Fusion Power Just Got Closer: New Software Runs on 10,000 GPUs

Fusion is the reaction that powers the sun, and if we could bottle it on Earth we would have clean energy with no long-lived radioactive waste. The leading design is the tokamak: a donut-shaped magnetic bottle that holds plasma (electrically charged gas) hotter than the sun's core. Building one costs billions, so scientists lean heavily on computer simulations to test ideas before committing to steel and concrete. Better simulations mean fewer expensive dead ends.

That is where TokaGLINT comes in. It is new software that splits the heavy math of electromagnetic fields — the magnetic forces that hold the plasma in place — across thousands of graphics chips (GPUs), the same kind of hardware that powers AI chatbots and video games. On a supercomputer, it ran on more than 10,000 GPUs at once and kept about 90% efficiency, meaning doubling the machines nearly doubled the speed. A single machine also ran 2.67 times faster than the standard method.

Why is that hard? The plasma inside a reactor is constantly shifting, and the magnetic fields have to be recalculated again and again over long stretches of time. Older approaches tend to drift and fall apart during long runs. This solver stays stable, which is what makes realistic, hours-long simulations possible. The team tested it against EAST, a real experimental fusion reactor in China, using a plasma simulation code called SymPIC.

The catch: this is research software, not a product, and it does not build a working power plant. Fusion electricity is still likely years — probably decades — away, and this tool only makes the design phase faster and cheaper. Still, it is a clear sign that the GPU boom driving AI is spilling into science, and that the tools for chasing fusion are getting sharper.