Data Wordlength Reduction for Low Power Signal Processing

  • Slides: 10
Download presentation
Data Wordlength Reduction for Low. Power Signal Processing Software Kyungtae Han May 3 rd,

Data Wordlength Reduction for Low. Power Signal Processing Software Kyungtae Han May 3 rd, 2004 The University of Texas at Austin May 3, 2004

Low-Power at Software Level § Portable wireless computing demands minimizing power dissipation due to

Low-Power at Software Level § Portable wireless 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 wordlength May 3, 2004

Objectives and Problems § Objectives § Power analysis of digital signal processing blocks for

Objectives and Problems § Objectives § Power analysis of digital signal processing blocks for data word length in low-power software § Software level power minimization with reducing data wordlength while not changing the hardware structures § Problems § How much can shorter data wordlength reduce power consumption? § Need formulation of power consumption according to reduced data wordlength for optimum design May 3, 2004

Power Consumption § Average power consumption § Switching power consumption May 3, 2004

Power Consumption § Average power consumption § Switching power consumption May 3, 2004

Multiply Unit for Digital Wireless Transceivers § Multiply unit is usually a major source

Multiply Unit for Digital Wireless Transceivers § Multiply unit is usually a major source of power consumption in typical DSP applications § Many digital blocks are required multiply unit for digital wireless communications § Digital filters, equalizers, FFT/IFFT, digital down/up converter, etc. § Multiplier core in DSP § TI 64 uses a Wallace tree with 3 -2 compressors § TI 62 uses a Booth multiplier May 3, 2004

Data Wordlength Reduction § Fixes the size of multiplication § Decreases input data wordlength

Data Wordlength Reduction § Fixes the size of multiplication § Decreases input data wordlength § Signed right shift (Least significant bit side) § Move toward LSB (i. e. left shift) § Arithmetic right shift (if negative) § Truncation (Most significant bit side) § Make LSB side to be zero values May 3, 2004

Example of Array Multiplier Transition Counts § Array multiplier with data flow model §

Example of Array Multiplier Transition Counts § Array multiplier with data flow model § § § 16 x 16 bit array multiplier 10, 000 random data Node transitions counted Without delay information Data Flow Modeling § 16 x 16 bit array multiplier transition counts Input data 16 bits Max 3152 Min 383 May 3, 2004 Std 319 Mean 1863

Wallace Multiplier Transition Counts § Simulation § Discrete Event Simulation: Verilog Compiler Simulator §

Wallace Multiplier Transition Counts § Simulation § Discrete Event Simulation: Verilog Compiler Simulator § Average transition counts of all gate output § All gates are assumed to have a unit gate delay § Variable size Wallace multiplier Wallace Size (bits) Avg. transition counts 8 16 670 May 3, 2004 4177 32 22021

Wallace Multiplier Transition Counts § Fixed-size Wallace multiplier (32 x 32 bits) § n

Wallace Multiplier Transition Counts § Fixed-size Wallace multiplier (32 x 32 bits) § n 2 rate reduction in truncation method Average transition counts for data wordlength reduction Input wordlength Truncation Signed Right Shift 32 22021 16 6138 28634 8 1383 31364 4 404 32996 May 3, 2004

Conclusion and Future Work § Conclusion § Data wordlength reduction in MSB side decreases

Conclusion and Future Work § Conclusion § Data wordlength reduction in MSB side decreases transition counts and power consumption § In Wallace multiplier, transition counts decrease n 2 rate Optimum wordlength Error (d) Power (p) [1/performance] § Future work § Find optimum wordlength for power and precision § Comparison with Booth Radix-4 § Apply for digital signal processing blocks May 3, 2004 Wordlength(n)