Reliability modeling of dynamic thermal management in multicore processor

With the continuous downscaling in semiconductor technology, the growing power density and thermal issues in multi-core processors are challenging and crucial. The system reliability associated with increased power dissipation affect the reliability of thermal management. High temperatures and large thermal variations on the die create severe challenges in system reliability, performance, leakage power, and cooling costs. Dynamic thermal management (DTM) methods regulate the operating temperature based on the provided temperature profile from thermal sensors, which is transmitted using network-on-chip (NoC) in multi-core systems. DTM efficiency is highly dependent on the accuracy of thermal data. Temperature profile inaccuracies are caused by various factors including sensor placement, sensor device imprecision, and interconnection deep sub-micron (DSM) noise. While temperature profile inaccuracies due to sensor placement and sensor device imprecision have been widely addressed, limited study performed on the impact of interconnection DSM noise on DTM efficiency. Hence, this thesis develops a comprehensive simulator model to investigate the impact of interconnect DSM noise on thermal data accuracy and DTM efficiency. The simulation results demonstrate that DSM noise severely affecting the MSbs of thermal data that leads to significant degradation of DTM performance. To mitigate the DSM noise impact on DTM efficiency, an NoC fault tolerance scheme, exploiting inherent characteristics of DSM noise impacting the thermal data, is proposed that comparing to the standard coding scheme achieves lower cost in term of area and power consumption while increasing DTM efficiency by 38%. The second source of chip reliability involves power delivery network (PDN). PDN suffers from long-term reliability threats such as electro- migration (EM). Loss of limited Controlled Collapse Chip Connection (C4) pads to electro-migration makes delivering a stable supply voltage more critical. C4 bumps failure mechanism depends on current density, on-chip voltage noise, and temperature. In this thesis, the C4 bumps failure mechanisms dependency on each individual bumps' temperature value is explored that leads to more accurate mean-time-to-failure (MTTF) of the whole system. The simulation results demonstrate that using uniform temperature leads underestimating the system MTTF by up to 16 times due to exponentially dependency of C4 bump failure to temperature.

Text SOMAYEH RAHIMI POUR - IR.pdf Download (1MB)

Actions (login required)

View Item