Doesn't the UEFI firmware map a GPU framebuffer into the main address space "for free" so you can easily poke raw pixels over the bus? Then again the UEFI FB is only single-buffered, so if you rely on that in lieu of full-fat GPU drivers then you'd probably want to layer some CPU framebuffers on top anyway.
Someone last winter was asking for help with large docker images and it came about that it was for AI pipelines. The vast majority of the image was Nvidia binaries. That was wild. Horrifying, really. WTF is going on over there?
You’re assuming a discrete GPU with separate VRAM, and only supporting hardware accelerated rendering. If you have that you almost certainly have more than 2MB of ram
The IBM PGC (1984) was a discrete GPU with 320kB of RAM and slightly over 64kB of ROM.
The EGA (1984) and VGA (1987) could conceivably be considered a GPU although not turning complete. EGA had 64, 128, 192, or 256K and VGA 256K.
The 8514/A (1987) was Turing complete although it had 512kB. The Image Adapter/A (1989) was far more powerful, pretty much the first modern GPU as we know them and came with 1MB expandable to 3MB.
Neither EGA or VGA were "GPUs", they were dumb framebuffers. Later VGA chipsets had rudimentary acceleration, basically just blitters - but that was a help.
The PGC was kind of a GPU if you squint a bit. It didn't work the way a modern GPU does where you've got masses of individual compute cores working on the same problem, but it did have a processor roughly as fast as the host processor that you could offload simple drawing tasks to. It couldn't do 3D stuff like what we'd call a GPU today does, but it could do things like solid fills and lines.
In today's money the PGC cost about the same as an RTX PRO 6000, so no-one really had them.
