We'll present a concept for managed GPU communication that offers GPU-based traffic sourcing and sinking. Vendor support for communication among multiple GPUs is still limited to relying on CPUs for the orchestration of kernel launches and data movements. Instead, we observe improvements of unified address spaces for multi-GPU systems, enabled either by NVLink or PCIe. Instead of exposing memory operations on non-uniform resources to the user, we propose to use managed communication schemes that support the user with tasks such as buffer allocation, data movement, and connection management. Our concept for GPU-managed communication is based on building blocks like ordered queuing, event notification, and support for matching semantics, and therefore composable for different communication models. Besides introducing our concept and the building blocks, we'll also present an initial communication API that demonstrates the value of this concept.