If it has embarrassingly parallel tasks that it can dispatch to a massively parallel subsystem dedicated to solving embarrassingly parallel tasks, is that abuse or smart use of resources?

That being said most simulation games are usually memory-latency and memory-bandwidth limited, not compute limited.

It looks a perfect match. At first. Then you realize that you are not alone.

Very much like using the Java stream parallel API in a webserver is doing wonders in dev, but not in production, as you have many other threads serving other requests also starved for CPU cores