The AMD Radeon R9 Fury X kind of backs that statement up since it was able to allocate dynamic VRAM for extra VRAM past its 4GB of dedicated VRAM capacity. We saw up to a 4GB utilization of dynamic VRAM. That allowed the Fury X to keep its 4GB of dedicated VRAM maxed out and then use system RAM for extra storage. In our testing, this did not appear to negatively impact performance. At least we didn't notice anything in terms of choppy framerates or "micro-stutter." The Fury X seems to be using the dynamic VRAM as a cache rather than a direct pool of instant VRAM. This would make sense since it did not cause a performance drain and obviously system RAM is a lot slower than local HBM on the Fury X. If you remember a good while ago that AMD was making claims to this effect, but this is the first time we have actually been able to show results in real world gaming. It is awesome to see some actual validation of these statements a year later. This is what AMD said about this in June of 2015.
Note that HBM and GDDR5 memory sized can’t be directly compared. Think of it like comparing an SSD’s capacity to a mechanical hard drive’s capacity. As long as both capacities are sufficient to hold local data sets, much higher performance can be achieved with HBM, and AMD is hand tuning games to ensure that 4GB will not hold back Fiji’s performance. Note that the graphics driver controls memory allocation, so its incorrect to assume that Game X needs Memory Y. Memory compression, buffer allocations, and caching architectures all impact a game’s memory footprint, and we are tuning to ensure 4GB will always be sufficient for 4K gaming. Main point being that HBM can be thought of as a giant embedded cache, and is not directly comparable to GDDR5 sizes.