Speaker
Description
We present a performance‑portable lattice gauge theory simulation library implemented using the Kokkos parallel programming model. The library supports efficient Monte Carlo simulations of SU(N) gauge theories across diverse hardware architectures—including CPUs (via OpenMP and Serial backends), NVIDIA GPUs (CUDA), AMD GPUs (HIP), and Intel GPUs (SYCL)—all from a single source code base. It incorporates pseudo‑heatbath algorithms for gauge field updates and provides measurements of key observables such as the plaquette and Polyakov loops. We demonstrate performance portability by recompiling the same source code without a single line of code change, with different Kokkos backends, achieving excellent scaling and computational efficiency on both CPU (x86 and Arm) and GPU (NVIDIA A100 and AMD) platforms. On a 96‑core Arm processor, the OpenMP backend attains a parallel efficiency of up to 57.8%, while on NVIDIA GPUs the performance matches that of a native CUDA implementation. The modular design and the Kokkos abstraction layer ensure that the code remains maintainable and readily extensible to future computing architectures.