Only 3.5go at full speed..
A commonly repeated misconception. All the RAM ran at full speed. The issue was that one bidirectional L2 cache/crossbar link was serving two memory controllers for the final 1GB (8 memory controllers total over 4GB, each serving a 512MB chunk). Because the L2 cache/crossbar links and memory controllers were partitioned into units of two (both with their own internal interconnect) then disabling one L2 cache/crossbar link meant two memory controllers were sharing the single crossbar link.
The only time performance was impacted was when those two memory controllers were both attempting to read or write
at the same time, in which case they were limited to the maximum
unidirectional bandwidth of that shared crossbar link. And in such a situation both of those 512GB chunks could receive a performance impact, depending on which received priority. If one was reading and the other was writing,
everything ran at full speed. Having full control over where memory is allocated, Nvidia scheduled tasks to avoid having simultaneous reads or simultaneous writes pass through those last two controllers (i.e. one chunk reads while the other writes, then switching off). The idea that the GTX 970 had a 'slow 512MB' is easy to state, but completely incorrect as to how things functioned.
As seen from the extensive benchmarking performed on the GTX 970, this also had no practical effect on performance. By the time you upped resolution and texture size enough for there to be a measurable performance impact from lack of available simultaneous read or write performance (e.g. in comparison to a downclocked GTX 980), you are already looking at completely unplayable FPS in the single-digit range regardless of card.
