>>6
>>4,8 is speaking the truth. Think about it. Modern GPUs are essentially massively parallel vector processors with thousands of in-order cores, with their own L1 and L2 cache, own on board RAM. The graphics driver is essentially an operating system for this whole other computer slotted into your PCIe bus.
It has to have a thread, warp, and wavefront scheduler (GPU cores can be merged into warps and wavefronts executing from a common instruction stream, or they can be split to execute different instruction streams, or you can have blocks of like 16 cores merged executing a common task, it can get pretty complex). The cores on modern GPUs are pretty much in-order general purpose CPU cores with a 1024-bit SIMD vector pipeline instead of an FPU.
It has to have it's own virtual memory system, and the driver has to manage page tables and mapping into the CPU address space on the PCIe bus.
And then the drivers have to implement both OpenGL and Direct3D as client interfaces, and across multiple versions of these specification, and across multiple platforms.
You better believe they're up in the 10's of millions of lines of code.