We have seen the improvements that the "Steamroller" will offer in performance per watt with its design improvements. In addition to those improvements, AMD will be using "dense" or "thin" libraries employed by its GPU design teams, but for CPU implementation.
AMD told us that products currently shipping with 32nm use a combination of automated place and route and hand-placed semi-custom design (top plot), which reduces power and area somewhat. To deliver more power efficient computations, AMD has employed a high-density cell library to reduce the area and power by 30 percent (bottom plot). The design yields a more portable and energy efficiency CPU core employing industry standard design methodologies well adapted to a foundry model. These improvements, according to AMD, are yielding a 15 to 30 percent lower energy per operation for power constrained designs, as compared to a full process node improvement.
Look for more details from AMD during Hot Chips Symposium on its Surround Computing and Steamroller.