This is a clean architecture, with the clear potential to get better for virtualization and running varied workloads. Right now AMD is unique in their ability to integrated CPUs and compute-capable GPUs, and they have used that advantage well.
Lets look at what others are doing.
Nvidia is ATI/AMD's traditional competitor, and has a multi-year lead in language, application, and ecosystem development. OpenCL is a poor second to CUDA, the language. And CUDA has grown to be far broader than a language extension. Nvidia didn't have a CPU to integrate, and thus concentrated development on mitigating the programmer's pain in dealing with two distinct memory spaces. They currently have "unified addressing", which essentially page faults from the other address space, rather than seamless, cache coherent virtual memory. But being stuck on the wrong side of the PCI bus isn't all bad. In exchange for not having cache coherence, the GPU has massive bandwidth to local memory, optimized for its access pattern.
ARM Holdings is the other competitor, albeit indirectly. The Mali 600 series GPUs are compute capable, and are expected to become cache coherent with the upcoming ARMv8 cores. But thus far they have proven to be difficult to evaluate, with only a few quirky implementations (e.g. the first Exynos 5) and OpenCL support available to the public more than a year after the benchmark results were filed.