Power Modeling for GPU Architectures Using McPAT

Lim, Ji‐Eun; Lakshminarayana, Nagesh B.; Kim, Hyesoon; Song, Won Kyung; Yalamanchili, Sudhakar; Sung, Wonyong

doi:10.1145/2611758

Cited by 58 publications

(19 citation statements)

References 33 publications

(33 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Ma et al use a statistical model to estimate the best GPU configuration to save power [16]. Lim et al propose a new power model for GPUs that is based on empirical data from a GTX 580 [13]. One simulation-based paper on thermal management for GPUs discusses methods for managing power through architecture manipulation (clock gating, fetch gating, dynamic voltage scaling, multiple clock domains, and floorplanning) [18], which is orthogonal to our study.…”

Section: Related Workmentioning

confidence: 99%

Effects of source-code optimizations on GPU performance and energy consumption

Coplin

Burtscher

2015

Proceedings of the 8th Workshop on General Purpose Processing Using GPUs

View full text Add to dashboard Cite

This paper studies the effects of source-code optimizations on the performance, power draw, and energy consumption of a modern compute GPU. We evaluate 128 versions of two n-body codes: a compute-bound regular implementation and a memory-bound irregular implementation. Both programs include six optimizations that can be individually enabled or disabled. We measured the active runtime and the power consumption of each code version on three inputs, various GPU clock frequencies, two arithmetic precisions, and with and without ECC. This paper investigates which optimizations primarily improve energy efficiency, which ones mainly boost performance, and which ones help both aspects. Some optimizations also have the added benefit of reducing the power draw. Our analysis shows that individual and combinations of optimizations can alter the performance and energy consumption of a GPU kernel by up to a factor of five.

show abstract

Section: Related Workmentioning

confidence: 99%

Effects of source-code optimizations on GPU performance and energy consumption

Coplin

Burtscher

2015

Proceedings of the 8th Workshop on General Purpose Processing Using GPUs

View full text Add to dashboard Cite

show abstract

“…Nagasaka et al [93] model GPU power of kernels using performance counters. Lim et al [94], GPUWattch [95] and GPUSimPow [96] are simulation based power models. McPAT [97] is the basis for Lim et al [94] and GPUWattch [95] uses GPGPUSim [98] to simulate execution time.…”

Section: Energy Modelingmentioning

confidence: 99%

Analytical Cost Metrics : Days of Future Past

Prajapati

Rajopadhye

Djidjev

2018

View full text Add to dashboard Cite

ANALYTICAL COST METRICS : DAYS OF FUTURE PASTFuture exascale high-performance computing (HPC) systems are expected to be increasingly heterogeneous, consisting of several multi-core CPUs and a large number of accelerators, specialpurpose hardware that will increase the computing power of the system in a very energy-efficient way. Specialized, energy-efficient accelerators are also an important component in many diverse systems beyond HPC: gaming machines, general purpose workstations, tablets, phones and other media devices. With Moore's law driving the evolution of hardware platforms towards exascale, the dominant performance metric (time efficiency) has now expanded to also incorporate power/energy efficiency.This work builds analytical cost models for cost metrics such as time, energy, memory access, and silicon area. These models are used to predict the performance of applications, for performance tuning, and chip design. The idea is to work with domain specific accelerators where analytical cost models can be accurately used for performance optimization. The performance optimization problems are formulated as mathematical optimization problems. This work explores the analytical cost modeling and mathematical optimization approach in a few ways. For stencil applications and GPU architectures, the analytical cost models are developed for execution time as well as energy.The models are used for performance tuning over existing architectures, and are coupled with silicon area models of GPU architectures to generate highly efficient architecture configurations.For matrix chain products, analytical closed form solutions for off-chip data movement are built and used to minimize the total data movement cost of a minimum op count tree.ii ACKNOWLEDGEMENTS I would like to thank my advisor, Sanjay Rajopadhye, for his guidance and support. He has been a continuous source of motivation. His brilliant ideas and farsightedness have awed me throughout my Ph.D. I learnt a lot from him and enjoyed being his student. I not only gained technical knowledge in the field but also learnt values of a good human being like patience, kindness, selflessness, compassion, and dedication. I feel honored to have the opportunity to work with him.

show abstract

“…As is the case with modeling efforts, simulations of power, temperature, and other computing characteristics were initially developed for CPUs (e.g., see Brooks et al [2000], , and Zhang et al [2003]), yet the complicated and evolving GPU architecture requires tailored simulation research. For example, see the power simulation work of Lim et al [2013] where GPU and CPU architectural differences are itemized. GPU power simulators are often built on performance and power models using hardware data that is predicted rather than observed, for example, from code analysis.…”

Section: Gpu Power Simulatorsmentioning

confidence: 99%

Understanding GPU Power

2016

View full text Add to dashboard Cite

Modern graphics processing units (GPUs) have complex architectures that admit exceptional performance and energy efficiency for high-throughput applications. Although GPUs consume large amounts of power, their use for high-throughput applications facilitate state-of-the-art energy efficiency and performance. Consequently, continued development relies on understanding their power consumption. This work is a survey of GPU power modeling and profiling methods with increased detail on noteworthy efforts. As direct measurement of GPU power is necessary for model evaluation and parameter initiation, internal and external power sensors are discussed. Hardware counters, which are low-level tallies of hardware events, share strong correlation to power use and performance. Statistical correlation between power and performance counters has yielded worthwhile GPU power models, yet the complexity inherent to GPU architectures presents new hurdles for power modeling. Developments and challenges of counter-based GPU power modeling are discussed. Often building on the counter-based models, research efforts for GPU power simulation, which make power predictions from input code and hardware knowledge, provide opportunities for optimization in programming or architectural design. Noteworthy strides in power simulations for GPUs are included along with their performance or functional simulator counterparts when appropriate. Last, possible directions for future research are discussed.

show abstract

Power Modeling for GPU Architectures Using McPAT

Cited by 58 publications

References 33 publications

Effects of source-code optimizations on GPU performance and energy consumption

Effects of source-code optimizations on GPU performance and energy consumption

Analytical Cost Metrics : Days of Future Past

Understanding GPU Power

Contact Info

Product

Resources

About