The worst part about now in hindsight is how they probably could've moved to a work-list system like Vulkan or io_uring and kept it all in separate processes to

The worst part about now in hindsight is how they probably could've moved to a work-list system like Vulkan or io_uring and kept it all in separate processes to retain that stability. Yes, memory wasn't as plentiful back then and with single-core machines there still would've been more unavoidable context switches but with queues the context switches could've still been reduced in total as opposed to crossing the kernel boundary at all.