The TOP500 Supercomputer list is out and once again, IBM has bragging rights. Big Blue not only tops the list, it pretty much owns it, with 232 of the 500 computers on the list.
The TOP500 list is compiled and published twice per year, in June and November, by supercomputing experts in the U.S. and Germany. For the last few years, IBM’s monster cluster Blue Gene/L, used by the Department of Energy at the Lawrence Livermore National Laboratory, has topped the list.
And for now, it’s staying there. Thanks to an expansion that involved adding 40 more racks to Blue Gene/L’s 64 racks, the monstrous system retains its leadership position with a peak performance of 600 teraflops and sustained performance of 478 teraflops. A teraflop is a trillion operations in one second.
The peak, listed as Rpeak in the list, is the theoretical limit on how fast the computer can go under benchmarks. The sustained performance, listed as Rmax, is how fast it has gone in real-world use.
The No. 2 computer is a new entry, the fastest in Europe, and also an IBM system, Blue Gene/P. This new system clocks in at 167 teraflops. SGI, also a new entrant, has the No. 3 spot. The Xeon-based supercomputer clocks in at 127 teraflops. No. 4 is an Indian computer system, a Xeon-based HP cluster, with 117 teraflops.
Four of the top 10 machines are IBM Blue Gene machines, and they don’t run IBM’s fastest chips, the POWER5 and POWER6. Instead, they use the older, slower PowerPC 440 chip. IBM did this because it needed a cooler chip to achieve the density required for Blue Gene/L, which contains 213,000 processors.
“It was becoming more obvious you couldn’t keep putting together faster clock speed processors because the heat generated is harder to dissipate and it takes more power,” said Herb Schultz, deep computing marketing manager at IBM. “The logic was if we want to get to these high levels, you really had to take a fork in the road. And instead of faster processors you got slower processors more densely packed.”
The Blue Gene/P supercomputer in Germany, for example, has 4,096 processors in each rack, 10 times what is normally found in a blade. “If you want to scale to these really high numbers, it’s really hard to do,” said Schultz. “The limiting factor becomes can you power and cool it and do you have space to cool it.”
IBM does have a plan to move forward to POWER5 and POWER6-based supercomputers, and it also has a new ace up its sleeve, called “Roadrunner.” This hybrid design will combine the AMD Opteron processor with the Cell Broadband Engine, the graphics processor in the Sony PlayStation 3.
Roadrunner, also a DoE project, will be delivered to Los Alamos National Laboratory in the summer of 2008 and will be capable of exceeding a petaflop
Much of what helps IBM achieve this performance is the software written for the systems. The code at Lawrence Livermore was written to promise a near one-to-one gain, so doubling the number of processors would result in doubling the performance.
“Blue Gene was designed in collaboration with Lawrence Livermore National Labs. They had the apps and knew how they wanted them to run. If you’re going to get to multiple teraflops, you have to optimize right,” he said. “With careful thinking of the apps and a balanced design of the hardware, if you do them right, you can put new hardware on and keep getting linear scaling.”
