I've just done some experimenation and can confirm Eric Appleman's statements that Bumblebee is not necessary for CUDA. (more about Bumblebee: Is a NVIDIA GeForce with Optimus Technology supported by Ubuntu?)

However, when you need to show graphical examples using OpenGL, you do need something like Bumblebee for Optimus systems, otherwise you either don't see anything on your display or get the error:

ERROR: Support for necessary OpenGL extensions missing.

When running a CUDA program, you need to install the CUDA toolkit and a nvidia driver. If you intend to compile programs, you also need the SDK. The installers can be found on http://developer.nvidia.com/cuda-downloads, please read the below instructions before borking your Optimus laptop.

Installing CUDA

Driver

I recommend to install the nvidia driver from Ubuntu's package manager. If you install Bumblebee, you don't need to worry about the driver. Otherwise, after installation, disable the nvidia libraries as described on https://askubuntu.com/a/107746/6969. If you do not do so, you'll loose 3D acceleraration and possibly get stuck on a low resolution.

Toolkit

Basically you have to download the installer, make it executable and run it. - Download the installer. As of this writing, 4.2.9 is the most recent driver. As I've a 64-bit OS, I use the 64-bit 11.04 Ubuntu package (though I'm running 12.04) - Make it executable and allow installation to /usr/local/cuda:

chmod +x cudatoolkit_4.2.9_linux_64_ubuntu11.04.run
sudo ./cudatoolkit_4.2.9_linux_64_ubuntu11.04.run

When the installation message occurs which asks you where to install CUDA, just press Enter to accept the default /usr/local/cuda:

......................................
Enter install path (default /usr/local/cuda, '/cuda' will be appended):

After installation, it'll print some messages that suggest to put the cuda library directory to your library search path:

========================================

* Please make sure your PATH includes /tmp/cuda/cuda/bin
* Please make sure your LD_LIBRARY_PATH
*   for 32-bit Linux distributions includes /tmp/cuda/cuda/lib
*   for 64-bit Linux distributions includes /tmp/cuda/cuda/lib64:/tmp/cuda/cuda/lib
* OR
*   for 32-bit Linux distributions add /tmp/cuda/cuda/lib
*   for 64-bit Linux distributions add /tmp/cuda/cuda/lib64 and /tmp/cuda/cuda/lib
* to /etc/ld.so.conf and run ldconfig as root

* Please read the release notes in /tmp/cuda/cuda/doc/

* To uninstall CUDA, remove the CUDA files in /tmp/cuda/cuda
* Installation Complete

You may skip this step if you want, but then you've to set LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda/lib:$LD_LIBRARY_PATH later when running a program.

SDK

If you want to be able to compile CUDA applications, you can install the SDK in a similar way as described above. Download, make it executable and run it (not as root, e.g. without sudo!). When compiling parts of it, I had to make some changes to NVIDIA_GPU_Computing_SDK/C/common/common.mk though:

• after about line 189, OPENGLLIB := -lGL -lGLU -lX11 -lXi -lXmu, add:

  OPENGLLIB += -L/usr/lib/nvidia-current -L/usr/lib32/nvidia-current
  

• after about line 271, swap ${RENDERCHECKGLLIB} with ${OPENGLLIB} such that it looks like:

  LIB += $(RENDERCHECKGLLIB) ${OPENGLLIB} $(PARAMGLLIB) ${LIB} -ldl -rdynamic
  

• a few (5) lines further, do the same, but remove -lcuda too.
• a few (7) lines further, do the same as you did in line 271.

Using CUDA

CUDA does not need a nvidia-driven X server to work. In that case you can run your random test program like:

LD_LIBRARY_PATH=/usr/lib/nvidia-current:/usr/lib32/nvidia-current:$LD_LIBRARY_PATH someComputallyIntensiveProgram

If you have not added CUDA to your library path, you will need:

LD_LIBRARY_PATH=/usr/lib/nvidia-current:/usr/lib32/nvidia-current:/usr/local/cuda/lib64:/usr/local/cuda/lib:$LD_LIBRARY_PATH someComputallyIntensiveProgram

(you can strip the 32-bit paths from it if your program is 64-bit).

If the CUDA program does have something to display using OpenGL, you have to use optirun:

 optirun blender

Or, if you did not have CUDA added to your default path:

LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda/lib:$LD_LIBRARY_PATH optirun blender

Try running

vblank_mode=0 optirun glmark2

Apparently there's a cap in place to limit the FPS to the refresh rate of the monitor.