Canada's Trillium supercomputer cracks a 50-year stellar mystery months after going online

A 50-year-old question in astrophysics now has an answer, and the story behind it is as much about compute infrastructure as it is about stars.

In a paper published in Nature Astronomy, researchers from the University of Victoria, the University of Minnesota, and the Texas Advanced Computing Center (TACC) used 3D hydrodynamic simulations to show that stellar rotation amplifies internal mixing in red giant stars by more than 100x. That result explains chemical signatures astronomers have observed since the 1970s but couldn't account for.

The compute that made it possible? Canada's Trillium supercomputer, a 34-petaflop system at SciNet (University of Toronto) that launched in August 2025. The simulations ran within months of the machine going live.

"We were able to discover a new stellar mixing process only because of the immense computing power of the new Trillium machine," said Falk Herwig, professor of physics and astronomy at the University of Victoria and director of the university's Astronomy Research Centre. "These are the computationally most intensive stellar convection and internal gravity wave simulations performed to date."

The problem nobody could simulate

Since the 1970s, observers have tracked changes in the surface chemistry of red giant branch (RGB) stars. Carbon-12 to carbon-13 ratios decline in ways that indicate material from deep nuclear-burning regions is somehow reaching the stellar surface. The trouble is that a stable radiative barrier sits between the outer convective envelope and the inner hydrogen-burning shell. Material shouldn't cross it.

Several mechanisms have been proposed over the decades. Thermohaline mixing, driven by molecular weight inversions, was the leading candidate for years, but 3D simulations showed it was too slow on its own. Magnetic buoyancy, meridional circulation, and internal gravity waves each had proponents. None could quantitatively reproduce the observed chemistry.

Rotation was long suspected as a factor, but testing the hypothesis required 3D simulations at resolutions that simply weren't possible until now.

"Until recently, while stellar rotation was thought to be part of solving this conundrum, limited computing abilities prevented us from quantitatively testing the hypothesis," Herwig said. "These simulations allow us to tease out small effects to determine what actually happens, helping us to understand our observations."

What Trillium brings to the table

Trillium replaced SciNet's previous Niagara and Mist clusters. The hardware details matter here:

• CPU cluster: 1,224 nodes, each with 196 AMD Turin cores at 2.6 GHz, totaling 235,008 CPU cores
• GPU cluster: 63 nodes with 96 AMD Genoa cores and 4x NVIDIA H100 GPUs (80 GB each), for 252 GPUs total
• Memory: 768 GB RAM per node
• Interconnect: NDR InfiniBand at 400 Gb/s (CPU) and 800 Gb/s (GPU), 1:1 non-blocking
• Storage: 29 PB of VAST NVMe across home, scratch, and project tiers
• Peak performance: Roughly 34 petaflops
• Power draw: About 1,700 kW

The system received more than $52 million in federal funding through the Digital Research Alliance of Canada. A Nature Astronomy paper within the system's first operational months is a concrete return on that money.

The team also used resources at TACC, which operates NSF-funded leadership-class systems including Frontera and Stampede3.

The software side: PPMstar

Hardware alone doesn't produce science. The simulation code behind this work, PPMstar, has been developed over many years by Paul Woodward at the University of Minnesota and Herwig at UVic. It's an explicit, compressible gas-dynamics solver that tracks two fluids while accounting for radiation pressure and radiative diffusion.

PPMstar has powered a progressive series of papers with increasing physical complexity (2015, 2023, 2024, 2025). The 2023 precursor study by lead author Simon Blouin established the simulation framework, showing that internal gravity waves alone moved too little material to explain observations. This latest work added rotation to the picture, and the results were dramatic: mixing rates jumped by more than two orders of magnitude.

That kind of sustained software investment, spanning institutions and decades, is easy to overlook when a headline says "supercomputer solves mystery." But PPMstar is what converts Trillium's raw FLOPS into astrophysical insight. Without either half, this paper doesn't exist.

Beyond stellar physics

The computational fluid dynamics methods used here aren't limited to the insides of stars. The same physics governs ocean currents, atmospheric circulation, and even blood flow. Herwig is actively collaborating across these domains to build shared simulation tools, which means HPC investment in astrophysics can pay dividends well outside the field.

Blouin, a CITA National Fellow at UVic who previously held a Director's Fellow position at Los Alamos National Laboratory, plans to extend the work to other stellar types and rotation profiles. Those future simulations will need even more compute.

The infrastructure argument

Every outlet covering this story has focused on the astronomy. Fair enough: solving a 50-year puzzle about how stars work is genuinely interesting. But there's a separate story here about infrastructure planning and public investment.

Canada funded a 34-petaflop supercomputer. Within months, it produced results that were published in one of the field's top journals. The researchers were explicit that the work was impossible before Trillium existed. This isn't a case where bigger hardware made existing calculations faster. It made them feasible for the first time.

For anyone building the case for national HPC investment, that's a data point worth remembering.