Plate Nº 56 · recorded November 18, 2026
Biology & EvolutionReported finding
El Capitan Supercomputer Cranks Out 10,700 Protein Structures Per Second
LLNL's ElMerFold workflow generated 41 million protein structures in 2.5 hours on El Capitan, a 71.7-fold speedup that earned a Gordon Bell Prize finalist nod.
By Nathan Brooks4 min read752 words
In brief
- ElMerFold generated over 10,700 predicted protein structures per second on 45,200 AMD MI300A APUs.
- The workflow cut generation of a ~41 million-structure data set from eight days to 2.5 hours, a 71.7-fold speedup.
- The team is a finalist for the ACM Gordon Bell Prize; the winner is announced November 19, 2026 at SC26 in Chicago.
- The data set trains OpenFold3, an open-source protein structure prediction model, and is publicly available.
- The ElMerFold data set is publicly available to the broader research community.

A Lawrence Livermore National Laboratory-led team has squeezed more than 10,700 predicted protein structures per second out of the El Capitan supercomputer. The optimized workflow, called ElMerFold, cut the time to generate roughly 41 million structures from an estimated eight days down to just two and a half hours — a 71.7-fold speedup.
The achievement made the team a finalist for the Association for Computing Machinery's Gordon Bell Prize, one of computing's top honors. The winner will be announced November 19 at the SC26 conference in Chicago.
The team ran ElMerFold across 45,200 AMD Instinct MI300A accelerated processing units — chips that combine a CPU and GPU with shared memory. Partners included researchers from AMD and Columbia University. El Capitan is the National Nuclear Security Administration's first exascale machine, meaning it can perform at least a quintillion calculations per second.
Why does anyone need millions of protein structures?
Proteins carry out essential jobs in nearly every biological process, and their 3D shapes determine how they work and what they bind to. Accurate structure predictions matter for understanding disease and designing new medicines.
The catch is data. Experimental techniques for determining structures are expensive and slow, and only a small fraction of known proteins have experimentally solved shapes.
"There is nowhere near enough real data to train state-of-the-art protein structure prediction models," said LLNL computer scientist and lead author Nikoli Dryden. "It takes a lot of time, expense and effort to get new experimental structures, so we supplement that with synthetic data."
ElMerFold tackles the shortage through distillation: existing prediction models generate synthetic structures used to train new models. At the scale modern biomolecular models demand, producing that data can cost more computation than training the models themselves.
What does the data set improve?
The structures feed OpenFold3, an open-source biomolecular structure prediction model. The distillation set broadly improves OpenFold3's performance, with a notable effect on antibody-antigen complexes — the interactions between immune proteins and the targets they recognize.
"The distillation set broadly improves OpenFold3 performance because it enhances its understanding of the biophysics of protein structure," said Mohammed AlQuraishi, a Columbia assistant professor of systems biology and computer science and a project collaborator. "It is particularly important for antibody-antigen complexes, because they typically do not have a strong evolutionary signal, requiring stronger physical understanding."
That matters because understanding how antibodies bind to targets is central to developing therapeutics against infectious diseases. Predicting how drugs and proteins interact could also speed up early-stage drug discovery.
How did they make it so fast?
The performance took deep engineering on both software and hardware. Key moves included:
- Scaling across nearly the entire El Capitan system
- Writing specialized computational kernels and improved memory management
- New scheduling techniques to keep processors busy
- Persistent inference servers that avoided reloading prediction models
- Rapid data access via El Capitan's Rabbit near-node storage system from HPE
- Automatic fault recovery, letting jobs survive individual node failures across thousands of nodes
"We went as far as writing machine-level instructions to maximize the performance of the system," said LLNL computer scientist Tal Ben-Nun. "A lot of what we learned will make its way back into other applications and programs at the lab."
The team worked closely with Columbia's protein folding experts throughout to make sure speed gains didn't hurt prediction accuracy. Some techniques are being contributed to compiler frameworks such as DaCe, which could benefit other workloads on the MI300A architecture.
"Working with LLNL and Columbia, we optimized ElMerFold to harness El Capitan's AMD Instinct MI300A architecture and AMD ROCm software to produce larger, richer data sets for protein structure prediction," said Nicholas Malaya, co-author and senior fellow for HPC and sovereign AI at AMD.
What's next?
The ElMerFold data set is publicly available to the broader research community. Next steps include generating more diverse training data, predicting protein complexes made of multiple interacting molecules, and folding the latest OpenFold3 model into the workflow.
For principal investigator Brian Van Essen, the project showed how exascale systems can accelerate the creation of training data for scientific models across materials science, chemistry and physics. It also addresses a national security problem: scarce data on rare or anomalous events that limits AI tools in NNSA mission work.
"Our goal is to refine the core scalable inference framework so that it can be more easily adapted to other complex scientific SI inference workloads," Van Essen said.
via phys.org/rss-feed/ (Source)
More from Nathan Brooks
Show full bio
Market editor covering companies, deals and strategy in general science news at SciBeat.
456 articles
Nearby plates
- Oxford Team Watches Virus-Like Particles Build Themselves in Real Time
- Google DeepMind Embeds Invisible Watermarks in AI-Designed Proteins
- Berkeley Lands Schmidt Grant to Build AI Platform on Doudna Supercomputer
- New 'Computational Microscope' Simulates DNA Packaging at 10x Previous Scale
- Could Dead Planets Fake a Biosignature? Complexity Test Holds Up