Unintuitively, your two questions are somewhat at odds with each other. The more work you do on the GPU, the more you need to shuffle because the more GPU memo

Unintuitively, your two questions are somewhat at odds with each other. The more work you do on the GPU, the more you need to shuffle because the more GPU memory you’d use AND the more state you’d need to check back on the CPU side, causing sync stalls. It’s not insurmountable, and macOS puts a lot more of its work on the GPU for example. Windows is a little more conservative in that regard. Here are some more confounding factors: - Every app needs one or more buffers to draw into. Especially with hidpi screens this can eat up memory quick. The compositor can juggle these to try and get some efficiency, but it can’t move all the state to the GPU due to latency. - you also need to deal with swap memory. You’d ultimately need to shuffle date back to the system ram and then to disk and back which is fairly slow. It’s much better theoretically on APUs though. Theoretically, APUs stand to solve a lot of these issues because they blur the lines of GPU and CPU memory.