VRAM. As of now, the GPU has AFAIK no access to the machine’s RAM. So all data relevant for rendering has to be transferred into the VRAM before the rendering process can begin.
Whenever the VRAM is full, the render (but not necessarily Blender) crashes. Since the GPU has no other resources available but its own, it will abort the render as soon as those resources are depleted. That’s why as many VRAM as you can get is so important when deciding on a graphics card and that’s why memory consumption has always been the Achilles’ heel of GPU rendering.