Workshop: Thirteenth Workshop on Accelerator Programming and Directives (WACCPD 2026)
Authors: Pedro Valero-Lara (Oak Ridge National Laboratory); Doris Chiang (University of California, Berkeley); and Philip Fackler, William Godoy, Keita Teranishi, and Jeffrey Vetter (Oak Ridge National Laboratory)
Abstract: We present JACC.Async, a new feature of Julia for ACCelerators (JACC), the performance-portable metaprogramming framework built on the just-in-time (JIT) and LLVM compiled Julia language. JACC.Async enables scientific applications to exploit asynchronous executions across multiple GPUs through a high-level, portable, and easy-to-use interface. Despite the fact that asynchronous multi-GPU execution is essential for achieving high performance on modern systems, existing solutions often rely on low-level, vendor-specific programming models and complex synchronization mechanisms. JACC.Async addresses this reliance by providing the first JIT-based solution for asynchronous multi-GPU computing, offering a unified abstraction that overlaps computation, communication, and data movement across diverse accelerator architectures. JACC.Async supports the two major GPU vendors (NVIDIA and AMD). Scientific workloads based on the conjugate gradient algorithm were used to evaluate JACC.Async on modern multi-GPU platforms featuring NVIDIA A100 GPUs and AMD MI250X GPUs. The results demonstrate that JACC.Async achieves scalable performance with minimal modifications to existing applications.
Back to Thirteenth Workshop on Accelerator Programming and Directives (WACCPD 2026) Archive Listing
Back to Full Workshop Archive Listing