I think it would be interesting, especially for dynamic use cases where you can recompile the graph but use it many times. Although it's also probably pretty niche to get to that level of performance from these libraries.

At that point I'd guess most projects start to build up their own set of primitives (e.g., containing subsets of particular graphs) that they can hand-optimize better than these general purpose computation graphs. That's what I ended up doing so I could better control the tradeoff around which operations should be batched, heuristics around how cheap certain operations are (especially based on the elements at the time), more complex dependency tracking when working with data outside the graph, etc.