Many attempts were done in this area in the late 80s/early 90s (and before if you don't focus on lazy evaluation; search lisp machine).
Search hardware graph reduction.
The major problem has always been that producing a dedicated CPU is expensive and there's no way to keep up with the progresses in general purpose processor that can justify investments with a hugely larger user base.
At some point there were even attempts of placing computation in the memory itself (typically using very fine grain combinators, SKI reduction and the such)
People have given up even on hardware support for small subproblems, such as garbage collection.
As for modifying the instruction set of an Intel processor, I don't know how feasible it is. But even if it is, consider that the entire architecture, pipelining, caching, predictions, speculative everythinbg etc. is hugely optimized for the typical workflow. You change that and all bets are off w.r.t performance and you may or not be ahead of the same CPU executing normal code out of a haskell compiler.