A team led by Lawrence Livermore National Laboratory runs a project called ElMerFold on the El Capitan supercomputer, mass-producing predicted protein structures as training data for AI models, work that made the group a finalist for the ACM Gordon Bell Prize

2

1

. The prize is described as the most prestigious honor in high-performance computing

2

1

.

The data problem

The reason a supercomputer is needed at all is scarcity. Only a small share of known proteins have experimentally determined structures, and getting them is expensive

2

1

. "There is nowhere near enough real data to train state-of-the-art protein structure prediction models," said Nikoli Dryden, a Livermore computer scientist and the project's lead author

2

1

.

So the field uses distillation: existing prediction models generate synthetic protein structures, and those synthetic structures train the next generation of models

2

1

. The catch is that making these datasets is itself expensive: per the researchers, generating distillation data at the scale modern biomolecular models need can sometimes cost more than training the models

2

1

. The lab says the project was an opportunity to advance open research in protein folding and build new capabilities for large-scale AI research and infrastructure

2

1

.

Squeezing the machine

El Capitan, the National Nuclear Security Administration's first exascale supercomputer, ran the workflow across nearly all of its 45,200 AMD Instinct MI300A APUs, which generated more than 10,700 predicted structures per second

2

1

. Against an extrapolated baseline, that is a 71.7-fold speedup, and the team estimates a dataset of roughly 41 million protein structures would take two and a half hours instead of eight days

2

1

. One caveat sits under the headline figure: the 71.7-fold speedup is measured against an extrapolated baseline, and the eight-day comparison is the team's estimate for a dataset of about 41 million structures

2

1

.

Getting there took aggressive engineering. The MI300A chips combine a CPU and GPU with shared high-bandwidth memory. "We went as far as writing machine-level instructions to maximize the performance of the system," said Livermore computer scientist Tal Ben-Nun

2

1

. Throughout, the team worked with protein folding experts at Columbia University to make sure the speed did not come at the cost of prediction accuracy

2

1

.

What the data buys

The dataset has been made publicly available and feeds OpenFold3, an open-source biomolecular structure prediction model

2

1

. Columbia collaborator Mohammed AlQuraishi said the distillation set broadly improves OpenFold3's performance and matters most for antibody-antigen complexes, which lack a strong evolutionary signal and need stronger physical understanding to predict correctly

2

1

. The team plans to generate more diverse datasets, including multi-molecule protein complexes

2

1

.

Source: Phys.org

Source: Phys.org

The same approach points beyond biology. "A majority of the problems in our national security mission have very few interesting data samples, and many of the programmatic needs require exploring or interpreting rare or anomalous events," said the lab's Brian Van Essen

2

1

. He said the approach offers a potential fix for data-scarce problems in stockpile modernization and other National Nuclear Security Administration applications

2

1

.

The Gordon Bell decision

Whether the performance work earns the field's top honor is still open. The Gordon Bell winner will be announced Nov. 19 at the SC26 conference in Chicago

2

1

.

© 2026 TheOutpost.AI All rights reserved