The University’s High-Performance Computing Research Center (HPCRC) has acquired a new supercomputer, named Traverse, which will aid research at the University’s Plasma Physics Laboratory (PPPL), as well as other University programs.
The addition joins six other computing clusters: Tiger, Dell, and Perseus, which are the largest and reserved primarily for faculty, as well as Nobel, Adroit, and Tigressdata, which are available to students. All the clusters are housed in a building on the Forrestal campus, about three miles from the main campus.
Supercomputers require high amounts of energy, and HPCRC typically uses 1.8 megawatts of electricity and is equipped with backup generators. The clusters can also overheat, which requires ventilating them with cooled air. The facility is efficient enough to have earned a LEED Gold rating.
Thanos Panagiotopoulos, the chair of the chemical and biological engineering department, said that Traverse will allow Princeton’s Chemistry in Solution and at Interfaces (CSI) lab to model the interactions of a few hundred molecules at a time.
“We do problems involving very large-scale calculations that connect quantum mechanics with … the collective properties of water and aqueous solutions,” Panagiotopoulos said. The simulations usually last only on the order of a few picoseconds but can help CSI understand the atomistic dynamics of various materials.
Roberto Car, director of CSI and the Ralph W. *31 Dornte Professor in Chemistry at the University, said that his group of researchers now uses a new, more efficient mathematical construction, called a deep neural network, which uses machine learning to compute the classical mechanics forces in any number of arrangements that share the same statistical probability. Researchers derive the interaction potentials from density functional theory, which considers the quantum mechanics of the atoms in their ground states.
“Having access to that kind of machine at Princeton will allow us to do this work on our code and experiment with the capabilities offered by this architecture,” Car said.
Traverse has a similar architectural structure to Summit, the most powerful supercomputer in the world, housed at Oak Ridge National Laboratory. Traverse is a 1.4-petaflop system, making it capable of 1.4 million billion floating-point calculations per second. It is on the TOP500 list, a ranking of the 500 most powerful supercomputers based on standard tests.
Panagiotopoulos and Car noted that Traverse will soon be overtaken by more powerful supercomputers. Car predicted that exascale systems, which would be capable of a billion billion calculations per second and function 1,000 times faster than petascale ones, will be built in the next few years. He noted that PPPL will likely be able to use technology developed at Oak Ridge.
What sets Traverse apart from the previous HPCRC clusters is its architecture — described by Car as “a hybrid architecture that consists of CPU [central processing unit] and GPUs [graphics processing units].” The clusters were built by IBM, and the GPUs were supplied by Nvidia, which sells GPUs for many personal computers and gaming systems.
Car said the first exascale supercomputers will share a similar architecture to Traverse, meaning that the work required to adapt the researchers’ current algorithms to Traverse will remain useful.
Traverse will help PPPL model the movement of plasma in its tokamak NSTX-U, the largest of its kind in the world, to better understand how to control the plasma on a millisecond timescale. PPPL was founded in 1951 and has been working, among other projects, to create a viable fusion reactor potentially capable of generating virtually unlimited energy.
Traverse was financed by the University, and it will be used by graduate students, postdoctoral researchers, and faculty at the University, as well as PPPL, which is managed by the Department of Energy.