The GPU uses some coalescing to minimize latency and maximize utilization of memory bandwidth. Isn't that likely to merge accesses and thus amortize?
Excuse me we are getting in area of hardware that I am not well researched.
Coalescing patterns are variable across implementations a common point across most hardware is occupancy ratio or hardware-based, context-switch-free multi-threading of sort. Intel calls it HyperThreading®, it's guaranteed to work on two threads whereas GPUs accomodate as many threads they can fit. In my experience is considerably more consistent and easy to use than exploiting coalescing patterns.