From the Nai's Benchmark, assuming if the allocation is caused by disabled of SMM units, and different bandwidth for each different gpus once Nai's Benchmark memory allocation reaches 2816MiBytes to 3500MiBytes range, I can only assume this is caused by the way SMM units being disabled.
Allow me to elaborate my assumption. As we know, there are four raster engines for GTX 970 and GTX 980.
Each raster engine has four SMM units. GTX 980 has full SMM units for each raster engine, so there are 16 SMM units.
GTX970 is made by disabling 3 of SMM units. What nvidia refused to told us is which one of the raster engine has its SMM unit being disabled.
I found most reviewers simply modified the high level architecture overview of GTX 980 diagram by removing one SMM unit for each three raster engine with one raster engine has four SMM unit intact.
First scenario
What if the first (or the second, third, fourth) raster engine has its 3 SMM units disabled instead of evenly spread across four raster engine?
Second scenario
Or, first raster engine has two SMM units disabled and second raster engine has one SMM unit disabled?
Oh, please do notice the memory controller diagram for each of the raster engine too. >.< If we follow the first scenario, definitely, the raster engine will not be able to make fully use of the memory controller bandwidth