Performance Analysis and Power Estimation of ARM Processor

1
Performance Analysis and Power Estimation of ARM
Processor

• Team
• Ajayshanker Krishnamurthy
• Swathi Tanjore Gurumani
• Zexin Pan
• Project Advisor
• Dr.Alexander Milenkovic

2
Agenda

• Overview
• Tools Used
• Performance Analysis - Results
• Power Estimation - Results
• Conclusion

3
Overview
Target Binaries
Performance Metrics
Compile
Benchmarks
Exe
Power Dissipated
Performance Metrics

• MiBench
• SimpleScalar
• PowerAnalyzer

4
Tools Used

• Benchmarks
• Critical part of design process due to
  performance based designs
• Embedded BenchmarksFastest growing market
  segment in the u-processor industry
• MiBench (University of Michigan)
• Free, commercially representative embedded
  benchmark suite
• Set of 35 embedded applications of six categories
  
• Automotive and Industrial Control, Network,
  Security, Consumer Devices, Office Automation and
  Telecommunications
• Security Algorithms Rijndael, Blowfish, Sha, Pgp
• Small data set represents a light-weight, useful
  embedded application
• Large data set provides a more stressful,
  real-world application

5
Tools Used

• SimpleScalar (Born 1982, _at_ University of
  Wisconsin)
• Provides an infrastructure for simulation and
  architectural
• modeling
• Can model a variety of platforms
• - unpipelined processors to detailed micro
  architectures
• Suited to the needs of researchers and
  instructors
• - meets the critical requirements
  Performance, Flexibility Detail
• Supports popular instruction sets -Alpha,
  Power PC, x86 ARM
• Baseline simulator models
• Sim-safe, Sim-fast, Sim-cache, Sim-profile,
  Sim-bpred,
• Sim-fuz, Sim-outorder

6
Tools Used

• PowerAnalyzer SimpleScalar-Arm Power Modeling
  Project
• Joint venture of U Michigan U Colorado
• Estimator that allows power/performance
  trade-offs to be examined
• Tightly Coupled with SimpleScalar Toolset for ARM
• Gives Power dissipation for each component
  individually
• Switching, Internal Leakage
• Can be configured based on two models
• Analytical Empirical

7
Measurement Methodology

• Configured for Current (SA 110) and Next (PXA
  250) generation
• Input Same dataset (gt3M) for all algorithms to
  achieve fair comparison and reliable result
• Output raw data related to performance and power
  consumption are obtained from PowerAnalyzer
  report
• Data Processing (digesting) and visualizing

8
Performance Analysis

• Configured Sim-outorder to represent current and
  next generation of embedded processors
• Intel SA-110 for current generation
• 32 bit general purpose micro processor
• On chip data cache(16K),instruction cache(16 K)
  and MMU
• Used in PDAs, Smart phones, digital cameras etc.
• Intel PXA-250 for next generation
• High performance Intel Xscale core
• On chip data cache(32 K),instruction cache(32
  K),branch target buffer and MMU
• Used in Multimedia Applications

9
Configuration
Current Next
I Fetch Q size 2 4
Branch Pred. Not Taken Bimod
I Issue Width 1 1
Cache dl1 163232 323232
Cache il1 163232 323232
TLB itlb 1640964 1640964
TLB dtlb 3240964 3240964
10
Results
11
Results
12
Results
Current generation predictor Not Taken Next
generation predictor Bimod
13
Results
14
Why use power as performances criteria?

• T. Mudge, Power A first class design
  constraint, Computer, vol. 34, no. 4, April
  2001, pp. 52-57
• Limiting power consumption is critical,
  particularly in portable and mobile applications
  such as cell phone and laptop due to limit
  battery life
• One of the major markets of ARM is portable and
  mobile products

15
Power Estimation

• Measurement Methodology
• ARM simulator power measurement tools
  PowerAnalyzer 1.1 from UMICH
• Configured for Current (SA 110) and Next (PXA
  250) generation
• Input Same dataset (gt3M) for all algorithms to
  achieve fair comparison and reliable result
• Output raw data related to performance and power
  consumption are obtained from Power Analyzer
  report
• Data processing (digesting) and visualizing

16
Power Estimation

• Difficulties using PowerAnalyzer
• Report gives power consumption for every ARM
  component, but no unit!
• Since all these numbers are huge, we have
  difficulties figuring out what they mean ??

17
Power Estimation
18
Power Estimation
19
Power Estimation
20
Conclusion

• The performance gain in next generation of
  processors is offset by the increase in power
  consumption. Intel Xscale almost doubles the
  power consumption with about 10 performance gain
  over SA- 110
• The next generation of processors with larger
  caches improve performance
• The bimodal branch predictor greatly reduces the
  number of miss predictions
• Power consumption not only depends on hardware
  architecture and system configuration (system
  clock,etc.), but also heavily relies on Benchmark
  and input dataset

21
Thank You

• Questions