Just read an interesting paper called "Guided Image Filtering for Interactive High-quality Global Illumination" by Bauszat, Eisemann and Magnor, which was accepted at EGSR 2011. The paper describes a novel method for edge-aware filtering of noisy path traced images using the normal and depth buffers instead of the noise image itself. The indirect lighting is filtered separately from the direct lighting and the whole pipeline (path tracing + filtering) is implemented on the GPU using CUDA (the numbers in the paper are obtained with a GTX 285). The new method is compared to cross bilateral filtering and the edge-avoiding à-trous wavelet filter described in a HPG paper from last year and the results seem very promising.
I very much agree with the introduction to the paper:
"...many approximations exist to reproduce certain effects of the real global illumination and try to incorporate it into a classic rasterizer. Unfortunately, in order to speed up the computation they are either aiming at simulating only a single effect, like soft shadows [ED08] or ambient occlusion [RGS09] or they are are limited to certain constraints, requiring low-frequency lighting environments [SKS02], static scenes [LSK07] or precise parameter adjustment for varying scenes [RGK08]. Code complexity increases tremendously for systems that simulate global illumination effects using rasterization because multiple approximations need to be combined to achieve a sufficient visual quality. On the other hand path tracing naturally incorporates all these effects with a very simple rendering algorithm.
Another drawback of approximative algorithms is that they not necessarily converge to the correct result. Therefore, they will only become faster with better hardware but do not necessarily produce a higher quality result. Instead of trying to incorporate global illumination effects into a rendering pipeline which is not capable of physically correct rendering, it is a much more promising design to take a physically valid rendering approach and implement approximative algorithms on top to speed up the rendering. Designing them in an adjustable way results in a very future-oriented rendering paradigm."
Research in fast interactive GPU path tracing with minimal noise has made enormous advancements over the past year and I'm confident that real-time high-quality path traced global illumination for games will be feasible soon.