Energy consumption represents an impor- tant issue with limited and embedded devices. Such devices, e.g. smartphones, process many images, both to render the UI and for application specific purposes.
We aim to evaluate the energy consumption of different image encoding/decoding algorithms.
We run a series of experiments on a ARM based platform and we collected the energy consumed in performing typical image encoding and decoding tasks.
We found that there is a significant difference among codecs in terms of energy consumption. Most of the energy con- sumption relates to the computational efficiency of the algorithm (i.e. the time performance) though the type of processing and the algorithm may affect the average power usage up to 37%, thus indirectly affecting the energy consumption.
We conclude that JPEG compression is significantly more energy efficient than PNG both for encoding and decoding. Further studies should focus on the additional features that affect energy consumption beyond computational complexity.
4. Plan
Object: image codec algorithms
Purpose: assessing differences
Focus: energy consumption
Context: Raspberry PI
5. Research Questions
• RQ1: Do different codecs consume different amounts
of energy for encoding/decoding images?
• Metric: total energy
• RQ2: How closely are energy consumption and
computational performance correlated?
• Metrics: Enegy vs. time, and Power
10. RQ1: codec consumption
Task ☞ DECODING ENCODING
CODEC☟
Energy Time Energy Time
PNG 1.78 9.6 6.16 39.0
JPEG, Q:10 0.94 4.5 1.05 7.6
JPEG, Q:40 0.76 3.4 0.84 2.7
JPEG, Q:80 0.68 2.9 0.78 2.7
Average values over the three images, values for five repetitions of the task
12. RQ2: Energy and complexity
0.0
2.5
5.0
7.5
10.0
0 20 40
Time [s]
Energy[s]
Process
Decode
Encode
E = P · t
13. Mean Power vs. Total Energy
A
B
C
A
B
C
DecodeEncode
jpg:Q80 jpg:Q40 jpg:Q10 png:LL
Codec:Quality
Image
0.15
0.17
0.20
0.23
Power A
B
C
A
B
C
DecodeEncode
jpg:Q80 jpg:Q40 jpg:Q10 png:LL
Codec:Quality
Image
2.5
5.0
7.5
Energy
14. Key Findings
• Decoding a PNG image consumes ~2 times as much
energy as for a JPG image
• ~6 times for encoding
• Energy consumption is strongly correlated to time
• Other factors affect the energy marginally
• The average power consumption variation among
algorithms is ~10%
15.
16. Open Questions
• Can we identify algorithms that with a similar
computational complexity consume less power?
• What are the features in software that can affect
power consumption?