Each 3d object motion and physics as well can be mapped in a lossy way as an object palette flow graph. You can use tricks like low color, grainyness, mixed resolution high frequency data reduction and A patchy bit of grey scale data for colour upscaling detail and light path adjustment in mixed detail and resolution when compressing and rendering the primer frame before the final output mastered frame. This technology would rely heavily on inference but could do a near ps4 level of detail in 300p 45 FPS on a pi4 level machine or near ps5 level at 4k 120 Hz on the next Nintendo. Basically a lot of very efficient low bit depth calculation with minimal accuracy relative to a good detailed dynamic 3d flow output. You can also integrate fractal data reduction.
So the GPU and video processing unit could be more integrated and more dependent on inference and low bit depth calculation.