> Last I checked, the arms race seemed to be almost over. GPUs are only 20% faster per year, and the progress seems to be getting even slower.

This is what happens when you try to make too much inferences of a trend in a short time within a step function.

The primary ingredient that fuels continued performance growth in semiconductors, including GPUs, is process shrinks. They are fundamentally what gives AMD and nVidia the additional transistors that they can use to increase performance. The foundry 20nm processes were a bust, meaning that GPUs effectively skipped a shrink. This means that the last 3 years, after 28nm became available, the only way they have had to increase performance was slight efficiency increases and growing the chips bigger.

GPUs will move to next gen foundry processes, the Samsung/GlobalFoundries 14nm and TSMC 16nm late this year or early next year. This will provide the most dramatic single change in the underlying manufacturing process of the history of making GPUs. Assuming no major architectural advances, expect twice the performance.

Process shrinks are giving diminishing returns as well, due to dark silicon [1]. We can fit more transistors in the same area which gives the possibility for more functionality but if we power them all at the same time we'll just melt the chip.

[1] http://en.wikipedia.org/wiki/Dark_silicon

Despite this general trend, the next process shrink will be the largest single step in GPU history.

4K and VR is going to drive the performance demands up quite a lot very soon.