Dedicated hardware makes sense in very specific scenarios, otherwise general-purpose programmable compute is the way to go. The primary reason we've landed on CPU + GPU is because they're aiming at different tradeoffs for dealing with general-purpose compute loads: CPU for maximizing throughput of complex, sequential tasks and GPU for more limited, but highly parallel tasks.
Dedicated lighting hardware is actually a step backwards. Fixed-function pipelines were typical of 3D hardware last century but we've moved on. Fluid simulation is a pretty good fit for GPGPU. Generalized physics simulations are a more complicated story that aren't a great fit for dedicated hardware for an end-to-end solution. Some aspects, like cloth simulations, that only impact visuals work well on the GPU. Anything that requires a round-trip back to the CPU to impact gameplay, like collision detection, becomes a much trickier proposition not for reasons of throughput, but of latency.
