Exploring the design space for 3D clustered architectures

Journal Article3D die-stacked chips are emerging as intriguing prospects for the future because of their ability to reduce on-chip wire delays and power consumption. However, they will likely cause an increase in chip operating temperature, which is already a major bottleneck in modern microprocessor design. We believe that 3D will provide the highest performance benefit for high-ILP cores, where wire delays for 2D designs can be substantial. A clustered microarchitecture is an example of a complexity-effective implementation of a high-ILP core. In this paper, we consider 3D organizations of a single-threaded clustered microarchitecture to understand how floorplanning impacts performance and temperature. We first show that delays between the data cache and ALUs are most critical to performance. We then present a novel 3D layout that provides the best balance between temperature and performance. The best-performing 3D layout has 12% higher performance than the best-performing 2D layout