Transcript Chapter # 5 - EECS Instructional Support Group Home Page

EECS 150 - Components and Design
Techniques for Digital Systems
Lec 24 –Power, Power, Power
11/27/2007
David Culler
Electrical Engineering and Computer Sciences
University of California, Berkeley
http://www.eecs.berkeley.edu/~culler
http://inst.eecs.berkeley.edu/~cs150
1
Broad Technology Trends
Moore’s Law: # transistors on
Bell’s Law: a new computer
cost-effective chip doubles every
18 months
class emerges every 10 years
Computers
Per Person
1:106
1:103
Mainframe
Mini
Workstation
PC
Laptop
1:1
Today: 1 million transistors per $
PDA
Cell
103:1
years
Mote!
Same fabrication technology provides CMOS radios
for communication and micro-sensors
2
Sustaining Moore’s Law
“If unchecked, the increasing power
requirements of computer chips could
boost heat generation to absurdly high
levels,” said Patrick Gelsinger, Intel’s
CTO is reported to have said.
“By mid-decade, that Pentium PC may
need the power of a nuclear reactor. By
the end of the decade, you might as
well be feeling a rocket nozzle than
touching a chip. And soon after 2010,
PC chips could feel like the bubbly hot
surface of the sun itself,”
3
Power, Power, Power
• IT devices represent 2% of global
CO2 emissions worldwide
LAN and office
telecom, 7%
Computers
Per Person
1:106
1:103
Printers, 6%
PCs and
Monitors, 39%
Mobile
telecom, 9%
Mainframe
Fixed-line
Telecom, 15%
Servers, 23%
Mini
Workstation
PC
Laptop
1:1
Source Gartner
PDA
Cell
103:1
years
Mote!
4
What is EECS150 about?
Pgm Language
CS 61C
Deep Digital Design Experience
Asm / Machine Lang
Fundamentals of Boolean Logic
Instruction Set Arch
Synchronous Circuits
Machine Organization
Finite State Machines
HDL
Timing & Clocking
Device Technology & Implications
FlipFlops
Gates
Circuits
EE 40
Devices
Transistor Physics
Transfer Function
Controller Design
Arithmetic Units
Bus Design
Encoding, Framing
Testing, Debugging
Hardware Architecture
HDL, Design Flow (CAD)
5
Data Centers
Client
Computers
Per Person
1:106
1:103
• 1.5% of total US energy
consumption in 2006
• 60 Billion kWh
• Doubled in past 5 years
and expected to double in
next 5 to 100 Billion kWh
– 7.4 B$ annually
EPA report aug 4 2007 delivered to
congress in response to public law
109-431
Mainframe
Mini
Workstation
PC
Laptop
1:1
PDA
Cell
103:1
years
Mote!
48% of IT budget spent on energy
50% of data center power goes into
cooling
1 MW DC => 177 M kwH + 60 M
gals water + 145 K lbs copper +
21 k lbs lead
6
Servers: Total Cost of Ownership (TCO)
Machine rooms
are expensive …
removing heat
dictates how
many servers can
fit
Reliability: running computers hot
makes them fail more often
Electric bill adds
up! Powering the
servers +
powering the air
conditioners is a
big part of TCO
7
M. K. Patterson, A. Pratt, P. Kumar,
“From UPS to Silicon: an end-to-end evaluation of datacenter efficiency”, Intel Corporation
8
P watts = I amps * V volts
1A
1V
+
-
This is how electric tea pots work ...
Heats 1 gram of water
0.24 degree C
0.24 Calories per Second
1 Joule of Heat
Energy per Second
1 Ohm
Resistor
20 W rating: Maximum power
the package is able to
transfer to the air. Exceed
rating and resistor burns.
9
Basics
• Warning! In everyday language, the term
“power” is used incorrectly in place of “energy”
• Power is not energy
– E=P*T
• Power is not something you can run out of
• Power can not be lost or used up
• It is not a thing, it is merely a rate
• It can not be put into a battery any more than
velocity can be put in the gas tank of a car
10
Data Center Power Usage Today
11
PC
• HPxw4200
Client
– 180 w active with two LCDs
– 130 w w/o monitor, 110 w idle,
– 6 w suspend
J2EE
SOAP
• 60% are left on around the clock
• 15% of all office power
• US:
Enterprise
Server
Computers
Per Person
1:106
1:103
– 1.72 B$ & 15 M tons CO2 annually
• Mid size company:
– 165 K$ & 1400 tons of CO2
• Existing power mgmt (hibernation)
can reduce by 80%
Mainframe
Mini
Workstation
PC
Laptop
1:1
=> Do nothing well
PDA
Cell
PC Energy Report 2007, 1E
103:1
years
Mote!
12
Do Nothing Well
13
Notebooks ... now most of the PC market
Apple MacBook -- Weighs 5.2 lbs
8.9 in
1 in
12.8 in
Performance: Must be “close enough” to desktop
performance ... many people no longer own a desktop
Size and Weight: Ideal: paper notebook
Heat: No longer “laptops” -- top may get “warm”,
bottom “hot”. Quiet fans OK
14
Battery: Set by size and weight limits ...
Battery rating:
55 W-hour
46x energy than iPod nano.
iPod lets you listen to music
for 14 hours!
Almost full 1 inch depth.
Width and height set by
available space, weight.
At 2.3 GHz, Intel
Core Duo CPU
consumes 31 W
running a heavy load
- under 2 hours
battery life! And,
just for CPU!
At 1 GHz, CPU consumes
13 Watts. “Energy saver”
option uses this mode
...
15
Battery Technology
• Battery technology has developed slowly
• Li-Ion and NiMh still the dominate technologies
• Batteries still contribute significantly to the
weight of mobile devices
Nokia 61xx 33%
Handspring
PDA - 10%
Toshiba Portege
3110 laptop - 20%
16
55 W-hour battery stores
the energy of
1/2 a stick of dynamite.
If battery short-circuits,
catastrophe is possible ...
17
CPU Only Part of Power Budget
2004-era notebook running a
full workload.
“other”
GPU
LCD
Backlight
CPU
If our CPU took no power
at all to run, that would
only double battery life!
LCD
18
“X-Internet” Beyond the PC
Internet Computers
Internet Users
500
Million
Today’s Internet
1.5 Billion
Automobiles
700 Million
Telephones
4 Billion
X-Internet
Electronic Chips
60 Billion
Forrester Research, May 19
2001
Revised 2007
“X-Internet” Beyond the PC
Millions
15000
10000
5000
X
Internet
20
01
20
02
20
03
20
04
20
05
20
06
20
07
20
08
20
09
20
10
0
PC
Internet
Year
Forrester Research, May 20
2001
Cooling an iPod nano ...
Like a resistor, iPod relies
on passive transfer of heat
from case to the air
Why? Users don’t want
fans in their pocket ...
To stay “cool to the touch” via passive cooling,
power budget of 5 W
If iPod nano used 5W all the time, its battery would last
15 minutes ...
21
Powering an iPod nano (2005 edition)
Battery has 1.2 W-hour
rating: Can supply
1.2 W of power for 1 hour
1.2 W / 5 W = 15 minutes
More W-hours require bigger battery
and thus bigger “form factor” -it wouldn’t be “nano” anymore!
Real specs for iPod nano :
14 hours for music,
4 hours for slide shows
85 mW for music
300 mW for slides
22
0.55 ounces
12 hour
battery life
$79.00
1 GB
23
20 hour battery life for audio,
6.5 hours for movies (80GB version)
Up from 14 hours
24 hour
battery life for for 2005 iPod nano
audio
Thinner than 2005 iPod nano
5 hour battery
life for photos
Up from 4
hours for 2005
iPod nano
12 hour
battery life
24
What’s in the iPhone?
Motherboard
USB & GSM
Battery
WiFI antenna
GSM antenna
http://www.anandtech.com/printarticle.aspx?i=3026
25
What’s in your iPhone?
Main Processor
LCD i/f
ARM1176 + 1GB mem
WiFi & Most of Cell Phone
4 GB NAND Flash
• 3 ARM processors
26
iPhone Parts (?)
•
•
•
•
•
•
•
•
•
•
•
•
•
Baseband processor: Infineon – SGold3/ARM926?
Applications/video processor:
Samsung/ARM10 or 11
802.11 chip: Marvell/ARM9?
Touchscreen controller: Broadcom
Touchscreen: Balda/TPK
Bluetooth: CSR
USB IC: Alcor, Phison
Audio: Wolfson
Memory module: A-Data, Transcend
Flash memory: Samsung, Toshiba,
Hynix
Position sensor (MEMS?):
STMicroelectronics, Analog
devices?
Light sensor: ???
Proximity sensor: ???
•
•
•
•
•
•
•
•
•
•
•
•
Camera sensor: Micron?
Camera module: Altus or Lite-On
Technology, Primax Electronics
Camera lens: Largan Precision
Microphone: ???
Power management: NXP?
Passives: Cyntec
Quartz: TXC
Assembly: Foxconn, FIH
Casing & mechanical parts:
Foxconn & Catcher
Push button: Sunrex
Connectors & cable: Entery,
Cheng Uei, Foxlink, Advanced
Connectek
PCB: Unimicron & Tripod
27
UCB Mote Platforms
*
Computers
Per Person
1:106
1:103
*
Mainframe
Mini
Workstation
PC
Laptop
1:1
PDA
Cell
* Crossbow variation
103:1
years
Mote!
28
Key Design Elements
Flash Storage
data logs
pgm images
proc
Data
SRAM
timers
pgm
EPROM
WD
Sensor
Interface
analog sensors
digital sensors
Wireless Net
Interface
RF
transceiver
Wired Net
Interface
serial link
USB,EN,…
Low-power
Standby & Wakeup
•
•
•
•
•
•
ADC
antenna
Efficient wireless protocol primitives
Flexible sensor interface
Ultra-low power standby
Very Fast wakeup
Watchdog and Monitoring
Data SRAM is critical limiting resource
29
TinyOS-driven architecture
• 3K RAM = 1.5 mm2
• CPU Core = 1mm2
– multithreaded
• RF COMM stack = .5mm2
– HW assists for SW stack
•
•
•
•
Page mapping
SmartDust RADIO = .25 mm2
SmartDust ADC 1/64 mm2
I/O PADS
• Expected sleep: 1 uW
– 400+ years on AA
• 150 uW per MHz
• Radio:
– .5mm2, -90dBm receive sensitivity
– 1 mW power at 100Kbps
• ADC:
– 20 pJ/sample
– 10 Ksamps/second = .2 uW.
jhill mar 6, 2003
30
Microcontrollers
• Memory starved
– Far from Amdahl-Case 3M rule
• Fairly uniform active inst per nJ
– Faster MCUs generally a bit better
– Improving with feature size
• Min operating voltage
– 1.8 volts => most of battery energy
– 2.7 volts => lose half of battery energy
• Standby power
–
–
–
–
substantial improvement in 2003
Probably due to design focus
Fundamentally SRAM leakage
Wake-up time is key
• Trade sleep power for wake-up
time
DMA Support: permits ADC
sampling while processor is
sleeping
– Memory restore
31
What we mean by “Low Power”
• 2 AA => 1.5 amp hours (~4 watt hours)
• Cell => 1 amp hour
(3.5 watt hours)
Cell: 500 -1000 mW
WiFi: 300 - 500 mW
GPS: 50 – 100 mW
=> few hours active
=> several hours
=> couple days
WSN: 50 mW active, 20 uW passive
450 uW => one year
45 uW => ~10 years
* System design
* Leakage (~RAM)
* Nobody fools
mother nature
Ave Power = fact * Pact + fsleep * Psleep + fwaking * Pwaking
32
Mote Power States at Node Level
Active
Active
Sleep
WakeUP
Work
Sleep
WakeUP
Work
Telos: Enabling Ultra-Low Power Wireless Research, Polastre, Szewczyk, Culler, IPSN/SPOTS 2005
33
Radios
• Trade-offs:
– resilience / performance => slow wake up
– Wakeup vs interface level
– Ability to optimize vs dedicated support
34
Power to Communicate
140
120
100
80
60
40
20
0
0
1
2
3
4
5
35
Multihop Routing
• Upon each transmission, one of the recipients
retransmits
– determined by source, by receiver, by …
– on the ‘edge of the cell’
36
Energy Profile of a Transmission
Datasheet
Analysis
• Power up oscillator &
radio (CC2420)
• Configure radio
• Clear Channel
Assessment, Encrypt
and Load TX buffer
• Transmit packet
• Switch to rcv mode,
listen, receive ACK
20mA
10mA
5 ms
10 ms
37
Example: TX maximum packet
25
20
mA
15
10
5
0
-15
-10
-5
0
5
10
15
ms
38
The “Idle Listening” Problem
• The power consumption of “short range” (i.e., lowpower) wireless communications devices is roughly
the same whether the radio is transmitting, receiving,
or simply ON, “listening” for potential reception
– includes IEEE 802.15.4, Zwave, Bluetooth, and the many variants
– WiFi too!
– Circuit power dominated by core, rather than large amplifiers
• Radio must be ON (listening) in order receive anything.
– Transmission is infrequent. Reception α Transmit x Density
– Listening (potentially) happens all the time
Total energy consumption dominated by idle listening
39
Communication Power Consumption
Sleep
~10 uA
Transmit
~20 mA x 1-5 ms
[20 - 100 uAs]
I
Time
I
Time
Listen Receive
~20 mA ~20 mA x 2-6 ms
40
Announcements
• Project Check-offs this week
– TAs posting extra “office hours” for use of slip days
• Dr. Robert Iannucci, Nokia on Thurs
– Bring questions, show off projects
• Short HW 10 out tonight
– Due next wed.
• Wrap-up and Course Survey 12/4
• Project Demos Friday 12/7
– Signup sheet is posted
– 5 min demo + 5 min Q&A
– Set up 20 mins in advance
• Final Exam Group: 15: TUESDAY, DECEMBER 18,
2007 5-8P
41
Basics – Power and Digital Design
• Power supply provides energy for charging and discharging wires
and transistor gates. The energy supplied is stored & then
dissipated as heat.
Power: Rate of work being done wrt time
Rate of energy being used
P  dw / dt
Units:
P  E t
Watts = Joules/seconds
• If a differential amount of charge dq is given a differential increase
in energy dw, the potential of the charge is increased by:
• By definition of current:
V  dw / dq
I  dq / dt
dw dq
dw / dt 

 P V I
dq dt
t
w
 Pdt

total energy
A very practical
formulation!
If we would like
to know total energy
42
Recall: Transistor-level Logic Circuits
• Inverter (NOT gate):
Vdd
Gnd
what is the
relationship
between in and out?
in
0 volts
Vdd
out
Gnd
3 volts
43
Older Logic Families have Pullup R
nMOS Inverter
R
44
Power in CMOS
Switching Energy:
Vdd
energy used to
switch a node
Vdd
pullup
network
Calculate energy
dissipated in pullup:
i(t)
v(t)
0 1
pulldown C
network
v(t)
t0
t1
GND
t1
t1
t1
t0
t0
t0
Esw   P(t )dt   (Vdd  v)  i(t )dt   (Vdd  v)  c (dv dt ) dt 
t1
t1
t0
t0
 cVdd  dv  c  v  dv  cVdd  1 2cVdd  1 2 cVdd
Energy supplied
2
Energy stored
2
2
Energy dissipated
An equal amount of energy is dissipated on pulldown
45
Switching Power
•
Gate power consumption:
–
 f
Assume a gate output is switching its output at a rate of:
activity factor
1/f
clock rate
(probability of switching on
any particular clock period)
Pavg  E t  switching rate  Esw
Therefore:
Pavg    f 1 2 cVdd
2
Pavg
• Chip/circuit power consumption:
Pavg  n  avg  f 1 2 cavgVdd
number of nodes (or gates)
clock f
2
46
Other Sources of Energy Consumption
• “Short Circuit” Current:
•
Junction Diode Leakage :
Vout
I
Vin
Vin
Vout
Transistor drain regions
“leak” charge to substrate.
I
I
Vin
10-20% of total chip power
Diode
Characteristic
V
~1nWatt/gate
few mWatts/chip
47
Other Sources of Energy Consumption
•
Consumption caused by “DC leakage current” (Ids leakage):
Ids
Vin=0
Vout=Vdd
Ioff
Transistor s/d conductance
never turns off all the way
Vgs
Vth
Low voltage processes much worse
•
•
This source of power consumption is becoming increasing
significant as process technology scales down
For 90nm chips around 10-20% of total power consumption
Estimates put it at up to 50% for 65nm
48
Controlling Energy Consumption:
What Control Do You Have as a
Designer?
• Largest contributing component to CMOS power
consumption is switching power:
Pavg  n  avg  f 1 2 cavgVdd
2
• Factors influencing power consumption:
– n: total number of nodes in circuit
 : activity factor (probability of each node switching)
– f: clock frequency (does this effect energy
consumption?)
– Vdd: power supply voltage
• What control do you have over each factor?
• How does each effect the total Energy?
49
Example
A
add/sub
Operand Registers
B
and/or
cmp
MUX
R
Result Register
• What is the cost of optimistic compute and
select?
• How might we reduce it?
50
Discussion: Digital Design and Power
Pavg  n  avg  f 1 2 cavgVdd
2
• Think about…
–

–
–
–
n

f
c
Vdd
• In
–
–
–
–
Function units
Registers, FSMs, Counters
Busses
Clock distribution
51
Technology Scaling and Design Learning
52
Scaling Switching Energy per Gate
Moore’s Law
at work …
Due to
reduced V and
C (length and
width of Cs
decrease, but
plate distance
gets smaller)
Recent slope
reduced
because V is
scaled less
aggressively
From: “Facing the Hot Chips Challenge Again”, Bill Holt, Intel, presented at Hot Chips 17, 2005.
53
Device Engineers Trade Speed and Power
We can reduce CV2 (Pactive)
by lowering Vdd
We can increase speed
by raising Vdd and
lowering Vt
We can reduce leakage
(Pstandby) by raising Vt
From: Silicon Device Scaling to the Sub-10-nm Regime
Meikei Ieong,1* Bruce Doris,2 Jakub Kedzierski,1 Ken Rim,1 Min Yang1
54
Customize processes for product types ...
From: “Facing the Hot Chips Challenge Again”, Bill Holt, Intel, presented at Hot Chips 17, 2005.
55
Intel: Comparing 2 CPU Generations ...
Find enough
tricks, and
you can
afford to
raise Vdd a
little so that
you can raise
the clock
speed!
Clock speed
unchanged ...
Lower Vdd, lower C,
but more leakage
Design tricks:
architecture & circuits
56