One of the main things about Steamroller that interests me is the fact that single-threaded performance will get a massive boost, as well as them fixing the problem with a module's resource allocations when both cores are at full load. Those issues getting fixed will make it a much better product.

Each core of a module will get its own integer decoder, which will improve performance of a module when both cores are at full load. The floating point pipeline lengths will be reduced, which will grant a nice boost to single-thread IPC.

Those two are really some of the biggest flaws of the Bulldozer initial design, which Piledriver enhanced but didn't fix completely. Among those will be further perf/performance improvements and overall efficiency improvements. I'm really eager to see how Kaveri and the eventual Steamroller FX will pan out.