPUs also will continue to face competition from FPGAs for some opportunities.
Adoption barriers
HPC buyers consistently report the following main barriers to more extensive GPU deployment:
- Ease-of-programming: Despite the availability of useful tools, such as CUDA, OpenCL and The Portland Group’s directives-based compiler that’s designed to transform Fortran or C source code into GPU-accelerated code, HPC buyers and end-users typically report that programming GPUs remains more onerous than the more familiar approaches to programming x86 processors. Presumably, this barrier will continue to drop over time as familiarity with GPU programming methods grows — through the more than 200 universities offering GPU curricula, and elsewhere — and as the GPU programming methods themselves advance.
- Mediated communication: Another issue frequently cited by HPC users is the fact that GPUs today are typically implemented as co-processors that need to communicate with x86 or other base processors via PCI-Express channels that are comparatively slow — at least when weighed against implementing the CPU and GPU on the same die. This mediated communication affects some applications more than others.
- Waiting for future CPU generations: Some HPC users believe that waiting to see what improvements future-generation x86 processors deliver is a risk worth taking, compared with the effort of learning how to program GPUs and adapting portions of their codes to run on GPUs. And, because GPUs are still relatively new devices for high performance computing, some users worry that the substantial effort to re-write their codes could be wasted if GPU architectures evolve in a new direction or if GPUs are not a long-lasting phenomenon in the HPC market. This wait-and-see group has been declining as GPUs have increased their influence in the global HPC market.
The bottom line
GPUs, as one prominent HPC expert put it, “are generally still in the experimental phase.” But they have tripled their worldwide footprint at HPC sites in the past two years alone, they have become more indispensable for attaining prominence on the closely watched Top500 supercomputers lists, and they have enabled a growing number and variety of real-world achievements.
As GPU hardware and software technologies advance, as more university students and others learn how to exploit GPUs, and as more GPUs become available to the world’s most creative scientific, engineering and computational minds, it seems almost inevitable that GPUs will settle into an important, natural role, complementing x86 processors within the HPC ecosystem.
Steve Conway is Research Vice President, HPC at IDC. He may be reached at editor@ScientificComputing.com.
