1 / 50

Being Globally Energy-Aware in DSP Systems Design

SIPS 2009 Tampere. Being Globally Energy-Aware in DSP Systems Design. Chong-Min Kyung KAIST. Outline. Introduction Difficulties of Energy Optimization Many Global Views Dealing with Uncertainties Complexity, Distortion and Rate Optimization Conclusion. Outline. Introduction

tambre
Download Presentation

Being Globally Energy-Aware in DSP Systems Design

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. SIPS 2009 Tampere Being Globally Energy-Aware in DSP Systems Design Chong-Min Kyung KAIST

  2. Outline • Introduction • Difficulties of Energy Optimization • Many Global Views • Dealing with Uncertainties • Complexity, Distortion and Rate Optimization • Conclusion

  3. Outline • Introduction • Energy, as a new metric in DSP design • Why Globally Energy-Aware? • Difficulties of Energy Optimization • Many Global Views • Dealing with Uncertainties : • Complexity, Distortion and Rate Optimization • Conclusion

  4. Traditional DSP Design ; fast, accurate, and cheap • Performance (Throughput, Latency) • vs. energy • Accuracy of computation • vs. accuracy of transmission • Cost of manufacturing • vs. cost of operation • vs. cost of maintenance • vs. cost of environment

  5. Weight shift in DSP Design toward Energy • Performance (Throughput, Latency) • vs. energy consumption • Accuracy of computation • vs. accuracy of transmission • Cost of manufacturing • vs. cost of operation • vs. cost of maintenance • vs. cost of environment

  6. Three Issues on introducing Energy • Issue of design target shift : Energy or Information, which is cheaper to deploy? • Issue of currency ratio between energy and information : Dealing with abundance of unnecessary data by maximizing ‘information per energy’ • Issue of global objective function

  7. Energy ;a new metric in DSP Design • Issue 1 : Energy vs. Information, which is cheaper to deploy? • Energy is consumed for information deployment ; physical endeavor is needed for energy deployment. • Industrial Revolution Age : Iron (information) is processed where coal (energy) is cheaper. • ‘Energy consumption’ as a new design metric

  8. Energy ;a new metric in DSP Design • Issue 2 : excessive unnecessary data • save energy by claiming performance /accuracy only as much as needed • Why waste energy for unnecessary information?

  9. Energy ; a new metric in DSP Design • Issue 3 : Composite cost function • Consider energy for communication as well as processing • Do not process what cannot be sent. • Process information such that the composite cost is minimal. • Price of energy depends on the its location as well as its form.

  10. Why Globally Energy-Awarein Design ? • Various forms of considering energy consumption during the design process • Low-energy design • Energy-scalable design • Energy-aware design : run-time, self-adjusting

  11. WhyGloballyEnergy-Aware in Design ? • Global vs. Local ; Questions to ask • Looking for a long-term solution? • Is Design Platform/Approach (for global energy awareness) needed?

  12. Cost of communication vs. Cost of processing • Various forms of data for transmission • Raw data • Compressed data • Feature-extracted data • Recognition/decision

  13. Cost of communication vs. Cost of processing • ‘Use Energy as much as Information processing requires’ vs. ‘Process I such (where, how) that E consumption is minimal’ • Trade-off between • amount of cost (energy, area, latency, design time) and • amount of profit (saved energy, bandwidth) Cost of Energy Communication Computing Raw Compressed Feature Decision Form of Information

  14. Outline • Introduction • Difficulties of Energy Optimization • Many Global Views • Dealing with uncertainties : • Complexity, Distortion and Rate Optimization • Conclusion

  15. Difficulties of Energy Minimization • Energy minimization problem • Contexts as constraints • Performance (throughput, latency) • Cost (chip area, code size) • Output quality (error, distortion) • Variables ; too many • Uncertainty, variability

  16. Difficulties of Energy Minimization • Too many variables in different function blocks affecting the objective function channel coding block mode, search range resolution, frame rate Transmitter ECC H.264 encoder + alarm generator Camera Storage

  17. Difficulties of Energy Minimization • Dealing with uncertainty/variability • Manufacturing tolerance • Input statistics, run time variation • Environment • Temperature • Supply voltage • Demand on QoS varies (is often negotiable) according to the remaining energy

  18. Outline • Introduction • Difficulties of Energy Optimization • Many Global Views • ChipLife Cycle View • System Operation Cycle View • Chip Design Cycle View • System Architecture Option View • Circuit-Battery System View • Dealing with uncertainties : • Complexity, Distortion and Rate Optimization • Conclusion

  19. Chip Life Cycle View Dirty gas, chemical, used mask out-signal design fabrication test operation designer ‘s time power(energy) in-signal test vector clean gas, chemical, mask Energy

  20. System Operation Cycle View • Control variables : spatial/temporal resolution, precision, parameter of algorithm, … Stage i Sampling ADC Processing Storage Transmission ti : exec. time ei : energy di : distortion cost Performance, or output quality Minimize E + 1D + 2T

  21. Total Cost Minimization Problem Sampling ADC Processing Storage Transmission Stage i SIPS, Tampere

  22. Graded Cascade System • How to decide ei and di? • ei-1 • di-1 • ti-1 ei di ti • ei+1 • di+1 • ti+1 D-1 (PSNR) Energy Time • ei

  23. Graded Cascade System : examples • Cache/memory hierarchy • Alarm control system • Any hierarchical human organization

  24. Chip Design Cycle View • High-level : hierarchical view 1) General-purpose case • Architecture/compiler co-design 2) Application-specific case (platform-based) • Energy-aware algorithm/architecture co-design • Low-level : both general- and application- specific case • DVFS (Dynamic Voltage Frequency Scaling) • Power gating

  25. a video surveillance sensor node d t e Silent Resolution ↑ Bits per pixel ↑ Resolution ↑↑ Bits per pixel ↑↑ Frame rate ↑↑ Frame rate ↓ Bits per pixel ↓ Inter prediction RDO off DVFS Inter prediction RDO off DVFS Intra prediction RDO off Power gating Inter prediction RDO on Fast mode decision

  26. System Architecture Option View • Simulation for minimal energy consumption with performance/deadline constraint based on behavior/energy model for each implementation of each block • Single / multi-core (with accelerator ?) • Interconnection architecture • bus / network • Memory system • Implementation • SoC • SiP (system in a package) • 3D IC (3-dimensional IC)

  27. Circuit(Load)-Battery System View • Considering overall energy consumption from a energy provider (battery) point • Battery-aware task scheduling • DC/DC converter efficiency variation

  28. Battery-aware task scheduling • Battery lifetime is dependent on the load current pattern. Charge recovery Energy consumption at load is the same!! But, charge consumption at battery is different.  B case is better than A case w.r.t. battery charge consumption point. A Battery fail B

  29. Reducing DC/DC converter loss • DC/DC converter loss • Function of input voltage and output current Battery voltage degradation  Converter efficiency improvement Alkaline AA cell 3 series voltage *TI TPS60100 efficiency (%)

  30. Outline • Introduction • Difficulties of Energy Optimization • Many Global Views (Big Pictures) • Dealing with uncertainties : • Uncertainty-Aware Energy Optimization • Complexity, Distortion and Rate Optimization • Conclusion

  31. Energy Optimization with Uncertainties • Technology≥ 130nm • Switching power consumption  Dynamic supply voltage/frequency scaling • 65nm ≤ Technology≤ 90nm • Sub-threshold leakage current • Adaptive body biasing and power gating • Technology ≤ 45nm • Gate leakage current  High-k/metal gate • Variability  adaptive body bias control

  32. Where Uncertainty Comes from (1) • PVT variation ; from device, environment • Process • Voltage • Temperature [Source: Scott Thomson, U Florida]

  33. Where Uncertainty Comes from (2) • Workload variation : from application • Data dependency # of loop iterations, etc. • Control-flow dependency : if/else statement, etc. • Architectural dependency : cache hit/miss, DDR page miss, etc. [source: MPEG4 for decoding movie clip, Harry Potter, using LG XNOTE LW25]

  34. Uncertainty-Aware Optimization • How to deal with the workload variation • Conservative approach: worst-case prediction • Aggressive approach: average value prediction • Stochastic approach Prob. Workload Predicted Workload Requirements (e.g., deadline) Environment (e.g., temperature, hardware system) Stochastic workload predictor

  35. Workload Uncertainty-Aware DVFS • DVFS (dynamic voltage frequency scaling) • Case study: MPEG4, H.264 decoder • Source of workload variation • Frame type: I/P/B-frame • Amount of motion • Configuration: quantization, frame rate, image size, etc. • Stochastic DVFS • Set voltage/freq. level is set based on predicted workload.

  36. Example Motion Compensation Deblocking Filter High freq. High freq. Conservative Approach Freq. Aggressive Approach Stochastic Approach Time

  37. Outline • Introduction • Difficulties of Energy Optimization • Many Global Views (Big Pictures) • Dealing with uncertainties : • Complexity, Distortion and Rate Optimization • Conclusion

  38. CRD Optimization • Cost • Rate • Distortion

  39. R, D first : Rate and Distortion • Definition • Rate : Energy needed for storage and transmission • Distortion: 1/SNR • How to reduce distortionat low data rate? • Transmission : Speed, Accuracy • Storage : Space, Accuracy

  40. Compression? • What is compression? • Reduce the load with least loss using as little resource as possible • in transmission : bandwidth, energy • in storage : storage capacity requirement, energy • Compression in lower level • Anti-aliasing filtering • Sampling • A/D conversion

  41. R + D, or RD as objective function • Sampling ← kind of compression (in time) • Distortion S·∆t (S: signal slope) • Rate (number of sample) = • R · D = S·T uncertainty product • Non-uniform sampling : how to assign sampling points ∆ ∆ = = s∆t s : signal slope ∆t Slope-aware sampling Variable-rate sampling i i+1

  42. A/D conversion : kind of compression • Figure of merit • Distortion = Rate ( # of bits/sample) = N • R·D = ∆ = D + ∆ investment in chip area power(energy) design time R

  43. CRDOptimization • Code compression • Code expansion Compression Code Cost - Execution time - Energy - Chip area SNR / Rate Handling energy ECC Channel/Source Coding Code Cost • Execution time • Energy • - Chip area Error rate-1/ code size Handling energy

  44. CRDOptimization • CRD model for chip design Design Cost Chip - Cost (design man months) Performance / Area or Power

  45. CDOptimization : test, recognition • CD model decision (rate is constant) Test Tested chip Cost (Testing time) Error-1 Pattern recognition Decision Cost (processing time) Error-1

  46. CRD Optimizationin H.264/AVC Candidate modes Candidate modes High-complexity RDO Entropy coding Generate bit sequences Low-complexity RDO Choose the min. J Choose the min. J’ Entropy coding Entropy coding Generate bit sequences PSNR  Entropy coding Low-complexity RDO (pseudo-rate) High-complexity RDO rate

  47. CRD Optimizationin H.264/AVC Candidate modes 2) 1) Reducing # of cand. modes 2) Reducing entropy coding time thru an efficient rate model High-complexity RDO Entropy coding Low-complexity RDO Choose the min. J Entropy coding 1) Generate bit sequences PSNR  Entropy coding High-complexity RDO rate

  48. CRDO in Wireless Transmission • Encoding power vs. Transmission power maximum encoding power Power Total power Transmission power Optimal encoding power Encoding power Minimum encoding bit rate Optimal bit rate Bit-rate

  49. Outline • Introduction • Difficulties of Energy Optimization • Many Global Views (Big Pictures) • Dealing with uncertainties : • Complexity, Distortion and Rate Optimization • Conclusion

  50. Conclusion • Energy consumption must be considered • In the design process • In a ‘globally energy-aware’ manner • Being ‘Globally Energy-Aware’ includes considering • Life cycle, design cycle, operation cycle of DSP chip, as well as • Dealing with variability, and finally… • Run-time CRD optimization • Complexity (computing energy) • Rate (transmission energy) • Distortion (quality of processing results)

More Related