Last time I messed with software rendering I wanted to set the base performance expectation so I blitted the same memory contents to the screen (3840x2160@120) with no sync using OpenGL and got 200-250 FPS on a 2080 Ti/3900X desktop. That means the entirety of my program logic and CPU rendering had to be less than the time it took for the CPU to hand the memory to the GPU to draw or I'd start dropping frames. I'd agree this just isn't realistic anymore going forward from a performance perspective, not to mention the power usage.

I ended up targeting 1920x1080@60 and having a shader nearest neighbor scale that up.