To accurately simulate large biological systems, scientists need models that capture how electrons behave, but such models are too slow for millions of atoms. Faster models miss important chemical changes like bond breaking.
The researchers built Tessera, a system that runs an accurate machine learning model across many GPUs. It distributes atoms by sorting them along a curve so each GPU gets the same number, and it reuses calculations across simulation steps to save time.
On 512 GPUs, Tessera simulated 30.7 million atoms with 89.7% efficiency, meaning the GPUs are almost as productive as if each worked on a small problem. This lets researchers study large structures like viruses with accurate physics, offering new insights into biology.