Does disabling a graphical desktop significantly improve the performance of CUDA code?

Sorry, not expert in CUDA ... but ... I'd assume so, but just how significant is another issue: just how much of the GPU is actually used by the desktop? I guess that depends on the desktop, some might not even use much GPU at all (possibly even nothing). From something similar: I always turn off the decorative desktop elements in Windows (especially since Vista and later) because I use 3dStudio / Revit / AutoCAD many times throughout my workday. These are highly graphics intensive programs and I've noticed significant slowdowns due to stuff like Glass effects in Vista.

This is very interesting question. The answer depends a lot on two factors. 1) GPU in use 2) The OS's desktop/window manager. If you're using a low end GPU, such  as a GT620, which has limited compute resources, then it could mean that  the GPU's resources might be strained. In which case, an increase in  performance may be possible. Nvidia's  drivers limit the amount of time a kernel can run uninterrupted on a  GPU. Priority is always given to incoming rendering requests from the  OS. Also, for kernels that deal with large amounts of memory, having an active GUI limits the amount of memory available for dedicated compute use. Finally, should you're application have a lot of i/o, such as cuda streams that constantly move data in and out of the device, then your GPUs DMA and the PCIe bus will also have to deal with the additional load from the GUI. I've found I/O to be major bottleneck at times. The fact that nvidia drivers throttle PCIe to run on v2.0, due to stability issues, doesn't help either :) So, to sum up my answer, disabling the GUI MIGHT be of use to you. You'd need to evaluate it on a case-by-case basis. However, if consistent performance and accurate run-times are a concern for you, then I'd recommend that you allow the integrated GPU on your processor handle your desktop, and dedicate your discrete GPU to CUDA workloads.