Arriving with the new Fiji GPU, Fury X brings with it some new architecture, including High Bandwidth Memory.
Manufactured with perhaps the final high-end outing for TSMC's 28nm process, the Fiji is an 8.9m transistor monster measuring 596 square mm. However as the first GPU of any stripe to make the use of High Bandwidth Memory the overall package size is tiny, measuring just 1011 sq. mm when including the area of the memory chips and substrate rather than roughly 3x that. AMD believe that this will truly unlock new form factors, over and above what was possible when more bulky GDDR5 chips are arrayed around a high-performance GPU.
Fiji itself is a 4096-shader part, a little under 50% more shaders than the Hawaii-based R9 290X/390X; if you're counting in Compute Units it puts the Fiji at 64. This gives Fiji unprecedented processing power for an AMD GPU, especially considering the commensurately large number of Texture Mapping Units (256) and ROPs (64). It's built on a next-generation GCN architecture which will also make it marginally more efficient than Hawaii's older GCN 1.1 design, and naturally efficiency will be key in leveraging the best possible performance without blowing through power allowances.
Four HBM stacks are arrayed around the Fiji GPU, and with this first version of HBM you're currently limited to 1GB VRAM per stack, 4GB in total. Compared to other modern GPUs that doesn't seem like much - NVIDIA's TITAN X has a 12GB frame buffer, whilst AMD's only R9 390X will be shipped with 8GB GDDR5 - but it's worth noting that the performance of HBM is on a different level.
In total the memory bus width is 4096-bits wide, massive when compared to any GPU in its class. AMD are targeting a 512Gb/s Memory Bandwidth - 130GB/s more than the R9 390X, and almost 200Gb/s greater than that available to NVIDIA's Titan X - and that really changes the dynamic of high resolution gaming. Frame buffer will still be a factor, but presupposing the GPU can process the data quickly enough it may just be that the 4GB limitation will not be nearly as problematic at 4K as everyone is currently expecting.
Furthermore Fiji has a 4x memory bandwidth/watt improvement over the R9 290X, which allows more power to be funneled to the GPU when operating within a TDP envelope. That's critically important as GDDR5 hits diminishing returns at a certain performance level, costing ever more power to push frequencies ever higher, whilst the depths of HBM have yet to be plumbed. By being ahead of the curve AMD have the opportunity to steal a march in GPU efficiency, the first time they will have done so since the release of the HD 7970.