Accelerating the Particle-In-Cell code ECsim with OpenACC
The GPU-optimized code runs 5x faster and cuts energy use by 3x on pre-exascale systems.
A collaborative research team has successfully supercharged a critical tool for plasma physics. By applying OpenACC pragmas to the ECsim code—a semi-implicit, energy-conserving Particle-In-Cell (PIC) simulation—they have prepared it for next-generation exascale supercomputers with minimal code restructuring. The results are striking: on the pre-exascale Leonardo system, the GPU-accelerated code delivers a 5x performance speedup while simultaneously slashing energy consumption by a factor of three compared to the original CPU-based version.
The performance gains extend across hardware generations, with tests showing substantial benefits from NVIDIA's latest GH200 unified memory architecture. The code also demonstrates impressive scalability, maintaining 70% efficiency in strong scaling tests up to 64 GPUs and 78% efficiency in weak scaling tests up to 1024 GPUs. This work, detailed in a new arXiv preprint, represents a major step toward efficient, exascale-ready kinetic plasma modeling, enabling previously impractical high-resolution simulations of phenomena like fusion plasma confinement and astrophysical shocks.
- Achieved a 5x speedup and 3x lower energy use on GPU vs. CPU reference code.
- Demonstrated high scaling efficiency of 70-78% on up to 1024 GPUs on the Leonardo system.
- Leveraged OpenACC for minimal code changes, showcasing a portable path to exascale computing.
Why It Matters
Enables faster, greener, and more detailed simulations of fusion energy and space plasma, accelerating critical scientific discovery.