NERSC Supports 2024 Gordon Bell Prize Winner

The flagship Perlmutter supercomputer at the National Energy Research Scientific Computing Center (NERSC) played a role in the groundbreaking research selected for the 2024 Association for Computing Machinery (ACM) Gordon Bell Prize, awarded annually to recognize outstanding achievement in high-performance computing. The prize was awarded November 21 at the Conference for High Performance Computing, Networking, Storage, and Analysis (SC24).

In the paper, “Breaking the Million-Electron and 1 EFLOP/s Barriers: Biomolecular-Scale Ab Initio Molecular Dynamics Using MP2 Potentials,” the researchers, based at Australian and American institutions, addressed a key problem in the simulation of complex biological phenomena: quantum mechanical methods offer accurate simulations but are too computationally expensive to present a practical solution for large systems. The researchers’ new approach combines two existing methods in a way that offers high accuracy as well as the ability to scale to biological systems with thousands of atoms. Combined with modern GPU architecture, the method enabled them to break the million-electron and 1 EFLOP-per-second barriers for ab-initio molecular dynamics (AIMD) simulations with quantum accuracy.

AIMD is a form of molecular dynamics that uses force fields derived at each step from ab-initio, or “first principles,” calculations—meaning they’re derived directly from the laws of physics. In this work, the team performed first-principles quantum chemistry simulations using the highly accurate MP2 quantum chemistry method, which is known to address important many-body interactions. Breaking the one-million-electron and one-EFLOP barrier is an important milestone because it shows that exascale HPC will allow scientists to tackle very complex systems at real-world scales that have previously been the realm of less accurate models.

In addition to the Frontier system at Oak Ridge Leadership Computing Facility (OLCF), the team tested their algorithms on 1,024 nodes (4,096 A100 GPUs) on Perlmutter and achieved a record of 3.4 seconds per timestep while simulating a protein fragment with 1,496 atoms and over 5,500 electrons. This translates to a throughput of 25,000 timesteps per day—three orders of magnitude faster than any calculation at this level of theory for a biosystem of this dimension.

“Perlmutter played a critical role in scaling our ab initio molecular dynamics (AIMD) simulations,” said PI Giuseppe Barca, “particularly for testing and refining the new algorithmic innovations that ultimately led to our record-breaking results.” According to Barca, these calculations were essential for testing their methods across a variety of molecular systems and validating their scalability before extending them to larger runs on Frontier.

“NERSC’s resources were indispensable for prototyping and troubleshooting during the development phase,” said Barca. “Access to Perlmutter enabled us to iterate quickly on our molecular fragmentation and perturbation theory techniques, providing the performance insights necessary for full deployment on exascale platforms. We’re deeply grateful for the support from NERSC, whose infrastructure was instrumental in achieving this milestone in computational chemistry.”

NERSC is a DOE Office of Science user facilities based at Lawrence Berkeley National Laboratory.