GPU rendering programs are like regular rendering programs such as Vray and mental ray, but GPU rendering engines can run on GPUs. Current versions of Vray and mental ray include GPU engines, and there are purpose-built GPU renderers like Arion.
GPU renderers are like Maxwell Render, but run on a GPU which makes them in many cases as fast as (or even faster than) CPU renders, and like Maxwell, in many other cases are impractical.
GPU render engines are not like the ones you're used to. They don't run like mental ray and Vray, which first calculate lighting and then render each pixel. GPUs have the ability to simultaneously process a very large number of simple operations, so they are not suited to the type of computation that regular Vray and mental ray run. Instead, GPU renderers work by performing many ray traces in parallel. A ray fired to each pixel on the screen can be calculated many, many times, and each time its quality is improved. You can run the render for as long as you like - at first you will see a very grainy image, like a digital photo taken in low light, and over time it will resolve into a smoother image. How long that takes will depend on the materials and lighting and the size of the render output. It could be minutes or several hours.
The algorithm doesn't leave room for user optimizations, so it can be easier to use than mental ray and Vray which have many settings. GPU renderers don't have a lot of buttons and numbers to enter. But the flipside is that you can't really optimize the render engine to the scene. There are a lot more people using CPU renders in production than GPU. Also, GPU renderers are not as good for animation as CPU ones, because you can't run the lighting calculation once, save it and use it for multiple frames - each frame must be calculated completely on its own.
If you're in doubt, the answer is probably that you don't need to worry about buying hardware for GPU rendering right now.