The v0.4.1 is available with some bugs corrected, WinXP support, hybrid rendering, improved speed and ATI support ( you need to download their OpenCL SDK because the Catalyst drivers don't include the OpenCL runtime... don't ask me why... ).
I want to share with you some preliminary benchmarks ( using the GPU exclusively and comparing it against the CPU ) :
A simple scene to test ambient occlusion.
A simple path-tracing scene, showing caustics and reflection.
A more complex scene.
Although I've spent a lot of time optimizing specifically for ATI cards, it seems my algorithms don't fit very well the Radeon's 5D-VLWI architecture or their SDK/JIT compiler is not very optimized yet.
On the other hand, the CPU seems to be very slow compared with the NVIDIA's GPUs. Although I allow the option to enable it to render, better keep it disabled and render exclusively with the GPU which is much much much faster.
So, you really should use an NVIDIA card. As you can see, the GT200 cards almost multiply exponentially the results of the GeForce 8/9 cards, so I recommend a GT200 or above card to get good results. If you want my personal choice, I would get a GTX 460 192 bits which delivers incredible performance for only 160$ or a GT240 if you want a sub-100$ card.
Oh, by the way, use the ForceWare 260 drivers if you can. The old 197 ones or the 25X ones have several problems.
If you're a 3dsmax user then you can try the program yourself. Go to http://www.ratgpu.com
Expect several speed optimizations and the Maya's version soon.