When using software-based rendering, the graphics processor in your render nodes won't make a bit of difference in the performance or final image. You can use an integrated GPU or a spare mainstream card you have laying around, but since you'll primarily be connecting to the render nodes using a VNC client, it won't matter.
However, if you are instead using GPU-based rendering, you need to plan on adding a graphics card that the rendering solution in each node supports. If you are using Nvidia Gelato, Nvidia's solution for GPU-based 3D rendering, any GeForce FX 5200 or higher is usable as your rendering device, while Gelato Pro requires a Quadro FX.
Support for Gelato appears to be trailing off in favor of Compute Unified Device Architecture (CUDA) development. Between CUDA and OpenCL, we can expect to see at least some support for GPU-based acceleration for rendering (and simulation/dynamics) in the near future, so that instead of all of the calculations taking place on the CPU, calculations particularly suited to the GPU (especially massively parallel floating-point operations) are offloaded to it. This is essentially the same thing that was done a few years ago with dedicated digital signal processor (DSP) farm cards like the ICE BlueICE, except commodity GPUs are used instead of more expensive DSPs or custom processors.
The caveat here is that GPU-based rendering (or acceleration) is still in its infancy. If you think you might make a transition in the future, then it is important to plan accordingly when buying 1U enclosures, as discussed previously. It is thus a good idea to go ahead and acquire the matching PCIe x16 riser cards so that you can upgrade later. Of course, this also means that the motherboards in your render nodes need at least a x16 PCIe slot.
In most cases, any individual 3D or compositing application will include its own support for network rendering. For example, LightWave offers ScreamerNet, 3ds Max includes Backburner (which Combustion uses as well), Maya includes Maya Satellite, and After Effects offers the Render Queue. But running all the needed separate render controllers on each one of your nodes is a comparative waste of system resources. Also, the limitations of the included render managers are numerous and they may or may not allow for monitoring of individual systems in order to check the status of a render or frame integrities on a per-system basis or to see if individual nodes have crashed.
In order to efficiently manage multiple systems and multiple renderers, centralized render controllers are thus required. There are several applications that can control multiple software programs from a single render controller. To name a few, Qube!, ButterflyNetRender, Deadline, Rush, and RenderPal are all commercial products. Moreover, there are literally dozens of shareware, freeware, and open-source controllers available as well. Going over the various features of these different render controllers would be an article in itself. Suffice to say that most of them actively monitor the systems for output, check the output frames for integrity, and notify you upon completion of a job. Some also support features like sending SMS messages. More advanced solutions allow for remote management, priority assignment, and handling of some of the local setup tasks themselves.
When shopping for a render controller, first look to see if your solution of choice supports the software you plan to use. Then, look at the features you need and compare your budget.