June 09, 2011
In case you were wondering if these new-fangled Chinese GPU-powered supercomputers can do anything useful, Thursday's announcement about the latest exploits of the Tianhe-1A system should give you some idea of the significance of these petascale beasts. On Thursday, researchers from the Chinese Academy of Sciences' Institute of Process Engineering (CAS-IPE) claimed to have run a molecular simulation code at 1.87 petaflops -- the highest floating point performance ever achieved by a real-world application code. The simulation is being used to help discern the behavior of crystalline silicon, a material used in solar panels and semiconductors.
According to NVIDIA, the application used just 2,000 lines of CUDA to accelerate the simulation -- not an inconsequential amount of source code, but considering the result, a pretty impressive ROI. In addition, all the reported FLOPS for this application were attributed to GPUs, in this case, 7,168 of them. The three-hour simulation modeled the behavior of 110 billion atoms, beating out the previous record for a molecular simulation code, which modeled 49 billion atoms at 369 teraflops. The latter was performed on Roadrunner, the original petaflop super, accelerated by IBM's souped up Cell processors, the PowerXCell 8i.
The 1.87 petaflop performance is quite an achievement for the top-ranked Tianhe-1A, especially considering the current number two system, the CPU-only Jaguar at Oak Ridge Lab, manages just 1.76 petaflops on Linpack, an artificial benchmark designed to show off a system's floating point muscles. In 2008, Jaguar delivered it own sustained petaflop for a real-world application, in this case a superconductor simulation code, which hit 1.35 petaflops*. That work nabbed the application team at Oak Ridge the Gordon Bell Prize that year.
Whether the CAS-IPE team wins any trophies for its molecular simulation application remains to be seen. The researchers will be presenting their work at the upcoming the NVIDIA GPU Technology Conference (GTC) in December in Beijing, and also next May in San Jose, California at the US GTC event.
Over and above the impressive FLOPS is the larger significance of using the technology to propel science and engineering forward. Last year, NVIDIA Tesla GM Andy Keane, penned an opinion piece warning that the lagging adoption of GPU in HPC could threaten the country's competitive edge. While that editorial could easily be construed as self-serving for his employer's interests, the fact is that the US and Europe have lagged countries like China and Japan in adopting this technology for their most elite systems. Those nations saw the revamped graphics chip as the most economical path to petascale machines.
Of course, there are valid reasons to be wary GPU computing for HPC -- programmability difficulties, over-hyping of performance, proprietary software, etc. -- leading many in the HPC community to be extra careful about adopting the technology. But the negative backwash from the original flood of hype can be as ill-informed as the initial exaggerations. In the current issue of HPCwire, Stone Ridge Technology CEO and GPU enthusiast Vincent Natoli, offers a nice set of rebuttals to the major objections to GPU computing. If you're a GPGPU fence-sitter, it's definitely worth a read.
Beyond the significance of GPU usage, the application work demonstrates that the Chinese are not just building these big machines for national prestige. Simulations such as these support basic science research that can be applied to designing and manufacturing better solar energy panels and semiconductor devices. These types of high-tech commercial applications are exactly what the US and other industrialized countries envision as the basis for their future economic growth, and their ability to compete in the global marketplace.
In that sense, even though today's announcement won't appear on the front page of the New York Times, as did the Tianhe-1A TOP500 news, this development is arguably much more significant.
It's also best to see this achievement in the larger context of what the Chinese scientific community is doing. A recent article in Forbes points out that China is quickly catching up to US in scientific output, and in some cases surpassing it:
In 2009, for the first time, Chinese researchers published more papers in information technology than those in the U.S., with both countries churning out more than 100,000 info-tech publications. In clean and alternative energy, Chinese researchers have likewise been publishing up a storm, not surpassing U.S. researchers but coming close.
The bottom line is that the US is in danger of losing its technological edge, which it has basically enjoyed, unchallenged, since the end of World War II. It's not that GPU computing is the magic bullet here. But news like this should be a wake-up call to American HPC'ers and policy-makers that sometimes being extra careful is the riskiest proposition of them all.
*The same superconductor simulation subsequently achieved 1.9 petaflops on the upgraded Jaguar supercomputer.
Posted by Michael Feldman - June 09, 2011 @ 6:01 PM, Pacific Daylight Time
Michael Feldman is the editor of HPCwire.
No Recent Blog Comments
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?