1. Basics of Energy-Efficient Design Lin Zhong ELEC424, Fall 2010 1

2. Outline General concepts Energy-saving mechanisms Integrated circuits (IC) – Processors Wireless interfaces 2

3. What is energy efficiency? Power Energy (Power * delay) Battery lifetime Energy * delay 3

4. Source of energy consumption IC – Computing/Switching – Radiation (wireless) Discrete components Display 4

5. Energy characterization Equipment setup 5 Power Supply - + Target 0.1 ohm Sense Resistor Voltage Sampling Device

6. 6 Differential measurement Extra energy/power consumption of an event obtained through differential measurements Extra energy consumption for writing “x” Write “x” with stylus/touchscreen 0 0.4 0.8 1.2 1.6 00.511.5 Time (s) Power (W) 0 0.4 0.8 1.2 1.6 00.511.5 Time (s) Power (W)

7. Rule No. 1 Focus on the big one! – Amdahl’s Law Reduction of the power of α % of the system by p% leads to α∙p % reduction of the total power 7 α

8. Rule No. 2 Minimize static energy consumption – IC consumes static power when it is merely powered 8

9. Rule No. 3 Minimize activity – When not doing things useful Turn it off Stop the clock Check the manual for power-saving modes Be aware of state transition overhead – Interrupt-driven instead of polling 9

10. Rule No. 4 Don’t forget parasites – More integrated solution leads to lower energy 10

11. Processors Dynamic voltage scaling Power-saving states – Clock gating – Power-down different subsets of components As-fast-as-possible or As-slow-as-possible? 11

12. Power consumption of processing Dynamic power 12

13. Busy power vs. delay vs. energy Analysis and Design of Digital ICs, Hodges et al 13

14. Core 2 Duo for example Intel® Core™2 Duo processor – T7800 at 2.6GHz – T7700 at 2.4GHz available on Thinkpad T61p – 0.75-1.35V, 35Watts Intel® Core™2 Duo Low Voltage – L7500 at 1.6GHz available on Thinkpad X61 – 0.75-1.3V, 17Watts Intel® Core™2 Duo Ultra Low Voltage – U7500 at 1.06GHz available on Dell D430 – 0.75-0.975V, 10Watts 14

15. Switching energy e=1/2∙C ∙V 2 Switching power P= b∙C ∙V 2 = a∙C ∙V 2 ∙f 15

16. Given workload L and deadline T L measured by # of CPU cycles Clock speed f ≥ L/T Time to finish: t = L/f Energy to finish: P ∙ t= a∙C ∙V 2 ∙f ∙t= a∙C ∙V 2 ∙L 16

17. Effect of lower clock speed (f) Power consumption P= a∙C ∙V 2 ∙f Energy consumption E=P ∙ t= a∙C ∙V 2 ∙f ∙t= a∙C ∙V 2 ∙L 17

18. Effect of lower supply voltage (V) Power consumption P= a∙C ∙V 2 ∙f=k∙V 3 =x∙f 3 Energy consumption E=P ∙ t= a∙C ∙V 2 ∙f ∙t= a∙C ∙V 2 ∙L Maximum clock speed f= b∙V 18

19. Given workload L and deadline T single processor The processor can run at any frequency (voltage) – f= b∙V The processor can be complete off when work is done (zero power when idle) To minimize energy consumption, at which frequency should the processor run? – f ≥ L/T (in order to meet the deadline) – E=P ∙ t= a∙C ∙V 2 ∙f ∙t= a∙C ∙V 2 ∙L – f=???? 19

20. time f T f 1 =L/T f 2 =L/(T/2)=2f 1 20

21. time P T P 1 =x∙f 3 P 2 =2 3 P 1 21

22. Given workload L and deadline T M processors The workload can be divided without overhead: L = L 1 +L 2 +…+L M (L ≥ Li≥0) To minimize energy consumption, at which frequency should processor i run? – f i = L i /T and V = u ∙ L i – E i = a∙C ∙V 2 ∙L i =w∙L i 3 22

23. Given workload L and deadline T M processors The workload can be divided without overhead: L = L 1 +L 2 +…+L M (L ≥ Li≥0) To minimize the TOTAL energy consumption, how should the workload be allocated? – E= E 1 +E 2 +…+E M = w∙L 1 3 +w∙L 2 3 +…+w∙L M 3 – = w(L 1 3 +L 2 3 +…+L M 3 ) 23

24. From high school [(a+b)/2] 2 ≤ (a 2 +b 2 )/2 ≥ ≥≥ Quadratic mean Arithmetic mean Geometric meanharmonic mean 24

25. From high school (Contd.) [(a+b)/2] 3 ≤ (a 3 +b 3 )/2 ( for a, b ≥0) – E= w(L 1 3 +L 2 3 +…+L M 3 ) ??? (L 1 +L 2 +…+L M ) 3 25

26. From college: Convex (Concave) By definition of “convex” 26

27. Jensen’s Inequality (finite form) ϕ (x) is convex – ϕ (t∙x 1 +(1-t)∙x 2 )≤ t∙ ϕ (x 1 )+(1-t) ∙ϕ (x 2 ) http://en.wikipedia.org/wiki/Jensen%27s_inequality#Proof_1_.28finite_form.29 27

28. a i =1/n ϕ (x) =x 2 (Convex) ϕ (x) =x 3 (Convex for x≥0) – E= w(L 1 3 +L 2 3 +…+L M 3 )=w∙M ∙ (L 1 3 +L 2 3 +…+L M 3 )/M – ≥ w∙M ∙[(L 1 +L 2 +…+L M )/M] 3 =w∙L 3 /M 2 ≥ 28

29. Homework Use Jensen’s Inequality to prove – for a i >0 ≥ 29

30. Check the assumptions Power consumption is zero when the processor is not active 30

31. Idle power (Static power) When IC is idle but not powered off, e.g. SRAM 31

32. Multiple power/clock domains TI OMAP 2 architecture, ISSCC 2005 Multimedia phone: NTT DoCoMo 3G FOMA 902i to be released with OMAP2420 32

33. time f T f 1 =L/T f 2 =L/(T/2)=2f 1 33

34. time P T P 1 =x∙f 3 34

35. time P T P 1 =x∙f 3 +P static 35

36. time P T P 1 =x∙f 3 +P static P 2 =2 3 x∙f 3 +P static 36

37. Given workload L and deadline T single processor One processor can run at any frequency (voltage) – f= b∙V The processor can be complete off when work is done (zero power when idle) Given P static – Given energy overhead of shutting down the processor (E overhead ) To minimize energy consumption, at which frequency should the processor run? 37

38. Check the assumptions (Contd.) The workload can be divided without overhead: L = L 1 +L 2 +…+L M (L ≥ Li≥0) Communication cost between processors!!! 38

39. Wireless interfaces Stay connected Establish connection Transfer data Transmit vs. receive 39

40. Energy per bit transfer Oppermann et al., IEEE Comm. Mag. 2004 40

41. 41 Power consumption (SMT5600)

42. 42 Power consumption (T-Mobile) Bluetooth Wi-Fi Cellular

43. 43 Power consumption (Contd.) Theoretical limits – Receiving energy per bit > N * 10 -0.159 N: Noise spectral power level Wideband communication Distance: d Propagation constant: a (1.81-5.22) P RX P TX ∝ P RX *d a

44. 44 Power consumption (Contd.) What increases power consumption – National regulation (FCC) Available spectrum band (Higher band, higher power) Limited bandwidth Limited transmission power – Noise and reliability – Higher capacity Multiple access (CDMA, TDMA etc.) – Security – Addressability (TCP/IP) – More……

45. Wasteful wireless communication 45 Time Micro power management Space Directional communication Spectrum Efficiency-driven cognitive radio

46. Space waste Omni transmission  huge power by power amplifier (PA) 46

47. Time waste Network Bandwidth Under-Utilization – Modest data rate required by applications IE ~ 1Mbps, MSN video call ~ 3Mbps – Bandwidth limit of wired link 6Mbps DSL at home 47

48. Spectrum waste 48

49. 49 Wireless system architecture Application Transport Network Data link Host computer RF front ends Baseband Network interface Network protocol stackHardware implementation Physical

50. 50 Power consumption (Contd.) Baseband processor Antenna interface LNA Low-noise amplifier PA Power amplifier Intermediate Frequency (IF) signal processing Local Oscillator (LO) Physical Layer IF/Baseband Conversion MAC Layer & above >60% non-display power consumed in RF RF technologies improve much slower than IC

51. 51 Power consumption (Contd.) Source: Li et al, 2004 ComponentsPower (mW) Power amplifier (PA) 246 Frequency synthesizer (VCO/FS) 67.5 Mixer30.3 LNA20 Baseband processing 5

52. 52 Circuit power optimization Major power consumers Baseband processor Antenna interface LNA Low-noise amplifier High duty cycle PA Power amplifier High power consumption Intermediate Frequency (IF) signal processing Local Oscillator (LO) Almost always on Physical Layer IF/Baseband Conversion MAC Layer & above Huge dynamic range 10 5

53. 53 Circuit power optimization (Contd.) Reduce supply voltage – Negatively impact amplifier linearity Higher integration – CMOS RF – SoC and SiP integration Power-saving modes

54. 54 Circuit power optimization (Contd.) Power-saving modes – Complete power off (Circuit wake-up latency + network association latency) on the order of seconds – Different power-saving modes Less power saving but short wake-up latency

55. 55 Power-saving modes Baseband processor Antenna interface LNA Low-noise amplifier PA Power amplifier Intermediate Frequency (IF) signal processing Local Oscillator (LO) Physical Layer IF/Baseband Conversion MAC Layer & above Radio Deep Sleep Wake-up latency on the order of micro seconds

56. 56 Power-saving modes (Contd.) Baseband processor Antenna interface LNA Low-noise amplifier PA Power amplifier Intermediate Frequency (IF) signal processing Local Oscillator (LO) Physical Layer IF/Baseband Conversion MAC Layer & above Sleep Mode Wake-up latency on the order of milliseconds Low-rate clock with saved network association information

57. 57 Network power optimization Use power-saving modes – Example: 802.11 wireless LAN (WiFi) Infrastructure mode: Access points and mobile nodes

58. 58 802.11 infrastructure mode Mobile node sniffs based on a “Listen Interval” – Listen Interval is multiple of the “beacon period” Beacon period: typically 100ms During a Listen Interval – Access point buffers data for mobile node sends out a traffic indication map (TIM), announcing buffered data, every beacon period – Mobile node stays in power-saving mode After a Listen Interval – Mobile node checks TIM to see whether it gets buffered data – If so, send “PS-Poll” asking for data

59. 59 Buffering/sniffing in 802.11 Gast, 802.11 Wireless Network: The Definitive Guide 802.15.1/Bluetooth uses similar power-saving protocols: Hold and Sniff modes

60. 60 Power profile of 802.11b

61. 61 Network variations Cellular Wi-Fi Cellular Wi-Fi Cellular Wi-Fi