Power Mapping of Computing Devices: Fundamentals and Applications

of “Power Mapping of Computing Devices: Fundamentals and Applications” by Abdullah Nazma Nowroz, Ph.D., Brown University, May 2014 Power consumption limits the maximum achievable performance of modern processors. Accurate power modeling is an essential step in both mobile and high-end processors. Large-scale transistor-level power modeling is computationally very challenging. As a result high-level power modeling is performed at the expense of accuracy. Pre-silicon power tools need to be complemented with post-silicon characterization to determine the true power consumption of circuits. Post-silicon power maps are developed during debugging and characterization phases of the first-silicon, and then applied to improve the design during re-spins and for future designs. Post-silicon power results can also improve the accuracy of power and thermal models. We provide fundamentals for post-silicon power validation using captured thermal infrared emissions from back-side of integrated circuits. We identify challenges associated with thermal-to-power inversion [8]: (1) spatial heat diffusion which blurs underlying power maps, and (2) measurement noise in thermal imaging systems. We address these challenges by devising optimization formulation that incorporates Tikhonov filtering. To further increase the accuracy of power mapping, we propose to use AC-based thermography which reduces the impact of flicker noise and spatial heat diffusion. The average power mapping error reduced from 40% using DC-based method to about 8.5% using the proposed AC-based method. A programmable circuit of micro heaters is implemented, and used to validate our methods and to quantify the improvements in power mapping attained from regularization techniques and AC-based methodology. Using our novel power mapping framework, we present versatile applications of postsilicon power mapping. First, we power map a soft processor embedded in a FPGA chip, and a real multi-core processor using DC-based techniques, where we decompose the