Multi-GPU systems lead the top 500 with a variety of vendors (AMD/Intel/Nvidia). We will conduct software development and experiments on multi-GPU systems to improve a portable and efficient software stack. Software of interest are - AML is a set of building blocks for portable memory management, including memory allocation, transfers, and layout transformations. - Kaapi/XKBlas - a tasking runtime system and a multi-GPU BLAS implementation built on top of it. - OpenMP - mainly focusing on its `target` interfaces