Your descprition is short enough to give a complete picture without taking too long time.
Let me give one remark. (Being Hungarian) I found it interesting that you brought Von Neumann into the picture. The Neumann architecture has nothing to do with the asymmetry inside a heterogenous environment. The major difference between the originally two competing architectures, the Harvard and Neumann architectures is that while in the latter data and application binary reside in the same memory (RAM), the former does not allow this. Neumann architecture makes room for overflow attacks and the likes, where function pointers may accidently point to ordinary data that can infact be interpreted as binary, and boom: the two notions do in reality mix. Harvard architectures would physically prohibit this, as application binary and data would not reside in the same memory space. This is completely a different topic as to how accelerators use an alternative memory address space and how their execution is subordinated to the CPU.
Other than this, your summary is fine.
