UC Riverside

Design and verification of today's nanometer very-large-scale integrated (VLSI) system remain a very challenging problem. For instance, the sub-90-nm technology has caused extremely large parasitic global interconnects and complicated models such as clock networks, power delivery networks and thermal models for packaged systems, which are difficult to be analyzed directly due to the limited computing resources. In addition, the high performance VLSI systems such as multi-core and emerging 3D stacked integrated systems, also lead to excessive high temperature on chip due to the elevated power densities. As a result, temperature should be explicitly managed both at design time through thermal-aware optimization and design techniques and at runtime through on-chip dynamic thermal management (DTM). Hence, accurate yet compact thermal models are required for thermal-aware design and optimization.

In this dissertation, we focus on those challenging issues and have proposed three novel techniques to facilitate the verification of the electronic and thermal effects of the nanometer integrated systems. Specifically, first, we have introduced a wideband model order reduction algorithm (WBMOR) to provide a general solution for the large system analysis problems. With the novel imaginary axis sampling technique and adaptive sample point placement, WBMOR is able to generate a reduced system accurate within the specified frequency band. Second, we have proposed a com- posable thermal modeling technique (ThermComp) for compact thermal modeling. ThermComp builds compact thermal models for each basic module, and uses these models to assemble different multi-core architecture thermal models, which improves the thermal modeling and analysis efficiency at design time. Last but not least, a runtime thermal estimation and prediction method (FRETEP) framework has been proposed to enable fine-grained DTM. With a thermal sensor based error compen- sation method utilizing only limited number of thermal sensors, FRETEP is able to estimate and predict the full-chip thermal behavior accurately with even inaccurate power estimation. Furthermore, a power-driven thermal sensor placement algorithm has been developed for FRETEP to further enhance the thermal estimation accuracy.