SYCL++: A Unified Programming Framework for Heterogeneous Supercomputers at Scale

With the rise of heterogeneous HPC systems, traditional scientific applications face increasingly severe challenges. Today’s supercomputers adopt vastly different designs in compute units, memory hierarchies, and interconnects, making application porting an expensive task that demands significant expert effort. This deep coupling results in fragmented, unmaintainable codebases. In particular, it creates substantial barriers for domain scientists, hindering algorithmic innovation across such diverse architectures. Consequently, there is an urgent need for a unified compiler and programming framework that enables true single-source performance portability across this diverse landscape. However, achieving this unification is challenging due to a central contradiction: while diverse architectures often converge on high-level abstractions, their low-level architectures frequently diverge in parallel execution models, memory management mechanisms, and network topologies. To address this, we propose SYCL++1, a unified programming framework that decouples the expression of computational semantics from platform-specific optimization strategies. SYCL++ introduces declarative primitives that allow a single, clean codebase to be efficiently mapped to different HPC systems. We validate our approach using a real-world CGFDM3D seismic simulation application. Our evaluation demonstrates that SYCL++ achieves performance portability across three distinct platforms, delivering up to 2.2 × speedups over baseline SYCL with minimal code divergence. The framework exhibits near linear scaling to 1,597,440 Sunway cores and 8,192 DCUs, enabling high-resolution simulation of realistic earthquake scenarios. We demonstrate this capability by modeling the March 2025 Myanmar earthquake at 125 m resolution. To our knowledge, this is among the first demonstrations of a unified HPC application achieving performance portability across three distinct leading supercomputers.