FrameBased Dynamic Voltage and Frequency Scaling for a MPEG Decoder

1
Frame-Based Dynamic Voltage and Frequency Scaling
for a MPEG Decoder

• Kihwan Choi Karthik Dantu
• Wei-Chung Cheng Massoud Pedram
• - ICCAD 02 -

2
Abstract

• DVFS technique for MPEG decoding to reduce energy
  consumption while QoS constraint.
• Frame has two parts
• Frame-Independent ( FI ) part
• Frame-Dependant( FD ) part
• ? Using DVFS in FD part
• ( higher than 50 CPU energy saving )

3
Introduction

• Key Idea
• just-enough speed to process the workload
  throughput constraints and reduce the energy
  dissipation.
• Effective DVFS Algorithm for MPEG decoding
• The future workload is accurately predicted by
  using a frame-type-based workload-averaging
  scheme at frame dependent part.
• Using frame independent part of decoding time as
  a BUFFER ZONE.
• Energy reduction of more than 50 on a
  StrongARM-1110 based platform.

4
Background(I)

• E CV2fT
• T1 ½T2, T3 ?T4
• W1? W1 75 energy saving
• W2? W2 89 energy saving

5
Background(II)

• Effective DVFS technique
• Accurate prediction of the CPU workload for a
  give task
• Prediction of CPU workload

• Moving average( MA )

• Weighted average( WA )

6
Background(III)
7
Background(IV)

• MPEG Terminology
• MPEG2 video stream consists of three frame types
• I-Frame( Intra-coded ), P-Frame( Predictive-coded
  ), and B-Frame( Bi-directionally-coded )
• Sequences of frames are grouped together to form
  a Group of Pictures( GOP ) which have 1215
  frames.
• Step of decode
• FD part Parsing ? IDCT ? Reconstruction
• FI part Dithering

8
Background(V)

• Prior Work
• Relationship between the decoding time and the
  data size
• ? cannot obtain code size of frame before
  starting IDCT step
• Using macro block in a frame
• Decoding time of second part of macro block is
  predicted based on the decoding time and the code
  size.
• ? workload prediction error by raising the supply
  voltage level during a memory-intensive step as
  opposed to a CPU-intensive step.

9
Algorithm(I)

• TDecoding TFD TFI EDecoding EFD EFI
• FI part Memory intensive( dithering part )
• ?TDecoding ?TFD ?TFI ?EDecoding ?EFD

10
Algorithm(II)
11
Algorithm(III)

• Error ratio
• I-frames, P-frames less than 10
• B-frames less than 20

12
Algorithm(IV)
13
Implementation(I)

• Mpeg_play MPEG player program( berkeley Univ. )
• ? insert calculating MA, clock speed and voltage

14
Implementation(II)
15
Result

• Power consumption while playing MPEG2 at fps 1
• Without DVFS 206MHz
• 2.94W _at_ 6V ( 0.49A )
• With DVFS 89 206MHz
• 2.46W _at_ 6V ( 0.41A )
• 53 reduce CPU energy
• ( CPU consume about 30 of the total energy )

16
Conclusion

• DVFS
• Each frame type is handled individually for more
  accurate decoding time prediction
• Error ratio is less than 20 in all frame type
• Using FI step is buffer zone
• Covered misprediction in FD step workload
• FI step is memory intensive
• ? CPU voltage increase does not affect the
  energy consumption
• More than 50 energy saving