Lawrence Livermore National Laboratory this week took a formal step toward a long promised era of ultra-high-performance computing, signing a $600 million contract with Cray, the Seattle-based computer company, to deliver an “exascale” computer in late 2022.
By late 2023, after a shakedown period, the computer, dubbed El Capitan, is expected to be at least 10 times more powerful than any computer today. Exascale computing means being able to perform at least a billion billion operations per second.
El Capitan should run at 1.5 “exaflops,” although past experience has shown that performance can be improved in certain uses with creative software design.
Late last year, for example, LLNL’s Sierra computer, the world’s third fastest when dedicated, reported speeding up by about one-third to become second fastest a few weeks after it became fully operational.
El Capitan will be a facility of the National Nuclear Security Administration, or NNSA – the semi-autonomous national security arm of the U.S. Department of Energy.
It will operate in classified mode, a powerful resource that serves the needs of weapons researchers at each of the three major NNSA laboratories. These include Los Alamos and Sandia National Laboratories in New Mexico and LLNL in California.
Resource-sharing is a common practice among the three laboratories for both cost and security reasons.
The U.S. nuclear arsenal was designed decades ago and has aged far beyond its expected lifetime.
The last full-scale U.S. nuclear test took place 27 years ago,
forcing weapons researchers to rely on simulations and models run on powerful computers in place of experimental data.
Since the researchers can’t run highly classified calculations on commercial computers or the cloud as academic and industrial scientists often do, NNSA has supported the development of ultra-powerful machines that come as close as possible to simulating realistic experimental conditions.
NNSA’s program for maintaining the weapons safely and reliably is called Stockpile Stewardship.
El Capitan will bring a vitally needed boost in capability to the program, according to LLNL director Bill Goldstein.
“Every component of both warheads and delivery systems must be redesigned and remanufactured to maintain the same capabilities that we had in 1992,” he said.
There is “incredible stress on our computational resources. El Capitan is designed to address that problem.”
He referred to the new capability as “cognitive simulation,” meaning it will combine machine learning, a branch of artificial intelligence, with established methods of modeling and simulation to give El Capitan a greater than tenfold boost in capability.
As others have done, he noted that the U.S. nuclear weapons were originally designed in two dimensions, but age in three.
Computationally, it is much more complex to simulate their slow, 3D decay over the years than it was to model their 2D physics over the millionths of a second that an explosion lasts.
In addition, he said, the interaction of nuclear package physics with delivery system engineering creates a “truly three-dimensional problem.”
“We simply can’t rely on two dimensional computing.”
In her remarks, Lisa Gordon-Hagerty, head of NNSA and a former LLNL employee, said that El Capitan will contribute significantly to the reliability and safety of the U.S. nuclear arsenal.
This in turn “will reassure our friends and allies and serve to maintain strategic stability in an increasingly unstable environment,” she said.
El Capitan will be the first NNSA exascale computer. Two other exascale computers are scheduled to be delivered earlier within the Department of Energy community, one to Argonne National Laboratory and the other to Oak Ridge National Laboratory.
It is also possible that the first exascale computer will not be American. China and Japan have extremely capable computer development programs, and both have declared their intention of dominating the field.