Frontier supercomputer runs largest universe simulation yet

Simulating the observable universe isn't just about rendering cool space pictures. It is one of the most brutal math problems in existence.

Historically, astrophysicists had to cut corners. To make the math fit on older machines, models mostly tracked gravity alone, ignoring the chaotic movements of gas, plasma, and expanding space.

Now, Frontier—the world’s second-fastest supercomputer—just delivered the largest and most intricate universe simulation ever recorded.

9,000 nodes tearing through cosmological physics

Frontier, housed at the Oak Ridge National Laboratory, pushed its 1.4 exaFLOPS muscle into full gear for the run, which took place in November 2024.

The run mobilized roughly 9,000 compute nodes, each powered by AMD Instinct MI250X GPUs. The researchers ran the Hardware/Hybrid Accelerated Cosmology Code (HACC), a core component of the ExaSky initiative inside the U.S. Department of Energy's $1.8 billion Exascale Computing Project.

The team spent seven years optimizing code for exascale hardware. The project benchmark called for a 50x speedup over previous milestones, but the system blew past that target, clocking in nearly 300 times faster than similar past hydrodynamic simulations.

Moving past gravity-only simulations

What makes this run special is that it tackles hydrodynamic cosmology.

Instead of simply mapping how gravity pulls dead weight across space, Frontier accounted for dark matter behavior, universe expansion, and the fluid dynamics of hot gas and cosmic plasma all at once.

Frontier has already shown off this raw capability in other fields. In April 2023, it ran the SCREAM model, calculating an entire year of global climate dynamics at an absurd 3-kilometer resolution. It has also been used for molecular-scale materials research and nuclear medicine.

Researchers are now eyeing how this sheer throughput can supercharge large-scale AI models and predictive scientific algorithms.

Why it matters

For astronomers, this simulation changes the game for hypothesis testing.

Instead of waiting years for deep-space telescopes to stumble upon rare cosmological anomalies, scientists can now compare telescope data directly against a dynamic, physics-complete digital twin of the universe.

For the hardware world, this is a massive validation badge for AMD.

While Nvidia dominates the commercial AI spotlight, Frontier proved that thousands of AMD Instinct MI250X accelerators can sustain monstrous scientific workloads without falling apart at scale.

We think the real payoff lies in the convergence of exascale computing and frontier AI. Training the next tier of foundational models and running high-resolution physics simulations require essentially the same muscle. Frontier is proving that exascale infrastructure is no longer just theoretical, though the immense power and capital required means only a handful of institutions on Earth will ever get to play in this sandbox.