This will be a solution primarily for the high-end APU and APU dedicated server markets, it promises to minimize the latency and overhead between the CPU and GPU and will be more cost-effective than on-die L3 cache.
The site claims the stacked memory will achieve higher speeds and lower latencies than DDR4, without the high cost of DDR4 memory. The die size of Carrizo will reportedly be smaller than Kaveri, the APU itself will be made on a 28nm process while the stacked DRAM will be fabbed on 20nm.