Exploring an Efficient Approach for Architecture-Level Thermal Simulation of Multi-core CPUs

In this work, an accurate and efficient thermal simulation approach based on proper orthogonal decomposition (POD) is applied to predict the temperature profile in an Intel Xeon E5-2699v3 CPU consisting of 18 cores. Using the POD method, the thermal problem is projected from a physical domain onto a functional space represented by a finite set of basis functions (or POD modes) that need to be trained by a large amount of thermal data. To generate the static and dynamic power consumption in space for the Xeon E5-2699v3 CPU as the heat source for thermal simulations in the training and validation, the cycle-level system simulator, gem5, and the power simulator, McPAT, are used. Gem5 is used to simulate the architectural characteristics of the CPU and gather performance counters. These statistics are gathered using a subset of the widely used SPLASH2 benchmark suite as the simulated workload. These performance statistics contain information regarding architectural events such as functional unit usage, cache accesses. With the generated power trace, thermal data is collected using FEniCS, an open-source platform that supports finite element method (FEM)-based simulation, subjected to a range of power variations in space and time. It has been demonstrated that the proposed approach is able to offer an accurate thermal prediction of the CPU with a reduction in degrees of freedom (DoF) by more than 5 orders of magnitude and a speedup of 3 orders of magnitude, compared to FEM.