1. Bhojan Anand‡, Karthik Thirugnanam†, Jeena Sebastian‡, Pravein G. Kannan‡, Akhihebbal L. Ananda‡, Mun Choon Chan‡ and Rajesh Krishna Balan† ‡ National University of Singapore (NUS) and † Singapore Management University (SMU) 1

2.  Problem: Display draws significant phone power  Key Challenge: No loss in end user experience 2 Display 45-50% Network 35-40% CPU 4-15% Measured on HTC Magic while streaming a Youtube Video

3.  Dynamically adjust image brightness and LCD display backlight levels  50-70% display power savings with no significant user experience impact 3

4. Key problem and Solution Background System Design Evaluation Discussion/Future Work 4

5.  LCD displays have two components:  Power consumed mostly by the Backlight  Thus brightening the image, and darkening the backlight saves power. 5 Backlight - Provides light, and consumes power LCD Panel - Filters light based on image to be displayed

6.  Method 1: Naively dim the display ◦ Creates visible artifacts (flicker, brightness loss, etc) ◦ Especially noticeable in high frame rate applications 6

7.  Method 2: Compensate with increased brightness ● Linearly apply same transform to entire image  Leads to saturated images 7

8.  Method 2: Compensate with increased brightness ● Non linear approaches prevent saturation but cause contrast loss  Our solution uses this approach intelligently 8

9.  Gamma Correction, or gamma, is a tone mapping function used to brighten scenes ● Very Low Saturation relative to linear ● Low computational overhead 9 Before After: Gamma 2

10. 10 Original Image after Gamma Increase (gamma=2) Image after Gamma Increase and backlight reduction

11.  Games are popular and resource intensive ● Extremely high frame rates ● Flicker and brightness changes very noticeable to users  We use two representative games ● Quake III – Commercial First Person Shooting (FPS) game ● Planeshift – Massively Multiplayer Role Playing Game (MMORPG) 11

12. Key problem and Solution Background System Design Evaluation Discussion/Future Work 12

13.  Ultimate goal: Save significant power with no loss in end user experience  Challenge 1: Understanding the relationship between the backlight intensity, gamma, image brightness, and the power consumed  Challenge 2: Identifying human thresholds for brightness compensation  Challenge 3: Dynamically applying the solution 13

14.  Relationship found to be linear.  No other major contributing factors 14 HTC Hero Laptop (W500) HTC Magic

15.  Perceived brightness kept constant  Non linear compensation necessary  Useful gamma range is 1 to 4 15

16.  Obtained via small user study ● 5 postgraduate students  Each user shown a range of images ● Covered a full range of brightness  For each image, users had to boost gamma to obtain two quality thresholds ● Described in next slide ● Tool provided boosted gamma at.1 intervals with automatic backlight compensation 16

17.  Conservative: Image quality comparable to original  Aggressive: Image quality is affected but acceptable 17

18. 18 Start Calculate Average Brightness of last X Samples. Is there a change? Mode + Brightness -> Gamma & Backlight Leave Settings as it is Sleep Thread for Y ms Yes No

19.  Test platform ● Most of the evaluation on Laptop ● Prototype Mobile implementation available (demo)  Objective Analytical Experiments ● Power measurements ● Measured power saved in different modes  Perceived User Impact ● Large scale user study (60 users) with Quake III ● Measured perceived quality loss in different modes 19

20. 20 Unmodified Conservative Two bounding modes tested but omitted for simplicity

21.  A recorded trace used to measure power 21

22.  Large Scale User study ● 60 Singapore Management University undergrads. ● 34 Male and 26 Female students with differing background and game experiences  Participants trained on an unmodified version of the game ● They then played the 3 different versions of the game ● Play order randomized with recalibration at every step 22

23.  Users rated each version by 6 criteria ● Covered different quality dimensions 23 Bad Good Strongly Agree = Strongly Disagree = Neutral =

24.  Difference between Aggressive and Conservative significant 24

25.  Difference betw. Conservative and Default 25

26.  Conservative (Dynamic Conservative) ● High Quality  Perceived quality comparable to default. ● Significant Savings – 49%  Aggressive (Dynamic Aggressive) ● High Power Saving – 68% ● Acceptable Quality 26

27.  OLED Displays ● Do not use Backlight ● Power consumption depends on displayed content ● Algorithm needs to be rethought  Other Limitations & Future Work ● Initial User studies were small scale, and in a controlled environment ● Power Measurements could be more accurate, esp for mobile phone ● Evaluation of mobile implementation 27

28.  General systematic approach to save power on LCD screens  Identified key parameters determining the quality vs power savings tradeoff  Implemented and tested with 60 end users  System achieves significant power savings with minimal overhead and quality loss 28

29. Any Questions 29