[Paper Review] Implications of Integrated CPU-GPU Processors on Thermal and Power Management Techniques

Kapil Dev, Indrani Paul+4 more|arXiv (Cornell University)|Aug 29, 2018

Parallel Computing and Optimization Techniques2 references4 citations

TL;DR

This paper investigates thermal and power management in integrated CPU-GPU processors using infrared thermal imaging and power mapping on real hardware. It demonstrates that DVFS and workload scheduling are tightly coupled in heterogeneous systems, and that application-level scheduling via OpenCL enables more efficient thermal and power management than OS-level scheduling, reducing peak temperature and power density by leveraging architectural differences and proximity effects.

ABSTRACT

Heterogeneous processors with architecturally different cores (CPU and GPU) integrated on the same die lead to new challenges and opportunities for thermal and power management techniques because of shared thermal/power budgets between these cores. In this paper, we show that new parallel programming paradigms (e.g., OpenCL) for CPU-GPU processors create a tighter coupling between the workload, the thermal/power management unit and the operating system. Using detailed thermal and power maps of the die from infrared imaging, we demonstrate that in contrast to traditional multi-core CPUs, heterogeneous processors exhibit higher coupled behavior for dynamic voltage and frequency scaling and workload scheduling, in terms of their effect on performance, power, and temperature. Further, we show that by taking the differences in core architectures and relative proximity of different computing cores on the die into consideration, better scheduling schemes could be implemented to reduce both the power density and peak temperature of the die. The findings presented in the paper can be used to improve thermal and power efficiency of heterogeneous CPU-GPU processors.

Motivation & Objective

• To analyze the thermal and power behavior of integrated CPU-GPU processors under real workloads using high-resolution thermal and power mapping.
• To investigate the interdependence of dynamic voltage and frequency scaling (DVFS) and workload scheduling in heterogeneous processors.
• To evaluate the impact of CPU core proximity to the GPU on thermal coupling, leakage power, and peak temperature.
• To identify design implications for thermal-aware scheduling and power management in CPU-GPU systems.
• To propose strategies for optimizing performance and energy efficiency by jointly considering DVFS and scheduling decisions.

Proposed method

• Employed infrared (IR) thermal imaging with an IR-transparent oil-based heatsink to capture high-resolution die-level thermal maps of a real CPU-GPU processor during workload execution.
• Inverted thermal maps to compute detailed power maps with high accuracy, enabling precise power and temperature breakdown per hardware module.
• Executed a range of OpenCL and SPEC CINT 2006 workloads across different DVFS settings and CPU core affinities to study thermal and power dynamics.
• Compared OS-based scheduling with application-based scheduling via OpenCL Runtime to evaluate thermal and power efficiency tradeoffs.
• Analyzed thermal coupling between CPU cores and GPU based on physical proximity, measuring peak temperature and total power for different core assignments.
• Used floorplan-aware scheduling to correlate spatial proximity with thermal impact, identifying optimal CPU cores for GPU kernel launch.

Experimental results

Research questions

• RQ1 How do DVFS and workload scheduling interact in terms of performance, power, and temperature in heterogeneous CPU-GPU processors?
• RQ2 What is the impact of CPU core proximity to the GPU on thermal coupling, leakage power, and peak die temperature?
• RQ3 How does application-based OpenCL scheduling compare to OS-based scheduling in thermal and power efficiency?
• RQ4 What are the implications of asymmetric power density and thermal profiles between CPU and GPU for thermal management?
• RQ5 Can localized cooling or frequency boosting be effectively applied based on thermal slack and architectural differences?

Key findings

• Application-based scheduling via OpenCL Runtime achieves better thermal and power efficiency than OS-based scheduling by exploiting architectural heterogeneity.
• Peak temperature increased by 11°C and total power by 4.1 W when launching a GPU kernel from Core3 (closest to GPU) versus Core0 (farther from GPU), due to stronger thermal coupling.
• At 3 GHz DVFS, the CPU reaches thermal limits before the GPU, despite GPU being the primary compute unit, due to higher thermal sensitivity and lower thermal headroom.
• The GPU exhibits lower peak temperature than the CPU at high DVFS settings, indicating available thermal slack that can be used for performance boosting.
• The GPU's lower thermal density allows for fewer thermal sensors per unit area compared to the CPU, enabling cost-effective thermal monitoring.
• Local cooling solutions (e.g., thermoelectric coolers) can be effectively targeted at CPU and GPU separately, given their distinct thermal profiles and power densities.