Trying to port our ray-tracing software to a single GPU was quite a success. Thus,encouraged by this enormous performance gain we purchased three GTX 280 to really boost the application. Unfortunately, it seems that CUDA’s multi-GPU support is not really good for Linux. First we thought to have a bug in our routines, but when porting the code to Windows Vista things speeded up as expected. I hope the NVIDIA guys are working on that problem in order to get a fix for CUDA to run under Linux as expected.