School of Mechanical Engineering
FACULTY OF ENGINEERING
From ZettaBytes to zeptoJoules – will
digital demand outstrip the physical
limits?
Dr Jon Summers ([email protected])
Institute of ThermoFluids (iTF)
Data Centre World, London, 15th to 16th March 2017
Agenda
 Information and Energy.
 Digital growth.
 Looking forward to 2030.
 Data centre power
consumption.
 Power versus demand based
on different technologies.
Waldrop, M. Mitchell. "The chips are down for Moore’s law." Nature News 530.7589 (2016): 144.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
What is a ZettaByte and a zeptoJoule?
 ZettaByte = ZB = 1,000,000,000,000,000,000,000 Bytes
 zeptoJoule = zJ = 0.000000000000000000001 Joules
 Bytes are a measure digital information.
 Joules are a measure of energy.
Exa is 18 zeros, Peta is 15 zeros, Tera is 12 zeros.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
Information and Energy
 Information is physical: writing on stone, printing text in a book – difficult
to reverse so thermodynamic entropy
increases.
 Rolf Landauer of IBM in 1961 demonstrated that
the minimum dissipation of energy in the erasure of
1 bit at room temperature is 3zJ.
 Bennett’s digital tape machine as discussed in
Feynman’s Lectures on Computation shows
that at room temperature a tape carrying a full
fuel load, 3zJ per bit, carries zero information.
Bennett, C.H., 1982. The thermodynamics of computation—a
review. International Journal of Theoretical Physics, 21(12), pp.905-940.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
Rolf Landauer (1961), "Irreversibility and heat generation in the
computing process" , IBM Journal of Research and Development 5 (3): 183–191
ZB to zJ
101101110101000010010001101110111101001
We are the cause of data centre growth.
Global Mobile Data Traffic Forecast
by Region
ExaByte =EB = 1,000,000,000,000,000,000 Bytes
Source: http://wearesocial.sg/
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
The Zettabyte era!
Cisco white paper on the Zettabyte (ZB) era:
– 2016 = 1.1ZB of traffic per annum
– Energy requirement of the network is growing
faster than data centres
– Metro traffic is growing faster than long haul
– Content delivery networks/systems
– Could grow micro-data centres
– What to do if everyone wants to stream 4k
video per year? And on the mobile network!
http://www.cisco.com/c/en/us/solutions/collateral/service-provider/visualnetworking-index-vni/VNI_Hyperconnectivity_WP.pdf
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
Store, transmit and compute digital
information
 Battle between digital growth and
energy efficiency of compute,
storage and transmission of
digital information.
 Xu based on Hilbert and Lopez:
Xu ZW. Cloud-sea computing systems: Towards thousand-fold
improvement in performance per watt for the coming Zettabyte
era. JOURNAL OF COMPUTER SCIENCE AND
TECHNOLOGY 29(2): 177–181 Mar. 2014.
Hilbert, M. and López, P., 2011. The world’s technological
capacity to store, communicate, and compute information.
science, 332(6025), pp.60-65.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
IPS = Instruction per second
101101110101000010010001101110111101001
Looking forward to 2030
 According to the paper by Xu, compute will be operating at 2588 ZIPS
 Can convert this to MW if we knew what power is required to perform an
instruction per second in 2030.
 Koomey’s law gives a prediction of computations per kWh, which can be
used to estimate how many computations can be done for a kWh in
2030 = 19.84 ExaComps => 5.5TIPS per W = 470GW of power =
1692TWh per year!
 Based on van Heddeghem et al, Storage and Communication in the DC
are consistently 20% and with a PUE of 1.1, 1692TWh = 2233TWh.
Van Heddeghem, Ward, et al. "Trends in worldwide ICT electricity consumption from 2007 to 2012." Computer Communications 50 (2014): 64-76.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
Predicted Data Center Electricity Usage
2233TWh based on
ZIPS by Xu.
Andrae, A.S. and Edler, T., 2015. On global
electricity usage of communication technology:
trends to 2030. Challenges, 6(1), pp.117-157.
Note that 2015 world electricity
production was 23,950TWh!
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
Data centre power consumption
Consider the contributing factors to Data Centre Power Consumption:
Power = Ndatacom x Power/datacom x PUE
Competition between:
Demand UP and
Consolidate DOWN
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
Efficiency of
ICT equipment
is a function of
Moore’s Law
ZB to zJ
10 years of PUE have
helped to reduce
overhead of a data
centre end use energy
consumption
101101110101000010010001101110111101001
Power consumption of IT hardware
Power/datacom = Ntr x Freq x Etr x CompUE
Number of transistors
per datacom has
increased for 50 years
doubled ever 2 years by
Moore’s law and indicates
performance.
Clock speeds have
not really
increased since
2005 as it has
a significant effect,
but is now variable.
Energy
consumption
per transistor
is key to total
power
consumption.
Note also that Power/datacom =  x C x V2 x Freq + leakage
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
Compute
Usage
Effectiveness
Overhead
from power
supply, xDD,
RAM, etc.
101101110101000010010001101110111101001
Energy consumption of a transistor
Etr
=
EFACTOR
Energy/Entropy Factor related
to the approach of state
changes in Field Effect
Transistors (FETs):
Depends on Voltage and
materials.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
x ( kB
Physical constant used
statistical mechanics,
called the Boltzmann
constant with a value of
1.38 x 10-23 J/K
ZB to zJ
x T)
Temperature
at which the
transistor is
operating.
101101110101000010010001101110111101001
History of EFACTOR
Processor
Architecture
Year
Feature Size
EFACTOR
Pentium 486
1989
600nm
9,932,000
Pentium M
2003
130nm
78,500
Core
2006
65nm
67,700
Nehalem
2008
45nm
18,900
Sandy Bridge
2012
32nm
4,500
Ivy Bridge
2014
22nm
1,750
Broadwell
2015
14nm
1,500
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
New
TriGate
FinFETS
~ 3D!
101101110101000010010001101110111101001
EFACTOR is linked to Moore’s Law
Cost of transistors is going
up. Peaked at 20 million
per $ in 2015
Moore’s Law: Self-fulfilling
prophecy to provide double
the number of transistors in
the same area every two
years.
Cross, T. "After Moore’s Law: Double, double, toil
and trouble." The Economist, Technology Quarterly,
Quarter 1 (2016).
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
Getting EFACTOR down.
 millivolt, transistor size and materials may reduce the EFACTOR , or going 3D
 Waldrop quotes
“My bet is that we
run out of money
before we run out
of physics”
[Rock’s Law]
Waldrop, M. Mitchell. "The chips are down
for Moore’s law." Nature News 530.7589
(2016): 144.
Carballo, Juan-Antonio, Wei-Ting Jonas Chan,
Paolo A. Gargini, Andrew B. Kahng, and Siddhartha
Nath. "ITRS 2.0: Toward a re-framing of the
Semiconductor Technology Roadmap." In
Computer Design (ICCD), 2014 32nd IEEE
International Conference on, pp. 139-146. IEEE, 2014.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
What are the practical limits of EFACTOR?
Etr
=
EFACTOR
x
kB
x T
 Frank argues that to measure a signal in the correct state with an error of
pe (<10-40) requires the signal energy to be greater than ln(1/pe)kBT, that is
around 100kBT.
 Bennett gave an interesting example of DNA polymerization that occurs in
cell division to use ~40kBT of energy per step.
 If we cannot get EFACTOR down, then we reduce temperature, T!
Frank, Michael P. "Approaching the physical limits of computing." Multiple-Valued Logic, 2005. Proceedings. 35th International
Symposium on. IEEE, 2005.
Bennett, Charles H. "The thermodynamics of computation—a review. "International Journal of Theoretical Physics 21.12 (1982): 905-940.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
Superconducting Computing!
• IBM ran a project from 19731983 on this – terminated due
to the success of Si.
• At 4K and an EFACTOR of
1,500, a cryotron (Buck’s
superconducting switch)
would use 83 zJ and
switching frequency of less
than 125 THz limited by
Planck Constant.
Image from:
Brock, David C. "The
NSA's frozen dream."
IEEE Spectrum 53,
no. 3 (2016): 54-60.
Buck, Dudley A. "The cryotron-a superconductive
computer component." Proceedings of the IRE 44,
no. 4 (1956): 482-493.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
What technology will be in the data centre
in 2030?
 CMOS with millivolt, reduce feature size, going 3D and using new materials
(error rate, leakage/quantum effects, heat issue, still in the lab)
[Etr = 100kBT]
 Superconducting (switch count per unit volume too low) [Etr > h/(tdelay)]
 Quantum (still the challenge of error correction) [Etr > h/(tdelay)]
 Reversible (complex logic) [ Etr = 0.04kBT]
 Dark silicon/multicore (software development needed) [ Ntr < Ntr ]
 Approximate computing (specialised application) [low bit operations]
 Neuromorphic (energy efficiency issues, application specific and
massively parallel)
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
Three technological scenarios.
1. At 5nm with 3D features, Nt could be 80 billion with the
clock at 5GHz requiring around 180W for a CPU that could
have 180 cores = 320GIPS per W
2. Reversible logic and Nt and same as above could operate
requiring 0.08W giving 180TIPS per W
3. Superconducting under same constraints as above would
require > 0.0014W giving 2.57PIPS per W
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
How does the power consumption
compare?
1. At 320GIPS per W yields
38,400 TWh – not possible!
2. At 180TIPS per W yields
68 TWh – possible!
2233TWh based on
Koomey.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
3. At 2.57PIPS per W yields
4.79 TWh – possible!
ZB to zJ
101101110101000010010001101110111101001
Reduce Ntr that are in use.
Power/datacom = Ntr x Freq x Etr x CompUE
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
So what does this all mean?
 The current roadmap to Etr = 100kBT will need excessive
amounts of power to meet the demands of 2030.
 Dark Silicon could become dominant and likely to cause
growth in ASICs [e.g. FPGAs, GPU like of special purpose
hardware] to keep the power consumption down.
 Increasing DC power consumption => candidates for decarbonising heat => stronger requirement for efficient
harvesting of heat using liquid cooling.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
What about energy consumption of the
networks?
Van Heddeghem, Ward, et al. "Trends in worldwide ICT electricity
consumption from 2007 to 2012." Computer Communications 50 (2014): 64-76.
Andrae, A.S. and Edler, T., 2015. On global electricity usage of communication
technology: trends to 2030. Challenges, 6(1), pp.117-157.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
DCs likely to become more distributed.
 Power consumption
of the networks
(in particular mobile,
end user + IOT) =>
Micro-Edge data
centres.
http://www.gsma.com/network2020/
wp-content/uploads/2015/01/
Understanding-5G-Perspectives-onfuture-technological-advancements-in-mobile.pdf
 Edge content in DC
at base of 5G masts!
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
Summary
 Current technology and demand trends will exhaust
electrical grids – “demand is likely to outstrip physical limits”.
Paradigm shift not yet evident – alternatives still in the lab.
 DCs are likely to become distributed – large at the core,
micro at the edge. Hybrid IT hardware using ASICs and
cooling using liquids.
 DCs are likely to be an important component of decarbonising heat.
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001
Many thanks.
Dr Jon Summers
[email protected]
e-  e-  e-  e-  e-  e-  e-  e-  e-  e-
ZB to zJ
101101110101000010010001101110111101001