A design methodology for energy aware neural networks

Dabbagh, Mehiar; Hajj, Hazem; Chehab, Ali; El-Hajj, Wassim; Kayssi, Ayman; Mansour, Mohammad M.

doi:10.1109/iwcmc.2011.5982732

Cited by 3 publications

(4 citation statements)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We also study how the different ranges of the numbers for the kernel operation affect its energy cost. An important consequence of this finding is that it provides evidence to support assumptions made in previous work [2] when exploring alternative kernel implementations for energy saving.…”

Section: Introductionsupporting

confidence: 78%

“…the number of times a kernel is executed, along with the corresponding kernel energy cost help in determining possible targets for energy optimizations. Further details about how to determine the kernels of the algorithms and how to know which kernel we should target for the highest energy optimizations are provided in our previous work [2].…”

Section: Kernelsmentioning

confidence: 99%

“…In this paper, we suggest solutions for the second and third stages. The first stage was addressed in [2]. To address the second stage, we designed a simple yet accurate algorithm that estimates the energy cost of a given kernel.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

An approach to measuring kernel energy in software applications

Dabbagh

Hajj

El-Hajj

2011

2011 International Conference on Energy Aware Computing

Self Cite

View full text Add to dashboard Cite

The large widespread of mobile devices and the increasing demand for high performance has made energy an important constraint in computer's technology. In this paper, we propose a simple software methodology that can be adapted by researchers to estimate the energy cost of kernels. Kernels are operations that can be implemented in hardware and are executed frequently in algorithms. This energy estimation helps in exploiting opportunities for energy optimizations in various algorithms. As a case study, we apply our methodology on back propagation (BP) algorithm and we estimate the energy cost of its kernels on two different architectures using: a) Simulation tools for the RISC architecture b) Physical measurement of the current and voltage using special boards for the CISC architecture. Our experiments show that the produced energy costs using our method have 5.14% error with respect to the costs of these kernels in real codes, which proves the high accuracy of our methodology. The results also show that different architectures exhibit different energy consumption when executing the same kernel which proves that alternative kernel implementations have impact on energy saving. Based on the estimated energy costs of the kernels, we further propose an energy optimization technique by using lookup tables (LUT). The proposed technique targets the exponential kernel because of its high energy consumption and replaces it with a LUT. The experimental results show a significant reduction of 99.97% in the energy consumed by the exponential kernel.

show abstract

Section: Introductionsupporting

confidence: 78%

Section: Kernelsmentioning

confidence: 99%

See 1 more Smart Citation

An approach to measuring kernel energy in software applications

Dabbagh

Hajj

El-Hajj

2011

2011 International Conference on Energy Aware Computing

Self Cite

View full text Add to dashboard Cite

show abstract

“…As a case study, we apply our suggested methodology to neural networks back-propagation (BP) (Haykin, 1998;Han and Kamber, 2006) since it is a good example of a compute-intensive application that is widely used in many domains. This paper is an extension to our previous work (Dabbagh et al, 2011), where we have tested the six-step methodology for BP, but the approach was limited to counter estimates. This paper extends the method by providing an approach for measuring actual kernel energy and then applying these measured energies instead of estimated counters.…”

Section: Introductionmentioning

confidence: 99%

An energy-aware design methodology based on kernel optimisations

Dabbagh

Hajj

El-Hajj

et al. 2014

IJAACS

Self Cite

View full text Add to dashboard Cite

The goal of this paper is to present a design methodology for developing energy aware algorithms. The key idea revolves around identifying operations called kernels, which are frequently used operations in the algorithm that can be implemented in hardware. Optimising these kernels for performance or energy would then lead to a major impact in energy saving. We propose a six-step methodology for design of energy aware algorithms. The method includes: high-level algorithm analysis, identifying high frequency kernels, determining the order of computation for each kernel via asymptotic analysis, prioritising kernels in terms of energy impact, proposing alternative implementations to the kernels that cause high energy consumption and investigating further opportunities for energy optimisation specific to the studied algorithm. We further propose a simple and efficient method for estimating a kernel's energy cost. The method was successfully tested with back-propagation (BP) neural network algorithm to identify the kernels targeted for energy optimisation. Based on the findings, we proposed several custom changes to the BP algorithms for lower energy alternatives to kernels, including options that trade off computational accuracy for higher energy saving.

show abstract