We think the design of the language will support that in the future, but we need to write a GPU-based runtime that lives inside of the runtime and get it to decide when to offload work to GPU first, and that will take time.