Processor Frequency Setting for Energy Minimization of Streaming Multimedia Application

1
Processor Frequency Setting for Energy
Minimization of Streaming Multimedia Application

• by A. Acquaviva, L. Benini, and B. Riccò, in
  Proc. 9th Internation Symposium on
  Hardware/Software Codesign, Apr. 2001.

2
Agenda

• Introduction
• Power optimization model derivation
• Power optimization algorithm
• Experimental results
• Conclusion

3
Introduction

• With technology enhancement, multimedia
  capabilities are being added to handheld devices.
  Examples
• Picture taking
• MP3 audio playback
• Video playback
• Audio recording
• A new problem arises ? power management

4
Introduction

• Power management options
• Shut down devices when in idle mode
• Problem Background tasks have to be stopped as
  well
• Better approaches clock frequency and voltage
  regulation
• Lowers system speed in idle states
• Reduces LCD display brightness

5
Introduction

• Real-time media streaming applications
• Retrieve stream data from off-CPU interface (e.g.
  discs, memory cards)
• Process data (e.g. decoding, decompression)
• Deliver processed data to output interface (e.g.
  display, speakers)

6
Power Optimization Model Derivation

• System settings
• The CPU must communicate with relatively slower
  I/O interfaces
• Clock frequency can be adjusted by software
• Frame-based media (e.g. MP3 audio, MPEG video)

7
Power Optimization Model Derivation

• Power consumption
• Energy per frame
• V supply voltage, C switched capacitance, f
  CPU clock frequency, Tf frame processing
  time, Nf number of cycles to process frame, t
  cycle time

8
Power Optimization Model Derivation

• Due to speed difference between the CPU and
  external hardware
• where Nidle is a non-decreasing function of f
• We now have

9
Power Optimization Model Derivation

• To satisfy real-time synchronization constraint
• Tmax maximum allowed time for a frame (e.g.
  1/30s in 30fps movie)

10
Power Optimization Model Derivation

• In words, target of power optimization is to
  reduce the term Nidle in
• under the constraint

11
Power Optimization Model Derivation

• This technique is more effective if application
  requires much memory access
• However, delay and energy spent for frequency
  adjustment must also be considered

12
Proposed Algorithm

• Define three curves FRB(f), FRA(f), FRW(f), the
  best-case, average-case and worst-case frame rate
  at f.
• Compute curves for all bit rate and sampling rate
  values and obtain FRoB(f), FRoA(f), FRoW(f)
• Compute FRreq by
• Nsample samples per frame, fixed at 576 for MP1
  and MP1 phase 2

13
Proposed Algorithm

• Normalize the curves FRoB(f), FRoA(f), FRoW(f) by
  FRoA(fmax) from a pre-calculated look-up table
• Intersect FRreq with three curves to obtain fmin,
  fav and fmax.

14
Proposed Algorithm

• CPU frequency can be set to
• fmin if we find constantly frames processed
  faster than the average rate
• fav if we want continuous playback, with some
  buffering storage for decoding rate jitter
• fmax to guarantee real-time performance on a
  frame-by-frame basis
• Greater than fmax if the processor is not
  dedicated to the application only

15
Experimental Results

• Energy consumption per frame

16
Experimental Results

• Energy consumption for 16KHz, 16KBit/s audio

17
Experimental Results

• Frequency setting

18
Example Calculation

• An audio stream of 16KHz, 16KBit/s
• Without any optimization, Ef 10.989mJ
• FRreq 16000 / 576 27.78 fps
• FRA(fmax) 65.36 fps
• Normalized FR 27.8 / 65.36 0.425 fps
• fmin 85.7MHz, fmax 106.7MHz
• Choosing fmax, Ef 8.9mJ ? 19 energy reduction

19
Conclusion

• Frequency-energy relationship is derived
• An energy optimization algorithm is proposed
• Experiment shows dramatic save in power
  consumption