ESE534 Computer Organization Day 6: February 12, 2014 Energy, Power, Reliability Penn ESE534 Spring2014 -- Mehta & DeHon.
Download ReportTranscript ESE534 Computer Organization Day 6: February 12, 2014 Energy, Power, Reliability Penn ESE534 Spring2014 -- Mehta & DeHon.
ESE534 Computer Organization
Day 6: February 12, 2014 Energy, Power, Reliability Penn ESE534 Spring2014 -- Mehta & DeHon 1
Today
• Energy tradeoffs • Voltage limits and leakage • Variations • [more of the gory transistor equations…] 2 Penn ESE534 Spring2014 -- Mehta & DeHon
At Issue
• Many now argue
energy
ultimate scaling limit – (not lithography, costs, …) will be the • Proliferation of portable and handheld devices – …battery size and life biggest issues • Cooling, energy costs may dominate cost of electronics – Even server room applications 3 Penn ESE534 Spring2014 -- Mehta & DeHon
Preclass 1
• 1GHz case – Voltage?
– Energy per Operation?
– Power required for 2 processors?
• 2GHz case – Voltage?
– Energy per Operation?
– Power required for 1 processor?
Penn ESE534 Spring2014 -- Mehta & DeHon 4
Preclass 1 Lesson
• What does Preclass 1 result tell us?
Penn ESE534 Spring2014 -- Mehta & DeHon 5
Energy and Delay
E
1 2 2
CV
t gd
=Q/I=(CV)/I I
d,sat
=(
m
C
OX
/2)(W/L)(V
gs
-V
TH
)
2 6 Penn ESE534 Spring2014 -- Mehta & DeHon
Energy/Delay Tradeoff
• • E V 2 t gd 1/V
E
1 2
CV
t gd =(CV)/I I d,sat (V gs -V TH ) 2 • We can trade speed for energy • E × ( t gd ) 2 constant Martin
et al.
Power-Aware Computing
, Kluwer 2001 http://caltechcstr.library.caltech.edu/308/ 7 Penn ESE534 Spring2014 -- Mehta & DeHon
Area/Time Tradeoff
• Also have Area-Time tradeoffs – HW2 spatial vs temporal multipliers – See more next week • Compensate slowdown with additional parallelism • …trade Area for Energy Architectural Option 8 Penn ESE534 Spring2014 -- Mehta & DeHon
Question
• By how much can we reduce energy?
• What limits us?
Penn ESE534 Spring2014 -- Mehta & DeHon 9
Challenge: Power
Penn ESE534 Spring2014 -- Mehta & DeHon 10
Origin of Power Challenge
• Limited capacity to remove heat – ~100W/cm 2 – 1-10W/cm 2 force air ambient • Transistors per chip grow at Moore’s Law rate = (1/F) 2 • Energy/transistor must decrease at this rate to keep constant power density • P/tr CV 2 f • E/tr CV 2 – …but V scaling more slowly than F 11 Penn ESE534 Spring2014 -- Mehta & DeHon
ITRS V dd Scaling: More slowly than F Penn ESE534 Spring2014 -- Mehta & DeHon 12
ITRS CV 2 Scaling: More slowly than (1/F) 2 Penn ESE534 Spring2014 -- Mehta & DeHon 13
Origin of Power Challenge
• Transistors per chip grow at Moore’s Law rate = (1/F) 2 • Energy/transistor must decrease at this rate to keep constant • E/tr CV 2 Penn ESE534 Spring2014 -- Mehta & DeHon 14
Historical Power Scaling
[Horowitz et al. / IEDM 2005] Penn ESE534 Spring2014 -- Mehta & DeHon 15
Microprocessor Power Density
Watts The Future of Computing Performance: Game Over or Next Level?
National Academy Press, 2011 16
Intel Power Density
Penn ESE534 Spring2014 -- Mehta & DeHon 17
Impact
Power Limits Integration Density Limit Constant Power Limit 45nm 32nm 22nm 16nm 11nm
Source: Carter/Intel Penn ESE534 Spring2014 -- Mehta & DeHon 18 18
Impact
• Power density is limiting scaling – Can already place more transistors on a chip than we can afford to turn on!
• Power is potential challenge/limiter for all future chips.
– Only turn on small percentage of transistors?
– Operate those transistors as much slower frequency?
– Find a way to drop V dd ?
19 Penn ESE534 Spring2014 -- Mehta & DeHon
How far can we reduce V
dd
?
20 Penn ESE534 Spring2014 -- Mehta & DeHon
Limits
• Ability to turn off the transistor • Parameter Variations • Noise (not covered today) Penn ESE534 Spring2014 -- Mehta & DeHon 21
MOSFET Conduction
From: http://en.wikipedia.org/wiki/File:IvsV_mosfet.png
Penn ESE534 Spring2014 -- Mehta & DeHon 22
Transistor Conduction
• Three regions – Subthreshold (V gs
Saturation Region
• (V gs >V TH ) • (V ds > (V gs -V TH ))
I
ds,sat
=(
m
C
OX
/2)(W/L)(V
gs
-V
TH
)
2 24 Penn ESE534 Spring2014 -- Mehta & DeHon
Linear Region
• (V gs >V TH ) • (V ds < (V gs -V TH )) I ds,lin =( m C OX )(W/L)((V gs -V TH )V ds -(V ds ) 2 /2) 25 Penn ESE534 Spring2014 -- Mehta & DeHon
Subthreshold Region
• (V gs I I Penn ESE534 Spring2014 -- Mehta & DeHon [Frank, IBM J. R&D v46n2/3p235] 26 • Concerned about I on and I off • I on drive (saturation) current for charging – Determines speed: T gd = CV/I • I off leakage current – Determines leakage power/energy: • P • E leak leak = V × I leak = V × I leak × T cycle 27 Penn ESE534 Spring2014 -- Mehta & DeHon • To avoid leakage want I off • Switch V from V dd to 0 very small • V gs in off state is 0 (V gs I I 10 28 Penn ESE534 Spring2014 -- Mehta & DeHon Leakage I 10 • S 90mV for single gate • S 70mV for double gate • For lowest leakage, want S small, V TH large • 4 orders of magnitude I VT /I off V TH >280mV Leakage limits V TH Penn ESE534 Spring2014 -- Mehta & DeHon in turn limits V dd 29 on off • Maximize I on /I off – for given V dd • Get to pick V TH , V dd ? E sw CV 2 d,sat m OX gs TH 2 I d,lin =( m C OX )(W/L)(V gs -V TH )V ds -(V ds ) 2 /2 I I 30 Penn ESE534 Spring2014 -- Mehta & DeHon • E = E sw + E leak • E leak = V × I leak × T cycle • E sw CV 2 I I • I chip-leak = N devices × I tr-leak 31 • E leak (V) ? • T cycle (V)? Penn ESE534 Spring2014 -- Mehta & DeHon 32 • Assign calculations – SIMD – each student computes for a different Voltage • Collect results on board • Group: Identify minimum energy point and discuss 33 Penn ESE534 Spring2014 -- Mehta & DeHon Penn ESE534 Spring2014 -- Mehta & DeHon 35 • Subthreshold slope prevents us from scaling voltage down arbitrarily. • Induces a minimum operating energy. 36 Penn ESE534 Spring2014 -- Mehta & DeHon Penn ESE534 Spring2014 -- Mehta & DeHon 37 Statistical Dopant Count and Placement Penn ESE534 Spring2014 -- Mehta & DeHon 38 [Bernstein et al, IBM JRD 2006] th Penn ESE534 Spring2014 -- Mehta & DeHon 39 [Bernstein et al, IBM JRD 2006] • Fewer dopants, atoms increasing Variation • How do we deal with variation? % variation in V TH (From ITRS prediction) 40 40 Penn ESE534 Spring2014 -- Mehta & DeHon • Higher V TH ? – Not drive as strongly – I d,sat (V gs -V TH ) 2 slower • Lower V TH ? – Not turn off as well I leaks more 10 41 Penn ESE534 Spring2014 -- Mehta & DeHon • Margin for expected variation • Must assume V TH can be any value in range I on,min =I on ( Vth,max ) I d,sat (V gs -V TH ) 2 V gs = V dd V TH Penn ESE534 Spring2014 -- Mehta & DeHon 42 • Must raise V dd to increase drive strength • Increase energy I on,min =I on ( Vth,max ) I d,sat (V gs -V TH ) 2 V gs = V dd V TH Penn ESE534 Spring2014 -- Mehta & DeHon 43 • Increasing variation forces higher voltages – On top of our leakage limits Penn ESE534 Spring2014 -- Mehta & DeHon 44 44 Old Variations New • Margins growing due to increasing variation Delay • Margined value may be worse than older technology? 45 Penn ESE534 Spring2014 -- Mehta & DeHon End of Energy Scaling? Black nominal Grey with variation [Bol et al ., IEEE TR VLSI Sys 17(10):1508 —1519] Penn ESE534 Spring2014 -- Mehta & DeHon 46 • Larger chips (billions of transistors) sample further out on distribution curve From: http://en.wikipedia.org/wiki/File:Standard_deviation_diagram.svg 47 Penn ESE534 Spring2014 -- Mehta & DeHon • Can trade time for energy – … area for energy • Variation and leakage limit voltage scaling • Power major limiter going forward – Can put more transistors on a chip than can switch • Continued scaling demands – Deal with noisier components • High variation • … other noise sources 48 Penn ESE534 Spring2014 -- Mehta & DeHon • Homework 2 due Tonight • Reading for Monday on web • Homework 3 due next Wednesday Penn ESE534 Spring2014 -- Mehta & DeHon 49Operating a Transistor
Leakage
How maximize I
/I
?
I
=(
C
/2)(W/L)(V
-V
)
Preclass 2
Preclass 2
In Class
Graph for In Class
Impact
Challenge: Variation
V
Variability @ 65nm
Variation
Impact of Variation?
Variation
Margining
Variation
Chips Growing
Big Ideas
Admin