I personally would love to see a frontend architecture that more closely resembles the backend: dataflow execution. It's been researched in the 80s/90s, with some hybrids popping up in literature from time to time (TRIPS, ADAM, etc)

Counterpoint: processors have gotten so good at extracting ILP from a sequential ISA that I wonder how much benefit you would gain from a dataflow ISA. The downside is that it would require to completely rethink and redo the software stack. (efficient par to seq is just as hard as efficient seq to par)