Supercomputer designers spent years chasing after the elusive petaflop barrier – the ability to perform one thousand trillion calculations per second.
IBM and Los Alamos National Laboratory finally achieved the milestone last summer, but a far more ambitious project is already underway: the building of a 20-petaflop machine.
“Moore’s Law says you get a doubling of performance every 18 months. We’ll do ten times that in 24 months,” boasts David Turek, IBM’s vice president of deep computing.
This time, IBM is teaming up with Lawrence Livermore National Laboratory in California, where a computer nicknamed Sequoia will be installed with 1.6 million processor cores across a relatively small physical footprint of 3,422 square feet. Lawrence Livermore needs incredible processing power in order to maintain the safety and effectiveness of the nation’s nuclear weapons stockpile, explains Mark Seager, assistant department head for advanced technology at the Lab.
“As weapons age … sometimes problems are found and then the question becomes ‘is this a big deal or not?” Seager says, noting that high explosives decay over time. “The way we answer this question is through simulation and experimental science.”
Livermore already has the fourth-fastest supercomputer in the world, an IBM Blue Gene/L system that goes up to 478 teraflops, or about half of one petaflop. The lab also has a 100-teraflop machine named ASCI Purple. Both machines were the fastest in the world when they were first made.
Sequoia is still being developed by IBM. The plan is to start delivering the system to Livermore in 2011 and have it fully deployed by 2012. In the meantime, IBM is supplying a 500-teraflop machine called Dawn in the next month or two, which will allow Livermore to develop the new physics models and algorithms that will be used on Sequoia.
Harnessing the 1.6 million processing cores on Sequoia will be a huge challenge.
“It’s developing new physics algorithms and new engineering algorithms,” Seager says. “To do that we need to do a lot of simulations and experiments and compare the results and stuff. Dawn is a vehicle that’s large enough where we can do a lot of these studies.”
When Sequoia arrives, Lawrence Livermore will be able to perform calculations anywhere from 12 to 50 times faster than it does with its current machines. In one instance, the lab used the ASCI Purple to perform 4,000 calculations related to one weapon system, and it took a whole month.
Sequoia’s 1.6 million IBM Power processor cores, 1.6 petabytes of memory and 98,304 compute nodes will be housed in 96 refrigerator-sized racks. Despite its massive power, the computer will use power 17 times more efficiently than Blue Gene/L, according to IBM.
The machine will be funded by the National Nuclear Security Administration for an as-yet undisclosed price. IBM is still designing the next-generation processors to be used in the supercomputer.
“You’re going to get a faster core, more cores on a chip, and more cores in total,” Turek says. “We’re also continuing advances in terms of the networking we put in the system to facilitate the communication among these different cores.”
Sequoia could potentially improve scientists’ ability to monitor and forecast weather by a factor of 40, allowing forecasters to “predict local weather events that affect areas from 100 meters to one kilometer in size, down from their current ten-kilometer ability,” according to IBM.
Moreover, “a 20-petaflop machine would deliver a 50x improvement in our capability to predict earthquakes and map out safe evacuation routes,” IBM states. “This would allow scientists to predict an earthquake’s effects on a building-by-building basis across an area as large as Los Angeles County.”