GPU Acceleration of Plasma and Physics Codes
Graphics processors run thousands of arithmetic operations at once, matching the structure of the particle and grid math in fusion codes.
Why GPUs fit this work
A GPU has thousands of simple cores optimized for doing the same operation on many pieces of data at once. Much of fusion computation, updating millions of particles or grid cells with the same equations, has exactly that shape. When the work maps well, GPUs deliver large speedups over CPUs for the same power and cost of hardware.
What accelerates well
- Particle pushes in plasma kinetic codes.
- Stencil updates on structured grids for fields and fluids.
- Monte Carlo neutron histories, which are naturally independent.
- Dense linear algebra inside solvers and surrogate training.
What is hard
GPUs are fast only when data movement is managed carefully. Irregular memory access, branching, and frequent host-device transfers erode the gains. Getting real speedups means restructuring algorithms around how the GPU reads memory, not just recompiling existing code.
In the Kronos stack
GPU acceleration underpins the HPC simulations for both the Hyperion breeder and the burner, and it is what makes generating large surrogate training sets and running ML-accelerated studies practical on a realistic timeline.
Precision and verification
GPU arithmetic can differ from CPU arithmetic in the last digits, so accelerated codes are checked against reference results to confirm the speedup did not change the physics, part of automated verification.
Payoff
Faster kernels mean more of the configuration space covered per unit of time, which is the point of acceleration.