GENE-X GPU upgrade targets full fusion plant simulation

Fusion plasma discharge glowing inside the JET tokamak chamber, showing the toroidal confinement geometry used in fusion plasma simulation research

(Image courtesy of UKAEA / EuroFusion)

Researchers at the Max Planck Institute for Plasma Physics are adapting the GENE-X fusion plasma simulation code to run on GPU-accelerated supercomputers, having already recorded performance gains of up to ten times over CPU-only systems in early testing. The work, carried out in collaboration with the Leibniz Supercomputing Centre in Garching, aims to make larger, more realistic whole-plant modelling of fusion reactors possible – a capability the team says is important for commercial reactor design.

Why fusion plasma simulation needs GPU power

Plasma confinement in a fusion reactor is governed by turbulent dynamics that play out across enormous ranges of scale. Capturing those interactions in software demands vast computational resources, and until recently, even the most powerful CPU-based systems have struggled to keep pace with the ambitions of reactor designers.

GENE-X is a gyrokinetic code built specifically for modeling plasma turbulence in magnetic confinement fusion devices. It handles both tokamak and stellarator geometries, making it applicable to the two dominant reactor concepts under active development. The original GENE code was first developed at IPP in 1999 and has been continuously extended since. GENE-X is its most recent development, targeting more complex and comprehensive plasma modeling than earlier versions could deliver.

The team is now porting GENE-X to run on Intel GPU hardware via the SuperMUC-NG Phase 2 system at the Leibniz Supercomputing Centre. The motivation goes beyond raw speed. Researchers want the code to run across all major processor architectures – Intel, NVIDIA, and AMD – to avoid dependence on any single hardware vendor. To achieve this, GENE-X’s original Fortran codebase is being extended with a C++ layer, which could also support future AI-assisted workflows alongside traditional high-performance computing methods.

Analyzing JET data and scaling fusion plasma simulation

The practical test case for the upgraded code is the Joint European Torus, the tokamak that operated in Culham, UK, until 2023. The LRZ release states that JET generated 69 megajoules – around 20 kilowatt-hours – from just 0.2 milligrams of plasma, enough to power an electric car for around 100 kilometres. JET produced more than 100,000 discharges over its operational life, generating a dataset that remains largely unanalysed. Its measurements of hydrogen isotope mixtures are considered particularly valuable for studying plasma turbulence under reactor-relevant conditions.

The LRZ release states the IPP team plans to run GENE-X simulations on three specific JET discharges, generating approximately 50 terabytes of data in total – a volume that illustrates why faster processing matters for this work. Incorporating more physical parameters into each model becomes practical only when the underlying code can handle that data throughput. The release states the resulting data is intended to help train AI models in the medium term, alongside accelerating the optimization of fusion concepts.

The longer-term ambition is more fundamental. The researchers aim to simulate an entire fusion facility – from the reactor walls to the core – rather than modelling sections in isolation. They are targeting power plants with a performance factor at least ten times that of JET. By autumn 2026, the team hopes to run the first GENE-X simulations on Intel GPUs. Growing interest from startups and national programs has added further motivation to make GENE-X widely available for those building the first commercial reactors.

Stay ahead in the fusion revolution explore more breakthroughs from leading innovators in clean energy technology.