AMD is pinning its hopes on APUs, which are a good way of using silicon efficiently and using the same package for the GPU and CPU. However, APUs require a lot of bandwidth, i.e. even cheap APUs benefit from somewhat faster memory modules, and let’s not forget about GDDR5 rumours, which did not pan out with Kaveri.
3D stacked memory should be used as high-bandwidth memory (HBM) that’s smaller yet faster than GDDR5. It can be used as system RAM or graphics memory. HBM was developed mainly for graphics applications and it should enter mass production in 2015. Built using new Wide I/O and TSV technologies, it should support bandwidths from 128GB/s to 256GB/s.
Such bandwidth makes a lot more sense when it comes to GPUs, at least at this point. Both Nvidia and AMD are expected to use stacked memory in their 20nm products, but probably not in first-generation 20nm products.
It remains unclear when AMD will embrace the technology for APUs. This would be somewhat more complicated, at least in consumer products that have to rely on off the shelf technology and components, which makes deployment of new technologies much slower.