Cyber-physical thermal management of 3D multi-core cache-processor system with microfluidic cooling

Abstract

Existing three-dimensional (3D) integration of multi-core processor-cache system is confronted with

the problem of die thermal run-away hazard. This teething problem is addressed by a real-time

demand-based thermal management with non-uniform microfluidic cooling in this paper. A novel

runtime temperature management is implemented to predict the real-time temperature demand

based on software-sensing with prediction-and-correction. A 3D thermal model is developed to cater

the real-time microfluidic thermal dynamics. An autoregressive (AR) prediction with correction is

implemented by Kalman filtering to predict and correct the runtime power, which is further used to

calculate the future temperature demand. With this software-sensing approach, thermal management

can be performed in a cyber-physical fashion with real-time sensing, prediction-and-correction

and fine-grained control. Compared to the existing works using on-chip temperature sensors, our

closed-loop controller with software-sensing avoids the cost of sensor implementation and deployment.

Our work analyzes and predicts the fine-grained temperature profile of multi-core processorcache

system. It enables a number of microfluidic channels to be adjusted adaptively with different

flow-rates to control the system temperature proactively as opposed to the static control with a uniform

flow-rate for microfluidic channels. With the proposed cyber-physical temperature management

scheme, it is shown that the temperature of multi-core system is suppressed below an acceptable

thermal threshold. In fact, the fine-grained flow-rate control also achieves a more even temperature

distribution and saves up to 72.1% of total flow-rate compared with uniform flow-rate controls.