I'm sure they do, as a lot of their low-range cards seem to use system
memory to increase their video memory size. I think Nvidia calls it
TurboCache, and AMD calls it HyperMemory.
Yousuf Khan
The system memory is used for buffer object storage. But that aside,
the problem isn't transfering the data but the resolve needed to
acquire the correct data to transfer. If the client cannot return
until the data transfrer is complete, this can consume considerable
amount of CPU time polling for the result. If this is done by sending
a request with non-blocking query, it's doable, but there isn't public
API for that.
Hardware can do what the OP is asking, but it needs someone to want to
write the software support.. FWIW, the compute shaders in coming APIs
and hardware might have something useful for this kind of activities.
p.s. I am assuming that the OP wants rendered image. If not, I'm so
sorry. =)