"To really answer the question “will discrete GPUs die out?” we need to look at the quantum level. Despite having the power budget for nearly unbounded performance, one of the bottlenecks for discrete GPUs is PCI Express, the interface to the system. In the case of graphics workloads, normally PCI Express does not present a constant bottleneck, however, everyone and their dog has heard of parallel computing. This is the case where there is a lot of traffic between the discrete GPU, CPU and main memory, and PCI Express becomes a liability in terms of bandwidth and latency.
Many people do not realize this, but the “life” of a CPU is dreadful and boring. It exists primarily waiting for data. The use of discrete GPUs for parallel computing through PCI Express will not improve the “quality of life” of the CPU. While some discrete GPUs will offer graphics and pure FLOP performance over an APU, the performance will be limited, in some cases, by the PCI Express interconnect.
This is where AMD Fusion APUs will shine. AMD Fusion APUs have not only been designed to offer great graphics performance, they also have been designed to offer great parallel compute performance. The fact that the CPU core resides next to the GPU core connected by a bus of mere nanometers, helps diminish the bandwidth and latency issues presented to parallel computing on a PCIE bus.
The design plan for successive generation of AMD APUs includes architectural innovation, as well as tighter and faster interconnects between the CPU cores and the GPU cores. One goal is to advance the parallel compute capabilities without sacrificing x86 and graphics performance."