It is always possible in theory but never in practice. The idea is stonrg but the silicon is weak. I will find this article tomorow and make some more detailed points then but it is far too alte now (1:30 in the morning!) And I have to get up early tomorow.
Note the practical applications from the XBOX?s nforce1 variant architecture e.g. MC0 (or MC1) for CPU operations MC1 (or/and MC0) for IGP operations. NVIDIA's dual 64bit memory controllers are linked via a crossbar feature.
Just do a google search for "nForce_TwinBank_Memory_Architecture_Tech_Brief.pdf" for
"The radical crossbar memory controller enables CPU and GPU to concurrently access the two 64-bit memory banks and is optimized for 64-bit CPU and GPU accesses to ensure near perfect bandwidth utilization ... Since the high-performance CPU and GPU data types are optimized for 64-bit access, both can access the two memory banks simultaneously and independently, fully utilizing available memory bandwidth " - NVIDIA's marketing materials.