Data Word Length Reduction for LowPower DSP Software

1
Data Word Length Reduction for Low-Power DSP
Software

• Kyungtae Han
• March 24, 2004

2
Low-Power Software

• Portable computing demands minimizing power
  dissipation due to limited power
• Minimizing power consumption
• Reduce supply voltage
• Decrease switching activity
• Software can reduce power consumption
• Ordering of operation
• Changing of number representation
• Reducing of data word length

3
Objectives and Problems

• Objectives
• Power analysis of multipliers for variable word
  length for low-power software
• Software level power minimization with reducing
  data word length while keeping hardware structure
• Problems
• How to analyze and measure power consumption for
  low-power software
• How much can shorter data word length reduce
  power consumption?

4
Power Analysis of Embedded Software

• Tiwari, Malik, and Wolfe, 1994
• Measuring power of instruction sets
• Formulated an instruction level power model
• Lee, Tiwari, Malik, and Fujita, 1997
• Found multiplier in the multiply and accumulate
  (MAC) unit is usually a major source of power
  consumption in typical DSP applications
• Observed the wide current variation of MAC
  instructions according to the two values being
  multiplied in MAC unit

5
Word Length for Low Power

• Chandrakasan et.al, 1995
• Showed that the word length affects all key
  parameters of a design, including speed, area,
  and power
• Switching activity is used for power estimation
• Erdogan and Arslan, 1996
• Demonstrated power reduction with different
  coefficient word lengths in a modified DSP
  processor
• Used data bus and coefficient bus for
  multiplication

6
Preliminary Results

• Array multiplier with SDF
• 16-bit array multiplier
• Different data word lengths
• 10,000 random data
• Node transitions counted
• Transition counts

7
Summary of Talk and Plans

• Summary
• Power analysis of embedded software
• Word length for low power
• Short data word length reduces power consumption
• Plans
• Formulate power model for different input word
  lengths being multiplied
• Find tradeoff with power consumption and
  precision
• Different architecture Booth Radix-4, Wallace
  multipliers
• Apply for digital FIR filter implementation using
  SDF domain

8
Backup Side Software Level Power

• Power consumption