I've sent the executable of the teapot in Sponza scene to a friend, szczyglo74 on Youtube, who has a much more powerful rig than my own (a PC with 3 GPUs: 1 GTX 590 (dual gpu card) + 1 GTX 460) and who made some very cool real-time videos of the sponza scene. Many thanks szczyglo!
The maximum path depth in each of these videos is 4 (= 3 bounces max):
32 spp per frame (480x360, 1/4th render resolution, 8 fps):
http://www.youtube.com/watch?v=fVAl-oKAL9I :awesome video, shows real-time convergence in most parts of the scene
32 spp per frame (640x480, 1/4th render resolution, 4.7 fps):
8 spp per frame (480x360, 1/4th render resolution, ~21fps):
4 spp per frame (640x480, 1/4th render resolution, ~18fps):
The above videos clearly show the importance of the number of samples per pixel per frame in indirectly lit areas: despite the low max path depth of 4, it is still possible to discern some details in the corridors in the first two videos (32 spp per frame) during navigation, while the last two videos (8 and 4 spp per frame) are obviously too dark in these regions, but clear up very fast with a stationary camera. Note that these tests were made with a kernel that is not optimized for indirect lighting (no multiple importance sampling is used here).
I'm quite happy with the sense of photorealism in these videos, especially when you consider that this is just brute force path tracing (no caching, filtering or interpolation yet nor anything like image reconstruction, adaptive sampling, importance sampling, bidirectional pt, eye path reprojection, ... which are all interesting approaches). A textured version of Sponza will probably further increase the realism, which is something I will try in a next test.