Which works fine for a renderer where you're typically serializing the scene and transferring it over anyway.
It's not as tenable for things like deformers etc that need to feed data back in to the scene every frame. Well.. at least not without severe performance repercussions.
As long as it can be streamed or batched you can get pretty much the same throughput with IPC. Only latency suffers slightly (microseconds, still within a frame budget even for realtime use), but you only pay that latency overhead once per batched action.
Chaosgroup solved that problem by using the process boundary. They have an integration plugin in GPL and then execute a standalone version of their renderer that they offer separately.