MultiGPU Performance-Portable Solution for Array Programming Based on Kokkos
Explore a cutting-edge solution for array programming on multiple GPUs in this conference talk from ARRAY 2023. Discover how researchers from Oak Ridge National Laboratory have developed a new application programming interface based on the Kokkos programming model to enable transparent and portable array computation across NVIDIA and AMD GPUs. Learn about the implementation of different techniques for exchanging stencils among GPU memory spaces and the use of autotuning to optimize GPU utilization. Gain insights into the performance evaluation conducted on Summit and Crusher supercomputers, comparing the new Kokkos solution against the traditional MPI + Kokkos approach. Understand the benefits of this innovative solution in terms of scalability, performance, and programming productivity for high-performance computing and data center applications.