Yes I can see it would be pretty fast, but the computation is defined at run time and its evaluation uses pointers to the various nodes of the computation graph. This limits what optimizations be applied (ie inlining, vectorization, reordering, etc). I'm just curious if anyone has taken this a step further and written something like PSAC except with a code-generation step that creates an optimized implementation of psac_propagate() for the specific computation graph.