With 2011 officially in the books, it’s time to offer a few predictions about the upcoming year in HPC. In general, I expect 2012 to continue the major trends we’ve seen over the past couple of years, namely the increased adoption of GPU computing into the mainstream and more parity of HPC capability around the world, as exemplified by China. There may, however, be one or two new trends to pop up.
More flops at the top
No big surprise here. By the end of the 2012, the top 10 supercomputers in the world will be dominated by 10 petaflop-and-above machines. Currently, Japan’s K computer is the one and only double-digit petaflopper on the planet. The move to multi-petaflops hardly qualifies as crystal-ball prediction, given that 10-petaflop-and-above contingent — Blue Waters (NCSA), Sequoia (LLNL), Titan (ORNL), Mira (ANL) and Stampede (TACC) — are all scheduled for boot-up in the second half of 2012.
The big machines mentioned above are all US-based systems, but it wouldn’t be too shocking to see a Chinese super or two in the 10 petaflop realm before the end of the year. In any case, the prevalence of so many multi-petaflop systems in the States will reestablish the US as the leader for elite supercomputing.
Meanwhile, expect China to continue to broaden its share of the planet’s top supercomputers. By next November, I’m guessing China will claim more than 100 of the top 500 systems in the world. Europe, on the other hand, looks like it could be heading for recession in 2012, and may have to tread water on the HPC front until governments regain their financial footing. In the US, federal spending for big-ticket HPC programs (e.g., UHPC) may also be a bit constrained given the apparent lack of political enthusiasm for publicly funded R&D.
Accelerators, especially GPUs, will continue their inroads into HPC. I expect NVIDIA’s introduction of its next-gen Kepler processor to be the biggest news in this area for 2012. Kepler looks like it will be in production no later than the second half of 2012, which is when Cray is supposed to start plugging the new parts into the Titan super at Oak Ridge National Lab.
Intel’s Many Integrated Core (MIC) debut in Knights Corner isn’t expected to go into full production until early 2013, but the Stampede system at TACC is supposed to get some early MIC silicon at the end of 2012. That won’t stop Intel from closing some more MIC deals for some big supers that will be on the drawing board this year.
Meanwhile, AMD should introduce its next-generation FireStream products for HPC with its latest and greatest Radeon GPUs. I say “should,” but given there has been so little activity in the FireStream group lately, I’m wondering if AMD is even planning to continue the product line. Even if AMD refreshes the FireStream parts, I’m not sure how much attention that will receive, given the lukewarm adoption of OpenCL, AMD’s preferred GPU computing framework. A CUDA port for the Radeon architecture, which is now more likely than ever, could change the game, however.
We’re all big data vendors now
The big data/analytics side of HPC heated up in 2011 and should continue to do so in 2012. The debut of IBM Watson was a pointed demonstration of the power of custom analytics running atop a supercomputer. The technology is already being adapted to healthcare analytics and appears to have a lucrative future.
In 2012 a lot more HPC vendors will try to leverage their natural advantage in parallel architectures and fast storage to capture big data customers. Even supercomputing stalwarts, like Cray, want a piece of the action. Those looking to expand into the business analytics space will do well to hook up with companies with some experience in this realm — SAS, Cloudera, and so on. Key hardware technology to watch is solid state storage; key software technology is Hadoop, especially the variants tweaked for performance.
After years of consolidating 10, 40 and 100 GbE technologies and products, Ethernet may now indeed be ready for prime time in high performance computing. The number of TOP500 system using 10 GbE as the system interconnect doubled from 7 to 14 over the last year, and I wouldn’t be surprised if it doubled again in 2012. Look for companies like Brocade, Cisco, Mellanox, and a handful of others to make some big HPC-related Ethernet news this year.
FDR InfiniBand, meanwhile, may get off to a somewhat slow start. Although the bandwidth is nearly twice that of QDR, latency is no better, at least in the current Mellanox product set. QLogic has yet to introduce it’s FDR InfiniBand products, so right now the technology is a single-source solution. Also, FDR adapters needs PCIe 3.0 to perform optimally and only a handful of chipsets in server ecosystem includes that support right now.
It’s not easy being green
Datacenter power is becoming a bottleneck everywhere, and that includes HPC. Power efficiency is just not keeping pace with performance, and for users who need to add capacity into existing facilities, that’s a problem. Besides, running a 10MW supercomputer to study global warming just seems like an oxymoron.
Every vendor from the chipmakers and OEMs up to workload managers and application developers will need a strategy to minimize energy use. Most of these, especially at the processor and system level, are in motion now, but we’ll see more progress on the software side in 2012.