AMD doesn't mention "next generation memory" in their roadmaps until after Navi, so it will take a while for AMD to move to something that is faster than HBM.
AMD doesn't mention "next generation memory" in their roadmaps until after Navi, so it will take a while for AMD to move to something that is faster than HBM.
What Nvidia has done well in recent years is compression, getting more out of less memory bandwidth. Using Delta colour compression and other technologies to lower bandwidth requirements. There is a reason why the R9 290X needed a 512-bit memory bus and the GTX 780Ti needed a 314-bit bus.
AMD is only starting to catch up in this regard. Nvidia is also ahead of the game when it comes to rasterization, eliminating certain (invisible) tasks before any bandwidth or clock cycles are wasted on it. AMD is only now catching up with DSBR in Vega.
AMD excels when it comes to pure compute performance, but Nvidia has a lot of underlying technology that has helped them get a lot more from less. This has allowed them to get more from cheaper chips, increasing their margins and delivering appealing power characteristics in many workloads.