As far as I understand it—and this is just from watching Apple's presentations on the architecture—the lack of a discrete GPU is a big part of how the Apple Silicon machines achieve good performance per watt.
Instead of having discrete RAM or a discrete GPU with its own VRAM, all of the RAM is accessible to the CPU and and the GPU in a unified memory architecture. On the M2 Ultra, this allows for 800 GB/s of memory bandwidth, and also eliminates a lot of the need to copy data from RAM to VRAM, as both the GPU and the CPU can access the same memory. In return, this allows the GPU to match the performance of discrete GPUs that have a lot more cores.
Of course, the big downside is that you can't expand the RAM or install a beefier GPU. It's all baked in to the logic board.