The Week in Review
Here is a collection of highlights from this week’s news stream as reported by HPCwire.
Welcome to the FutureGrid
Indiana University has been tasked with creating an experimental supercomputing network called FutureGrid. The IU researchers and nine national and international partners will focus on developing software and techniques to link together supercomputers in order to solve massive-scale scientific problems, such as climate system modeling and DNA analysis. Of the $15 million in funding, $10.1 million comes from the NSF and the remainder from partners. The linked systems will serve as a gigantic testbed:
“One way of looking at the FutureGrid is to think of it as an ‘experiment factory’ in which supercomputers can be interconnected in a myriad of different ways to find out which connections and software combinations work together and which do not,” said Brad Wheeler, IU vice president for information technology and CIO. “The ultimate goal is to create a system that researchers can use to most effectively match cyberinfrastructure and scientific needs in ways that help us make new discoveries most effectively.”
Geoffrey C. Fox, picked to lead the project, explained it thusly:
“FutureGrid will serve as a proving ground for new, distributed computing systems and will open up exciting new avenues for scientific collaboration and research. We envision the grids and clouds of the future not as a single system, but as many linked systems. For this reason we are engaging an incredible set of academic and commercial partners throughout the U.S. and in Europe to participate in FutureGrid.”
FutureGrid is expected to be operational next spring.
May the Best Chip Win
David Kanter at Real World Technologies took a stab at comparing the computational efficiencies of the current crop of CPUs and GPUs (and Cell processor). He does a nice job of outlining the fallacies of looking at peak performance metrics, but settles on looking at peak performance per watt and peak performance per die area (square mm). The latter metric seems a bit strange, but Kanter managed to come up with some pretty interesting observations regardless.
His basic conclusion: The differences between computation efficiency in GPUs and CPUs are smaller than most people might think. In particular, from an energy efficiency standpoint, low-power chips like Intel’s Atom are as good or better than top-end GPUs from NVIDIA and AMD, and even the server chips like Nehalem and Istanbul hold their own against their graphics brethren. The main weakness of the analysis is that it’s based on double precision (DP) floating point performance, a relatively new and underpowered feature in GPUs. If Kanter had used single precision, the results would have looked very different.
Here’s the sound bite conclusion from the article:
The bottom line of this rough analysis is that the gap between CPUs and GPUs isn’t quite as big as some have claimed, considering the power and die area. When taking these factors into account, GPUs seem to have a clear performance/mm2 advantage. However, performance/watt is more important and in that particular metric, CPUs can come much closer (at the cost of single threaded performance). GPUs are still very effective for certain workloads and clearly hold many advantages in terms of raw performance and bandwidth, but these advantages are not necessarily unassailable.
He goes on to say that processor architectures are undergoing rapid evolution (especially with regard to DP capability in GPUs) so this snapshot may not be worth much a year from now. Useful analysis nonetheless.