ESTIMATING TOTAL POWER CONSUMPTION BY SERVERS IN
THE U.S. AND THE WORLD
Jonathan G. Koomey, Ph.D.
Staff Scientist, Lawrence Berkeley National Laboratory and
Consulting Professor, Stanford University
Contact: JGKoomey@stanford.edu, http://www.koomey.com
Final report
February 15, 2007
2
i
EXECUTIVE SUMMARY
The amount of electricity used by servers and other Internet infrastructure has become an
important issue in recent years as demands for new Internet services (like music
downloads, video-on-demand, and Internet telephony) have become more widespread.
One of the weaknesses in the literature on data center electricity use has been the lack of
credible estimates of the aggregate power used by all servers and associated equipment in
the U.S. and the world. The data on the floor area and power densities of data centers are
anecdotal and limited by the proprietary nature of such data in most companies. Data on
the installed base of servers are also closely held by the companies who track it, and
server technology continues to change rapidly, necessitating constant updates to
measurements of power used by particular server models.
This study estimates total electricity used by servers in the U.S. and the world by
combining measured data and estimates of power used by the most popular servers with
data on the server installed base. These estimates are based on more detailed data than
are previous assessments, and they will be of use to policy makers and businesses
attempting to make sense of recent trends in this industry.
Aggregate electricity use for servers doubled over the period 2000 to 2005 both in the
U.S. and worldwide (
Figure ES-1
). Almost all of this growth was the result of growth in
the number of the least expensive servers, with only a small part of that growth being
attributable to growth in the power use per unit.
Total power used by servers represented about 0.6% of total U.S. electricity consumption
in 2005. When cooling and auxiliary infrastructure are included, that number grows to
1.2%, an amount comparable to that for color televisions. The total power demand in
2005 (including associated infrastructure) is equivalent (in capacity terms) to about five
1000 MW power plants for the U.S. and 14 such plants for the world. The total electricity
bill for operating those servers and associated infrastructure in 2005 was about $2.7 B
and $7.2 B for the U.S. and the world, respectively.
This study only assesses the direct electricity used by servers and associated
infrastructure equipment. It does not attempt to estimate the effect of structural changes
in the economy enabled by increased use of information technology, which in many cases
can be substantial.
ii
Figure ES-1: Total electricity use for servers in the U.S. and the world in 2000 and
2005, including the associated cooling and auxiliary equipment
1
ESTIMATING TOTAL POWER CONSUMPTION BY SERVERS IN
THE U.S. AND THE WORLD
Jonathan G. Koomey, Ph.D.
Staff Scientist, Lawrence Berkeley National Laboratory and
Consulting Professor, Stanford University
INTRODUCTION
Electricity used by information technology (IT) equipment has been a subject of intense
interest since the first E
NERGY
S
TAR
specification for personal computers was released in
the early 1990s (Johnson and Zoi 1992). The first detailed measurements of personal
computer electricity use were published in the late 1980s (Harris et al. 1988) followed by
estimates of total power used by office equipment (Koomey et al. 1996, Norford et al.
1990, Piette et al. 1991) and assessments of potential efficiency improvements in that
equipment (Ledbetter and Smith 1993, Lovins and Heede 1990).
As the 1990s came to a close, it was becoming clear that a new class of IT equipment was
increasing in importance. Computer servers, and the data center facilities in which they
were located, were becoming more numerous and more electricity intensive. The first
major efforts to understand server electricity use more deeply were spurred by a
controversy over the total power used by IT equipment in which dubious claims were
made about the contribution of IT equipment to total electricity use in the U.S. (Huber
and Mills 1999, Mills 1999). These claims were subsequently refuted (Baer et al. 2002,
Kawamoto et al. 2002, Koomey et al. 2002, Koomey et al. 2004, Koomey et al. 1999,
Roth et al. 2002) but out of the controversy grew the first peer-reviewed measurements of
data center electricity use (Blazek et al. 2004, Mitchell-Jackson et al. 2002, Mitchell-
Jackson et al. 2003). Later studies built upon that earlier work to created detailed
measurements of data center power use in multiple facilities (Greenberg et al. 2006,
Tschudi et al. 2004, Tschudi et al. 2006, Tschudi et al. 2003).
Recent growth in the Internet industry has led the popular press to report on increasing
demands for power from data centers (Delaney and Smith 2006, Markoff and Hansell
2006), but these reports are anecdotal and may not reflect aggregate trends. Reports of
large demand growth have also prompted interest from the policy community in
promoting higher efficiency in these facilities (Loper and Parr 2007). The purpose of this
study is to accurately characterize electricity used by servers in the U.S. and the world so
that public debate can proceed based on accurate data instead of the speculation and
hearsay that so often runs rampant in discussions of such topics (Koomey 2003, Koomey
et al. 2002).
PREVIOUS WORK
Several peer-reviewed estimates of power used by servers and data centers were
completed around the year 2000 (Kawamoto et al. 2001, Mitchell-Jackson et al. 2002,
2
Roth et al. 2002). The most detailed and comprehensive of these was that by Roth et al.
(2002), which used aggregate data from IDC (http://www.idc.com) by server class and
measured power data on a representative server for each class. The study also assessed
the electricity used by data storage (tape and hard disk drive) systems and network
equipment.
Unfortunately, little recent peer-reviewed work has been completed in this area. One
exception was an extension of the Roth et al. analysis to 2004 completed by Ton and
Fortenbery (2005) as part of their work on server power supplies. This work used the
same analytical structure as Roth et al. but updated the installed base and power use per
unit estimates to reflect more recent data.
This analysis improves on the Roth et al. analysis for servers by estimating power use for
servers in 2000, 2003, 2004, and 2005, and by using the latest IDC estimates of the
installed base of servers in each class (which are calculated using IDCâs detailed stock
accounting model, not available to Roth). In addition, this analysis relies on power
estimates from a much more detailed assessment of server power use for the most popular
server models in each size class. IDC supplied to me their detailed data on installed base
by server model, which allowed for a more sophisticated attribution of power use to the
most common server models in each class.
DATA AND METHODOLOGY
Data center power use consists of information technology (IT) loads (such as servers,
disk drives, and network equipment) and infrastructure loads (cooling, fans, pumps,
lighting, and uninterruptible power supplies or UPSs). This study focuses on the server
loads (which represent 60-80% of total data center IT loads) and the infrastructure energy
use associated with those servers.
Figure 1
shows conceptually the boundaries of the
study.
The analysis in this report relies on detailed data from IDC <http://www.idc.com/> on the
installed base and shipments of servers, plus measured data and estimates of the power
used per unit for the most common server models in each server class in the U.S. and the
world (including the U.S.). The IDC data (Cohen and Josselyn 2007) are widely
respected and used in the IT industry, but as with all data, they have strengths and
weaknesses, and these must be clearly understood before drawing conclusions from the
data.
Data made available by IDC included
1) Total installed base of servers by server class, historical and projected, for the U.S. and
the World, 1996 to 2010
2) Total shipments of servers by server class, historical and projected, for the U.S. and
the World, 1996 to 2010
3) Installed base of servers by model and manufacturer, for the U.S. and the World, 1998
to 2003
3
4) Shipments of servers by model and manufacturer, for the U.S. and the World, 1996 to
2005
One important component of this analysis is the size of the installed base of servers. IDC
estimates the installed base using data on shipments and equipment lifetimes derived
from manufacturer reporting and market surveys. The server lifetime estimates are based
on reviews of server service contracts and other survey data.
I relied on IDCâs data on aggregate installed base for 2000, 2003, 2004, and 2005, split
into three server classes (volume, mid-range, and high-end), and into U.S. and world
regions. IDC defines these server classes based on the cost of the system: volume
servers cost less than $25,000 per unit, mid-range systems cost between $25,000 and
$500,000 per unit, and each high-end system costs more than $500,000 per unit.
These data include servers in both enterprise and scientific (âhigh performanceâ)
computing applications, and exclude upgrades to existing servers. Blade servers, which
are an important component of recent growth in the total numbers of servers, are
subsumed under the volume server class, with one blade counting as one server. Some
servers that are incorporated into other equipment (such as network equipment) are
counted, depending on where in the supply chain they are sold. The IDC data also
include servers that are not housed in data centersâthe number and location of such
servers may affect the appropriate value for estimating power used by cooling and
associated infrastructure.
Another important category of servers that may be underrepresented in the IDC data is
that of custom servers used by some large Internet companies (such as Google) that are
ordered directly from the manufacturer as personal computer motherboards but are then
used as servers. One estimate reported in the New York Times in June 2006 (Markoff
and Hansell 2006) was that Google owns about 450,000 servers worldwide. It is not
known whether all of these servers are the custom-designed units described above and
how many are standard servers that would have fallen under the IDC âvolume serverâ
category. If all of these servers were added to the volume server category for the world
in 2005 they would increase the volume server installed base by 1.7%. It is also not
known how many other companies have followed Googleâs lead in purchasing such
custom designed âmotherboard serversâ.
Barroso (2005) of Google reported a âtypicalâ power use for low-end servers of 200
Watts, which is close to our estimates for volume server power use per unit in 2004 and
2005 sales in Table 4 (below). Assuming that this power use estimate is correct for
Googleâs servers, the total worldwide electricity use for volume servers would also go up
about 1.7% if these servers were added to our totals.
The general approach for calculating total power use was to estimate a power use per
server that could be multiplied by IDCâs total installed base. I assumed that the weighted
average power per unit of the six most popular models in each class in the installed base
would be a reasonable proxy for the average power use per unit. This approach assumes
that the models chosen accurately characterize the servers in a given class, and that
4
assumption should be assessed in the future as more accurate data become available. The
installed base by model was only available through 2003, so adjustments were required
for 2004 and 2005 to estimate power used by servers in 2005.
IDCâs total installed base estimates (plus shipments and implied retirements) are shown
in
Table 1
. Volume servers dominate the installed base, representing 90-95% of the
servers on a unit basis. Mid-range servers comprise most of the rest, with the high-end
servers only responsible for a few tenths of one percent of the total on a unit basis. The
U.S. houses between 30 and 40% of the servers in the world, depending on server class.
Table 2
shows the six most popular server models in each server class in terms of
installed base in 2000, based on IDC data for the U.S. and the world,
Table 3
shows the
same results for 2003, and
Table 4
shows the top three most popular server models
shipped in 2005. The exact installed base and shipment numbers by model are not
shown here because of confidentiality concerns, but the
total
installed base or % of total
shipments represented by the servers shown in Tables 2, 3, and 4 are shown in
Table 5
.
The most popular models comprise a larger share in the U.S. than in the world. These
results indicate that the U.S. market is more concentrated on a model basis than is the
world market. The most popular servers comprise significant percentages of the installed
base (between 16% and 38%) and an even larger share of the 2005 shipments.
I use the market share for each of the most popular servers to calculate a weighted
average power use per unit for each server class. For example, if each of the three most
popular US volume servers in Table 4 have 10% of the market, the âweightâ for each
serverâs power use will be 10% divided by 30%, or 33.33%.
Estimating power use for each server is not easy. The power use of electronic equipment
varies with hardware configuration, usage, and environmental conditions. The power
supplies for these devices are sized for the maximum loads expected when the server is
fully configured, so the actual measured loads observed in typical installations are much
lower than the rated power of the power supply.
Unfortunately, measured data on energy use are not commonly made available for most
servers, especially older ones (that is changing for more recent modelsâsee ASHRAE
(2004) and the discussion of âTypical power use per unitâ in the future work section). In
addition, estimating power use is complicated because each server can have multiple disk
drives, vary significantly in installed memory, and commonly have multiple processors
and redundant power supplies.
I assign a power use per unit for each server based on measured data, on-line server
configuration calculators, or estimates from manufacturer specification sheets. When
âtypicalâ measured power was not available, I multiplied the maximum measured
electricity use or the maximum rated input power of the power supply by factors taken
from industry experience to estimate typical power use.
Maximum measured electricity use is widely reported by some manufacturers (e.g. IBM)
and for others (e.g., HP, Dell) it is possible to calculate it using on-line configuration
5
tools. When I used such tools to estimate maximum measured power, I included the
maximum amount of RAM, the largest size and number of hard drives, n+1 redundant
power supplies, processor-intensive workloads, and two processors at the fastest clock
speeds. To convert such power estimates for high end servers to typical power use I
multiplied maximum measured power by 66%, which is the rule of thumb accepted by
IBM. For volume and mid-range machines, I used a factor of 40%, after discussions with
power experts at Dell, which reflects the lower utilization and less dense configurations
common with the smaller machines.
When neither typical nor measured maximum power was available, the last resort was to
use the maximum rated input power of the power supply taken from the specification
sheet for the server. Sometimes these sheets only report the
output
power of the power
supply, and in that case I divided the output power by 70% efficiency (which is typical
for server power supply efficiency) to estimate the maximum rated input power. To
estimate typical power, I multiplied rated input power by 25%, 30%, or 40% for volume,
mid-range, and high-end servers, respectively. This range of factors reflects recent
research (Ton and Fortenbery 2005) and industry experience (based on my conversations
with Intel, IBM, Dell, and other manufacturers). Use of these factors yields results
comparable to that for servers for which I do have measured data.
1
Table 6
summarizes the calculation of total power use. The installed base estimates are
taken from Table 1, and average power use per unit data for 2000 and 2003 are the
weighted averages from Table 2 and 3, respectively. The average power use per unit in
2004 is calculated assuming that the retirements in 2004 (from Table 1) have the same
power use per unit as the average for the year 2000 installed base, and that the shipments
in 2004 have the same power use per unit as the weighted average power of new
equipment from Table 4. The procedure is repeated in 2005, assuming that the
retirements in that year also use the same amount of power per unit as the average for the
year 2000 installed base.
Direct power consumption (million kW) is the product of installed base and average
power use per unit, while the direct electricity consumption (billion kWh) is calculated
from the direct power consumption assuming that servers operate 100% of the year (i.e.
with a load factor of 100%).
The total power consumption associated with servers in data centers also includes the
electricity use of cooling and auxiliary equipment. The Uptime Institute characterizes
such infrastructure loads using the Site Infrastructure Energy Efficiency Ratio or SI-EER
(Brill 2007a), which is the ratio of total data center electricity load to IT electricity load.
Malone and Belady (2006) call that same ratio âPower Usage Effectivenessâ, and itâs
typically about 2, based on the detailed results extracted from recent benchmarking
measurements for data centers (Greenberg et al. 2006). An SI-EER of 2 means that total
1
For details on the power calculations on a server-by-server basis, email Jonathan Koomey at
JGKoomey@stanford.edu.
6
loads are double the IT loadâsaid a different way, every kWh of electricity use for IT
loads means another kWh of electricity use for infrastructure. I apply an SI-EER of 2 to
direct server loads in Table 6 to get total loads associated with servers (this approach
assumes that all servers in IDCâs installed base are located in data centers, an assumption
that should be tested as more detailed data become available).
The total electricity bill is calculated assuming U.S. industrial electricity prices for 2000,
2003, 2004, and 2005, taken from the Energy Information Administrationâs
Electric
Power Annual
(http://www.eia.doe.gov/cneaf/electricity/epa/epat7p4.html) and adjusted
to 2006 dollars using the implicit GDP deflator. Most data centers are large enough to
qualify for industrial rates in the U.S. Because no comparable price data exist for the
world, I also apply the U.S. industrial prices to world consumption.
RESULTS
We explore different dimensions of the results below, beginning by summarizing total
electricity use for servers in the U.S. and the world, comparing year 2000 to previous
results, and then analyzing the changes in key drivers of the results from 2000 to 2005.
Total electricity use and expenditures
Electricity use associated with servers doubled from 2000 to 2005, representing an
aggregate annual growth rate of 14% per year for the U.S. and 16% per year for the
world. Almost all of this growth is attributable to growth in the numbers of servers
(particularly volume servers), with only a small percentage associated with increases in
the power use per unit.
Total direct power consumption for all servers in the U.S. in 2005 is about 2.6 million
kW. Including cooling and auxiliary equipment increases that total to about five million
kW, which is equivalent (in capacity terms) to five 1000 MW power plants. Total server
electricity consumption in the U.S. is 23 billion kWh in 2005. When electricity use for
cooling and auxiliary equipment is included, that total rises to 45 billion kWh, or about
1.2% of retail electricity sales in that year
2
, resulting in a total utility bill of $2.7 billion
(2006 dollars) when valued at U.S. industrial electricity prices (see
Figure 2
). Total
server power and electricity consumption for the world as a whole is about two and a half
times bigger than for the U.S.
Comparisons to previous analysis
Figure 3
shows ratios of results from this study to those from Roth et al. (2002) for the
U.S. in the year 2000 (Roth found total electricity used by servers in the U.S. in 2000 to
2
U.S.
retail
electricity
sales
in
2005
were
3661
billion
kWh
<http://www.eia.doe.gov/cneaf/electricity/epa/epat7p2.html >. World electricity sales in 2005 are estimated
to be about 14,700 billion kWh, derived from US DOE. 2006.
International Energy Outlook 2006
.
Washington, DC: Energy Information Administration, U.S. Department of Energy. DOE/EIA-0484(2006).
June. (http://eia.doe.gov/oiaf/ieo/).
7
be 10.1 billion kWh). The split between mid-range and high-end servers was different in
Roth et al. and so I lump those two classes together to create a consistent comparison
(Koomey 2001).
The estimates of installed base in 2000 in this study are about 20% greater for volume
servers and about 10% less for the larger servers. The bigger differences are in the power
use per unit, where this studyâs estimate for volume server power use per unit is about
50% greater than that for Roth, while the estimated power use for the larger servers is
about 30% less. Coincidentally, the weighted average power use per unit across all
servers is about the same, so the larger number of units is what drives this studyâs results
to be about 16% higher than the total electricity use estimates in Roth et al.
Changes from 2000 to 2005
Figure 4
shows ratios of 2005 to 2000 results for the U.S. The power use per unit for all
product types is higher in 2005 than in 2000, but the shifts in installed base (with the
number of units of volume servers doubling and the number of mid-range servers
declining about 40%) mitigate the effects of higher per unit power levels. The overall
increase in total electricity use is driven almost entirely by the increase in the number of
volume servers.
Figure 5
shows the same results for the world. The story is strikingly
similar, with the additional result that the total installed base is growing more quickly in
the world than in the U.S. over this period.
FUTURE WORK
Distribution of power use over models
The most popular models represent a significant portion of each server class, but it is
important to investigate the implications of this approach for the accuracy of the
estimated power use per unit. The gold standard of accuracy in this regard would be to
estimate power use for each and every model in the installed base and weight-average
that power use per unit across the entire installed base. Time and resource constraints
prevent such a comprehensive approach, so a reduced sample was needed, but choosing
that sample can be difficult. The accuracy of the choices depends in part on the shape of
the distribution and where the most popular models fall in that distribution. The popular
models may be the less costly and complex ones, but the relationship between cost of
information technology (IT) equipment and energy use is tenuous at best.
The most popular volume servers only cover 16% of the installed base in 2003, but
fortunately that class of servers tends to be more homogeneous than the larger systems.
The high-end systems are the most heterogeneous, but the most popular models for this
class of server cover a much larger portion of the installed base (29% to 36%), which
mitigates to some extent the effect on accuracy the shape of the distribution might
introduce in that instance. Further work is clearly needed on this issue as more accurate
data become available.
8
Estimating typical power use per unit
Further analysis is needed on the relationship among typical power, rated input power,
and maximum measured power. The factors used in this analysis are based on
recommendations from technical representatives of the major manufacturers, but
measured data using a standardized protocol is the most accurate way to arrive at typical
power use for servers in the field (ASHRAE 2004). Computational workloads also have
an important effect on power use per unit (particularly as new power-saving technologies
are incorporated into the equipment) but data on this effect are sparse. That situation
should improve as new protocols for energy measurements for servers come into effect in
coming months (for example, see <http://www.energystar.gov/datacenters> and
<http://www.spec.org/specpower/pressrelease.html>).
More detailed segmentation of the server markets
The aggregate segmentation of the server market used in this analysis masks some
important variations in that market. The high performance computing (HPC) market, for
example, has different usage patterns than do business servers, but even within the
business category there are large variations. For example, Internet search is much
different than web hosting, which is much different from application hosting. In addition,
variations in the physical characteristics of mid-range and high-end servers may interact
in complex ways with different usage patterns and have a substantial impact on power
use per unit. More detailed market segmentation may be helpful in disentangling some of
these effects.
Forecasts of future electricity use
It is particularly difficult to forecast trends in the IT industry. If the current IDC
worldwide forecast holds true, installed base for volume servers will grow by more than
50% from 2005 levels by 2010, while mid-range and high-end installed base will decline
20-30%. If power per server remains constant, those trends would imply an increase in
electricity used by servers worldwide of about 40% by 2010. If in addition the average
power use per unit goes up at the same rate for each class as our analysis indicates that it
did from 2000 to 2005, total electricity used by servers by 2010 would be 76% higher
than it was in 2005.
The IDC forecast incorporates several trends that will affect power used by servers,
including the move to more use of blade servers (which will tend to increase power use),
and the shift to consolidation and virtualization (which will tend to decrease power use by
reducing the number of physical servers that are needed). The industry has recently
become more sensitive to total cost of ownership for these facilities, driven by the
increasing importance of infrastructure and utility costs relative to IT costs (Brill 2007a).
The total cost of building a large data center is now on the order of $100 to $200M,
which is sufficient to get the attention of the CEO of most large organizations. That
visibility to corporate management is likely to drive operational and design
improvements that should over time improve the Site Infrastructure Energy Efficiency
9
Ratio and spur the adoption of energy metrics and purchasing standards for efficiency of
IT equipment within these companies (Brill 2007a, Brill 2007b).
Total power used by data centers
This analysis focused on the most important component of electricity used in data centers
(servers). Similar analyses are needed for data storage and network equipment, so that the
total power use of all data centers can be estimated. Roth et al. (2002) found that
electricity use for separate data storage devices in 2000 was 1.6 billion kWh or about
16% compared to the 10.1 billion kWh for U.S. servers alone. That study was not able to
separate network equipment energy into the component found in data centers and that
found in other facilities, and that is a key area where more data and analysis will be
needed. It will also be necessary to separate the installed base of servers into those
housed in data centers and those that are not, to more accurately estimate the SI-EER.
Opportunities for efficiency improvements
Previous work indicates substantial potential for improving the design and operation of
IT equipment and data center infrastructure (Brill 2007a, Eubank et al. 2004, Greenberg
et al. 2006, Tschudi et al. 2004, Tschudi et al. 2006), but additional data collection,
analysis, and policy changes are needed to realize those improvements in the real world.
Many of the changes required to capture those potential savings are institutional in
nature, and involve addressing the misplaced incentives that pervade the industry.
Current market structures and practices are aligned with minimization of first cost instead
of reduction of total cost of ownership, but that will change as new metrics are adopted
and companies change design and purchasing practices to reflect the new emphasis on
minimizing total costs.
CONCLUSIONS
The amount of electricity used by servers and other Internet infrastructure has become an
important issue in recent years. This study estimates total electricity used by servers in
the U.S. and the world by combining IDC data on the installed base with measured data
and estimates of power used by the most popular servers. These estimates are based on
more detailed data than previous assessments, and they will help policy makers and
businesses attempting to make sense of recent trends in this industry.
Aggregate electricity use for servers doubled over the period 2000 to 2005 both in the
U.S. and worldwide. Almost all of this growth was the result of growth in the number of
volume servers, with only a small part of that growth being attributable to growth in the
power use per unit. Total power used by servers represented about 0.6% of total U.S.
electricity consumption in 2005. When cooling and auxiliary infrastructure are included,
that number grows to 1.2%, an amount comparable to that for color televisions. The total
power demand in 2005 (including associated infrastructure) is equivalent (in capacity
terms) to about five 1000 MW power plants for the U.S. and 14 such plants for the world.
The total electricity bill for operating those servers and associated infrastructure in 2005
was about $2.7 B and $7.2 B for the U.S. and the world, respectively.
10
ACKNOWLEDGMENTS
This report was produced with a grant from Advanced Micro Devices (AMD), âin-kindâ
support and data from IDC, and independent review comments from experts throughout
the industry. All errors and omissions are the responsibility of the author alone.
I would like to thank Andrew Fox, Larry Vertal, Donna Sadowy, and Bruce Shaw from
AMD and Sarahjane Sacchetti from Bite Communications for their unwavering support
in the course of this project. I would also like to express my special thanks to Lloyd
Cohen and Vernon Turner at IDC, who supported the work by sharing key data and
patiently answering my numerous questions. In addition, I would like to express my
appreciation to colleagues at the technology companies who supplied me with data and
insights on the power used by server equipment, including (in alphabetical order by
company) David Moss and Greg Darnell at Dell, Christian Belady and Klaus Dieter-
Lange at HP, Joe Prisco at IBM, Henry Wong at Intel, and Ed Hunter and Dennis
Symanski of Sun.
Finally, I would like to thank the technical reviewers for their insights and comments.
The reviewers included (in alphabetical order by company):
Brent Kerby (AMD)
Neil Rasmussen (APC)
Michele Blazek (AT&T)
David Moss (Dell)
Chris Calwell (Ecos Consulting)
Steve Wade, John Cymbalsky, and Erin Boedecker, Energy Information Administration
Andrew Fanara (EPA)
Brian Fortenbery (EPRI Solutions)
Christian Belady (HP)
Joe Prisco (IBM)
Rebecca Duff (ICF)
Henry Wong & Michael Patterson (Intel)
Bruce Nordman and Bill Tschudi (LBNL)
Noah Horowitz (NRDC)
Mark Bramfitt (PG&E)
Joel Swisher & Amory Lovins (Rocky Mountain Institute)
Peter Rumsey (Rumsey Engineers)
Phil Reese (Stanford University)
Ed Hunter & Subodh Bapat (Sun)
Kurt Roth (Tiax)
Paul Komor (U of Colorado, Boulder)
Ken Brill (Uptime institute)
11
ABOUT THE AUTHOR
Jonathan Koomey is a Staff Scientist at Lawrence Berkeley National Laboratory and a
Consulting Professor at Stanford University. Dr. Koomey is one of the foremost
international experts on electricity used by computers, office equipment, and data centers,
and is the author or co-author of eight books and more than one hundred and fifty articles
and reports on energy and environmental economics, technology, forecasting, and policy.
He has also published extensively on critical thinking skills. He holds M.S. and Ph.D.
degrees from the Energy and Resources Group at the University of California at
Berkeley, and an A.B. in History of Science from Harvard University. In 1993 he won
the Fred Burgraff Award for Excellence in Transportation Research from the National
Research Council's Transportation Research Board. He was named an Aldo Leopold
Leadership Fellow in 2004 and an AT&T Industrial Ecology Fellow in January 2005. He
has been quoted in the New York Times, the Wall Street Journal, Barronâs, The
Washington Post, Science, Science News, American Scientist, Dow Jones News Wires,
and the Christian Science Monitor, and has appeared on Nova/Frontline, BBC radio,
CNBC, All Things Considered, Marketplace, On the Media, Tech Nation, and the
California Report, among others. His latest solo book is
Turning Numbers into
Knowledge: Mastering the Art of Problem Solving
<http://www.analyticspress.com>,
now in its third printing (and recently translated into Chinese).
For more biographical
details and a complete publications list, go to <http://www.koomey.com>.
12
REFERENCES
ASHRAE. 2004.
Thermal Guidelines for Data Processing Environments
. Atlanta, GA:
American Society of Heating, Refrigeration, and Air-Conditioning Engineers.
(http://www.ashrae.org)
Baer, Walter S., Scott Hassell, and Ben Vollaard. 2002.
Electricity Requirements for a
Digital Society
. RAND Corporation. MR-1617-DOE, ISBN 0-8330-3279-8.
(http://www.rand.org/publications/MR/MR1617/)
Barroso, Luiz AndrĂŠ. 2005. "The Price of Performance: An Economic Case for Chip
Multiprocessing."
ACM Queue, special issue on Multiprocessors
. vol. 3, no. 7.
September.
Blazek, Michele, Huimin Chong, Woonsien Loh, and Jonathan Koomey. 2004. "A data
center revisited: Assessment of the energy impacts of retrofits and technology
trends in a high-density computing facility."
The ASCE Journal of Infrastructure
Systems
. vol. 10, no. 3. September. pp. 98-104.
Brill, Kenneth G. 2007a.
Data Center Energy Efficiency and Productivity
. Santa Fe, NM:
The
Uptime
Institute.
(http://www.upsite.com/cgi-
bin/admin/admin.pl?admin=view_whitepapers)
Brill, Kenneth G. 2007b.
High-Density Computing: The Path Foward 2006
. Santa Fe,
NM:
The
Uptime
Institute.
(http://www.upsite.com/cgi-
bin/admin/admin.pl?admin=view_whitepapers)
Cohen, Lloyd, and Stephen L. Josselyn. 2007.
Market Analysis: U.S. and Worldwide
Server Installed Base 2006-2010 Forecast
. IDC. Draft report. January.
Delaney, Kevin J., and Rebecca Smith. 2006. "Surge in Internet Use, Energy Costs Has
Big Tech Firms Seeking Power."
The Wall Street Journal
. New York, NY. June
13. p. A1.
Eubank, Huston, Joel Swisher, Cameron Burns, Jen Seal, and Ben Emerson. 2004.
Design Recommendations for High-Performance Data Centers: Report of the
Integrated Design Charrette (conducted 2-4 February 2003)
. Old Snowmass, CO:
Rocky Mountain Institute.
Greenberg, Steve, Evan Mills, William Tschudi, Peter Rumsey, and Bruce Myatt. 2006.
Best Practices for Data Centers: Results from Benchmarking 22 Data Centers
.
Proceedings of the 2003 ACEEE Summer Study on Energy Efficiency in
Buildings. Asilomar, CA: American Council for an Energy Efficient Economy, in
Washington, DC. August.
13
Harris, Jeff, J. Roturier, L.K. Norford, and A. Rabl. 1988.
Technology Assessment:
Electronic Office Equipment
. Lawrence Berkeley Laboratory. LBL-25558.
November.
Huber, Peter, and Mark P. Mills. 1999. "Dig more coalâthe PCs are coming." In
Forbes
.
May 31. pp. 70-72.
Johnson, Brian J., and Catherine R. Zoi. 1992. "EPA Energy Star Computers: The Next
Generation of Office Equipment." In
Proceedings of the 1992 ACEEE Summer
Study on Energy Efficiency in Buildings.
Edited by Asilomar, CA: American
Council for an Energy Efficient Economy. pp. 6.107-6.114.
Kawamoto, Kaoru, Jonathan Koomey, Bruce Nordman, Richard E. Brown, Maryann
Piette, Michael Ting, and Alan Meier. 2002. "Electricity Used by Office
Equipment and Network Equipment in the U.S."
EnergyâThe International
Journal (also LBNL-45917)
. vol. 27, no. 3. March. pp. 255-269.
Kawamoto, Kaoru, Jonathan Koomey, Mike Ting, Bruce Nordman, Richard E. Brown,
Maryann Piette, and Alan Meier. 2001.
Electricity Used by Office Equipment and
Network Equipment in the U.S.: Detailed Report and Appendices
. Berkeley, CA:
Lawrence Berkeley National Laboratory. LBNL-45917. February.
(http://enduse.lbl.gov/Projects/InfoTech.html)
Koomey, Jonathan. 2001.
Turning Numbers into Knowledge: Mastering the Art of
Problem
Solving
.
Oakland,
CA:
Analytics
Press.
(http://www.numbersintoknowledge.com)
Koomey, Jonathan. 2003. "Sorry, Wrong Number: Separating Fact from Fiction in the
Information Age." In
IEEE Spectrum
. June. pp. 11-12.
Koomey, Jonathan, Chris Calwell, Skip Laitner, Jane Thornton, Richard E. Brown, Joe
Eto, Carrie Webber, and Cathy Cullicott. 2002. "Sorry, wrong number: The use
and misuse of numerical facts in analysis and media reporting of energy issues."
In
Annual Review of Energy and the Environment 2002
. Edited by R. H. Socolow,
D. Anderson and J. Harte. Palo Alto, CA: Annual Reviews, Inc. (also LBNL-
50499). pp. 119-158.
Koomey, Jonathan, Huimin Chong, Woonsien Loh, Bruce Nordman, and Michele
Blazek. 2004. "Network electricity use associated with wireless personal digital
assistants."
The ASCE Journal of Infrastructure Systems (also LBNL-54105)
.
vol. 10, no. 3. September. pp. 131-137.
Koomey, Jonathan, Kaoru Kawamoto, Bruce Nordman, Mary Ann Piette, and Richard E.
Brown. 1999.
Initial comments on 'The Internet Begins with Coal'
. Berkeley, CA:
Lawrence Berkeley National Laboratory. LBNL-44698. December 9.
(http://enduse.lbl.gov/projects/infotech.html)
14
Koomey, Jonathan, Mary Ann Piette, Mike Cramer, and Joe Eto. 1996. "Efficiency
Improvements in U.S. Office Equipment: Expected Policy Impacts and
Uncertainties."
Energy Policy
. vol. 24, no. 12. December. pp. 1101-1110.
Ledbetter, M, and L. Smith. 1993.
Guide to Energy-Efficient Office Equipment
.
Washington, DC: American Council for an Energy-Efficient Economy.
Loper, Joe, and Sara Parr. 2007.
Energy Efficiency in Data Centers: A New Policy
Frontier
. Washington, DC: Alliance to Save Energy. February.
Lovins, Amory, and H. Heede. 1990.
Electricity-Saving Office Equipment
.
Competitek/Rocky Mountain Institute. September.
Malone, Christopher, and Christian Belady. 2006.
Metrics to Characterize Data Center
& IT Equipment Energy Use
. Proceedings of the Digital Power Forum. Dallas,
TX: September 18-20.
Markoff, John, and Saul Hansell. 2006. "Hiding in Plain Sight, Google Seeks an
Expansion of Power."
The New York Times
. New York, NY. June 14.
Mills, Mark P. 1999.
The Internet Begins with Coal: A Preliminary Exploration of the
Impact of the Internet on Electricity Consumption
. Arlington, VA: The Greening
Earth Society. May.
Mitchell-Jackson, Jennifer, Jonathan Koomey, Michele Blazek, and Bruce Nordman.
2002. "National and Regional Implications of Internet Data Center Growth."
Resources, Conservation, and Recycling (also LBNL-50534)
. vol. 36, no. 3.
October. pp. 175-185.
Mitchell-Jackson, Jennifer, Jonathan Koomey, Bruce Nordman, and Michele Blazek.
2003. "Data Center Power Requirements: Measurements From Silicon Valley."
EnergyâThe International Journal (also LBNL-48554)
. vol. 28, no. 8. June. pp.
837 - 850.
Norford, Les, A. Hatcher, Jeffrey Harris, Jacques Roturier, and O. Yu. 1990. "Electricity
Use in Information Technologies." In
Annual Review of Energy 1990
. Edited by
J. M. Hollander. Palo Alto, CA: Annual Reviews, Inc. pp. 423-53.
Piette, Maryann, Joe Eto, and Jeff Harris. 1991.
Office Equipment Energy Use and
Trends
. Lawrence Berkeley Laboratory. LBL-31308. September.
Roth, Kurt, Fred Goldstein, and Jonathan Kleinman. 2002.
Energy Consumption by
Office and Telecommunications Equipment in Commercial Buildings--Volume I:
Energy Consumption Baseline
. Washington, DC: Prepared by Arthur D. Little for
the U.S. Department of Energy. A.D. Little Reference no. 72895-00. January.
(http://www.eren.doe.gov/buildings/documents)
15
Ton, My, and Brian Fortenbery. 2005.
Server Power Supplies
. Berkeley CA: Report to
Lawrence Berkeley National Laboratory by Ecos Consulting and EPRI Solutions.
December.
Tschudi, Bill, Tengfang Xu, Dale Sartor, Bruce Nordman, Jon Koomey, and Osman
Sezgen. 2004.
Energy Efficient Data Centers
. Berkeley, CA: LBNL, for the
California Energy Commission. LBNL-54163. March 30.
Tschudi, William, Evan Mills, Steve Greenberg, and Peter Rumsey. 2006. "Measuring
and Managing Data Center Energy Use: Findingsâand resulting best practicesâ
from a study of energy use in 22 data centers." In
Heating, Piping, Air-
conditioning (HPAC) engineering
. March. pp. 45-51.
Tschudi, William, Priya Sreedharan, Tengfang Xu, David Coup, and Paul Roggensack.
2003.
Data Centers and Energy Use â Letâs Look at the Data
. Proceedings of the
2003 ACEEE Summer Study on Energy Efficiency in Industry. Rye Brook, NY:
American Council for an Energy Efficient Economy, in Washington, DC. July.
16
Figure 1: Boundaries of this study
This study focuses on the largest single component of data center power use: servers and
the associated cooling and auxiliary equipment (C&A) needed to support them. Other
important components of data center electricity use include data storage and network
equipment, which together represent 20 to 40% of total data center load, but accurate
recent data on power use has not yet been developed for this equipment. C&A power
includes losses associated with the backup power systems, power conditioning, power
distribution, air handling, lighting, and chillers.
Server power
Data
storage
power
Network
equipment
power
Cooling and auxiliaries (C&A) associated
with server power
C&A for
data storage
C&A for
networking
Total Data Center Power
The focus of this study
17
Figure 2: Total electricity use for servers in the U.S. and the world in 2000 and
2005, including cooling and auxiliary equipment
18
Figure 3: Comparison of year 2000 estimates in this study and those of Roth et al.,
including U.S. installed base, power use per unit, and total server electricity use.
19
Figure 4: Comparison of 2005 to 2000 installed base, power per unit, and total
electricity use for the U.S.
20
Figure 5: Comparison of 2005 to 2000 installed base, power per unit, and total
electricity use for the World
Table 1: Installed base, shipments, and retirements of servers for the U.S. and the World
Units
Volume
Mid-range
High-end
Total
Volume
Mid-range
High-end
Total
Total Installed Base (1, 2, 3)
2000
Thousands
4,927
663
23.0
5,613
12,240
1,808
65.6
14,114
2001
Thousands
5,907
701
22.5
6,630
15,596
1,890
69.1
17,555
2002
Thousands
6,768
574
23.1
7,365
16,750
1,683
59.0
18,492
2003
Thousands
7,578
530
21.4
8,130
18,523
1,540
62.3
20,125
2004
Thousands
8,658
432
23.3
9,113
23,441
1,238
66.0
24,746
2005
Thousands
9,897
387
22.2
10,306
25,959
1,264
59.4
27,282
Total Shipments (2, 3)
2000
Thousands
1,659
111
4.8
1,774
3,926
283
13.0
4,223
2001
Thousands
1,492
66
3.6
1,562
3,981
206
10.4
4,198
2002
Thousands
1,714
67
3.1
1,784
4,184
204
9.4
4,397
2003
Thousands
2,069
76
2.9
2,148
5,017
211
8.8
5,237
2004
Thousands
2,517
53
2.8
2,572
6,083
184
8.6
6,275
2005
Thousands
2,721
62
2.6
2,786
6,822
187
8.5
7,017
Total Retirements (4)
2000
Thousands
300
116
5
420
1,631
264
10
1,905
2001
Thousands
513
28
4.1
545
626
125
6.9
757
2002
Thousands
853
194
2.5
1,049
3,030
411
19.6
3,461
2003
Thousands
1,259
120
4.6
1,383
3,243
355
5.5
3,603
2004
Thousands
1,437
151
0.9
1,589
1,165
485
4.9
1,655
2005
Thousands
1,482
106
3.7
1,592
4,304
161
15.1
4,481
For questions, contact Jonathan Koomey, 510-708-1970 C, jgkoomey@stanford.edu
(1) Installed base is measured at the end of any given year (December 31st).
(2) Installed base and shipments for all years from IDC data, filename IDC_QShare_InstalledBaseForecast2006.xls, release date February 3, 2006.
(3) Installed base and shipments include both enterprise and scientific servers. They do not include server upgrades.
(4) Retirements are derived from the installed base and shipments numbers. Retirements in 2000 derived using 1999 installed base (not shown) and year 2000 shipments.
(5) World includes the U.S.
All servers: U.S.
All servers: World (5)
Table 2: Top six models by server class for the U.S. and the World based on IDC 2000 installed base data
Volume servers: US
Volume servers: World
Typical
Typical
Brand
Model
Power (W) Notes
Brand
Model
Power (W)
Notes
Compaq
ML530
273
3, 6
Compaq
ML370
268
3, 6
Compaq
ML370
268
3, 6
Compaq
ML530
273
3, 6
Dell
2300
107
2, 7
Compaq
200 Total
114
3, 6
Compaq
DL380
131
3, 6
IBM
3000 Total
100
4, 6
Compaq
ML350
139
3, 6
HP
E Series
114
3, 6
Dell
1300
118
2, 7
Compaq
ML350
139
3, 6
Weighted average
186
Weighted average
183
Mid-Range servers: US
Mid-Range servers: World
Typical
Typical
Brand
Model
Power (W) Notes
Brand
Model
Power (W)
Notes
Sun
450
499
5, 7
Sun
450 Total
499
5, 7
Sun
3000/3500
263
5, 7
Sun
3000/3500
263
5, 7
Sun
4000/4500
432
5, 7
IBM
POWERSERVER C10/C20/E20/E30
106
4, 6
IBM
POWERSERVER C10/C20/E20/E30
106
4, 6
HP
rp 5400 Series / D CLASS 200/U
847
3, 6
IBM
9406 300
327
4, 6
Sun
4000/4500
432
5, 7
HP
K CLASS 100
872
3, 6
IBM
9406 300
327
4, 6
Weighted average
424
Weighted average
423
High-end servers: US
High-end servers: World
Typical
Typical
Brand
Model
Power (W) Notes
Brand
Model
Power (W)
Notes
Sun
10000 HE
13,456
5
IBM
POWERSERVER S80
1,335
4, 6
IBM
POWERSERVER S80
1,335
4, 6
Sun
10000 HE
13,456
5
IBM
SP Mid B9
2,640
4, 6
IBM
SP Mid B9
2,640
4, 6
IBM
4381
4,813
4, 6
IBM
4381
4,813
4, 6
HP
rp 8400 Series / V CLASS 2200/2250
2,920
3
HP
rp 8400 Series / V CLASS 2500 MR
2,920
3
HP
991 995 996
1,724
3
IBM
9406 640
1,327
4, 6
Weighted average
5,534
Weighted average
4,874
For questions, contact Jonathan Koomey, 510-708-1970 C, jgkoomey@stanford.edu
(1) For each server class, models shown are ranked in order of their share of installed base in 2003. Shares not shown because of confidentiality concerns.
(2) Dell 1300 and 2300 typical power based on maximum rated input power from Dell spec sheets.
(3) Compaq ML 370, ML350, ML530, and DL380 all assume G1 versions. Typical power based on max. measured power from HP online
configurator <http://h30099.www3.hp.com/configurator/calc/Power Calculator Catalog.xls> (N.B., HP and Compaq merged after 2000).
HP rp8400 V class and 991/995/996 typical power given in spec sheets.
(4) IBM models' typical power based on max. measured power from IBM estimates (for SP mid B9) or from spec sheets (all others).
(5) Sun models' typical power based on maximum rated input power from spec sheets, except for the 10000 HE for which typical power was given.
(6) Max. measured power multiplied by 40% and 66% to get typical power use for volume and mid-range/high end servers respectively.
(7) Max. rated input power multiplied by 25%, 30%, and 40% to get typical power for volume, mid-range, and high end servers, respectively.
Table 3: Top six models by server class for the U.S. and the world based on IDC 2003 installed base data
Volume servers: US
Volume servers: World
Typical
Typical
Brand
Model
Power (W)
Notes
Brand
Model
Power (W)
Notes
HP/Compaq
1600/ML 370
293
3, 6
HP/Compaq
1600/ML 370
293
3, 6
Dell
2650
178
2, 6
HP/Compaq
DL380 G2
150
3, 6
HP/Compaq
DL360
107
3, 6
HP/Compaq
DL360
107
3, 6
HP/Compaq
800/ ML350
165
3, 6
HP/Compaq
DL380 G3
212
3, 6
HP/Compaq
DL 380 G1 DL 380
131
3, 6
Dell
2650
178
2, 6
HP/Compaq
3000/ML 530
345
3, 6
HP/Compaq
3000/ML 530
345
3, 6
Weighted average
207
Weighted average
214
Mid-Range servers: US
Mid-Range servers: World
Typical
Typical
Brand
Model
Power (W)
Notes
Brand
Model
Power (W)
Notes
Sun
450
499
5, 7
Sun
450
499
5, 7
Sun
420R
183
5, 7
Sun
420R
183
5, 7
Sun
V480
432
5, 7
Sun
V880
450
5, 7
Sun
V880
450
5, 7
Sun
V480
432
5, 7
Sun
4000/4500
432
5, 7
HP
rp 7400 Series / N CLASS (rp7410)
1700
3
HP
rp 7400 Series/N CLASS (rp7410)
1700
3
IBM
9406-270
287
4, 6
Weighted average
524
Weighted average
522
High-end servers: US
High-end servers: World
Typical
Typical
Brand
Model
Power (W)
Notes
Brand
Model
Power (W)
Notes
Sun
10000 HE
13,456
5
IBM
p680-S85
1,335
4, 6
IBM
p690-681
11,286
4, 6
IBM
p690-681
11,286
4, 6
IBM
POWERSERVER S80
1,335
4, 6
IBM
POWERSERVER S80
1,335
4, 6
IBM
p680-S85
1,335
4, 6
Sun
10000 HE
13,456
5
HP
rp 8400 Series/V CLASS 2200/2250
2,920
3
HP
rp 8400 Series / V CLASS 2200/2250
2,920
3
IBM
SP Mid B9
2,640
4, 6
IBM
SP Mid B9
2,640
4, 6
Weighted average
6,428
Weighted average
5,815
For questions, contact Jonathan Koomey, 510-708-1970 C, jgkoomey@stanford.edu
(1) For each server class, models shown are ranked in order of their share of installed base in 2003. Shares not shown because of confidentiality concerns.
(2) Dell 2650 typical power based on maximum measured power from Dell online configurator <http://www.dell.com/calc>
(3) HP 1600/ML 370, 800/ML350, 3000/ML530, and DL360 all assume G2 versions. Typical power based on max. measured power from HP online
configurator <http://h30099.www3.hp.com/configurator/calc/Power Calculator Catalog.xls>. HP rp8400 V class and rp7400 N class typical power
given in spec sheets.
(4) IBM models' typical power based on max. measured power from IBM estimates (for SP mid B9) or from spec sheets (all others).
(5) Sun models' typical power based on maximum rated input power from spec sheets, except for the 10000 HE for which typical power was given.
(6) Max. measured power multiplied by 40% and 66% to get typical power use for volume and mid-range/high end servers respectively.
(7) Max. rated input power multiplied by 25%, 30%, and 40% to get typical power for volume, mid-range, and high end servers, respectively.
Table 4: Top three models by server class for the U.S. and the world based on IDC 2005 shipments data
Volume servers: US
Volume servers: World
Typical
Typical
Brand
Model
Power (W)
Notes
Brand
Model
Power (W)
Notes
Dell
2850
231
2, 6
HP
DL380
222
3, 6
HP
DL380
222
3, 6
Dell
2850
231
2, 6
HP
DL360
187
3, 6
HP
DL360
187
3, 6
Weighted average
217
Weighted average
218
Mid-Range servers: US
Mid-Range servers: World
Typical
Typical
Brand
Model
Power (W)
Notes
Brand
Model
Power (W)
Notes
IBM
i5-520
495
4, 6
IBM
i5-520
495
4, 6
IBM
p5 570
858
4, 6
IBM
p5 570
858
4, 6
Sun
V490
554
5, 7
Sun
V490
554
5, 7
Weighted average
641
Weighted average
638
High-end servers: US
High-end servers: World
Typical
Typical
Brand
Model
Power (W)
Notes
Brand
Model
Power (W)
Notes
IBM
p5 595
14,190
4, 6
IBM
p5 595
14,190
4, 6
HP
rp 8420
2,303
3, 6
HP
SUPERDOME
6,968
3
Sun
E25K
15,840
5, 7
Sun
E25K
15,840
5, 7
Weighted average
10,673
Weighted average
12,682
For questions, contact Jonathan Koomey, 510-708-1970 C, jgkoomey@stanford.edu
(1) For each server class, models shown are ranked in order of their market share in 2005.
Market shares are not shown because of confidentiality concerns.
(2) Dell 2850 typical power based on maximum measured power from Dell online configurator <http://www.dell.com/calc>
(3) HP volume servers assume G4 versions. Typical power based on max. measured power from HP online
configurator <http://h30099.www3.hp.com/configurator/calc/Power Calculator Catalog.xls>
HP rp8420 based on maximum measured power from spec sheets. Superdome typical power taken from spec sheets.
(4) IBM models' typical power based on max. measured power from spec sheets.
(5) Sun models' typical power based on maximum rated input power from spec sheets.
(6) Max. measured power multiplied by 40% and 66% to get typical power use for volume and mid-range/high end servers respectively.
(7) Max. rated input power multiplied by 25%, 30%, and 40% to get typical power for volume, mid-range, and high end servers, respectively.
Table 5: Percentage of shipments and installed base represented by the models used to estimate power levels
2000 installed base
2003 installed base
2005 shipments
US
Volume servers
20%
16%
23%
Mid-range servers
23%
26%
38%
High-end servers
28%
38%
39%
World
Volume servers
18%
16%
19%
Mid-range servers
20%
23%
28%
High-end servers
21%
31%
24%
For questions, contact Jonathan Koomey, 510-708-1970 C, jgkoomey@stanford.edu
(1) Table shows the percentage of total 2000/2003 installed base or 2005 shipments represented by the most
common models shown in Tables 2, 3, and 4, based on IDC data.
Table 6: Electricity used by servers for the U.S. and the World
Units
Volume
Mid-range
High-end
Total/Avg
Volume
Mid-range
High-end
Total/Avg
Total Installed Base (1)
2000
Thousands
4,927
663
23.0
5,613
12,240
1,808
66
14,114
2003
Thousands
7,578
530
21.4
8,130
18,523
1,540
62
20,125
2004
Thousands
8,658
432
23.3
9,113
23,441
1,238
66
24,746
2005
Thousands
9,897
387
22.2
10,306
25,959
1,264
59
27,282
Average power use per unit (2, 3)
2000
W/unit
186
424
5,534
236
183
423
4,874
236
2003
W/unit
207
524
6,428
244
214
522
5,815
255
2004
W/unit
213
574
6,973
248
216
578
6,783
252
2005
W/unit
219
625
7,651
250
222
607
8,106
257
Direct power consumption (4)
2000
Million kW
0.9
0.3
0.1
1.3
2.2
0.8
0.3
3.3
2003
Million kW
1.6
0.3
0.1
2.0
4.0
0.8
0.4
5.1
2004
Million kW
1.8
0.2
0.2
2.3
5.1
0.7
0.4
6.2
2005
Million kW
2.2
0.2
0.2
2.6
5.8
0.8
0.5
7.0
Total power consumption (7)
2000
Million kW
1.8
0.6
0.3
2.6
4.5
1.5
0.6
6.7
(including cooling and aux equipment)
2003
Million kW
3.1
0.6
0.3
4.0
7.9
1.6
0.7
10.2
2004
Million kW
3.7
0.5
0.3
4.5
10.1
1.4
0.9
12.5
2005
Million kW
4.3
0.5
0.3
5.2
11.5
1.5
1.0
14.0
Direct electricity consumption (5, 6)
2000
Billion kWh
8
2.5
1.1
12
20
7
3
29
2003
Billion kWh
14
2.4
1.2
17
35
7
3
45
2004
Billion kWh
16
2.2
1.4
20
45
6
4
55
2005
Billion kWh
19
2.1
1.5
23
50
7
4
61
Total electricity consumption (7)
2000
Billion kWh
16
4.9
2.2
23
39
13
6
58
(including cooling and aux equipment)
2003
Billion kWh
27
4.9
2.4
35
69
14
6
90
2004
Billion kWh
32
4.4
2.9
40
89
13
8
109
2005
Billion kWh
38
4.2
3.0
45
101
13
8
123
Total electricity bill (8)
2000
Billion 2006 $
0.9
0.3
0.1
1.3
2.1
0.7
0.3
3.1
(including cooling and aux equipment)
2003
Billion 2006 $
1.5
0.3
0.1
1.9
3.9
0.8
0.4
5.0
2004
Billion 2006 $
1.8
0.2
0.2
2.2
5.0
0.7
0.4
6.1
2005
Billion 2006 $
2.2
0.3
0.2
2.7
6.0
0.8
0.5
7.2
For questions, contact Jonathan Koomey, 510-708-1970 C, jgkoomey@stanford.edu
All servers: U.S.
All servers: World
Notes to Table 6
(1) Installed base is measured at the end of any given year, taken from Table 1, based on IDC data.
(2) Average power use in 2000 & 2003 for each server class taken from Tables 2 & 3, based on an installed base weighted average of power use for the top 6 servers in each class.
(3) Average power use in 2004 calculated assuming that retirements in 2004 (from Table 1) use the same amount of power per unit as does the year 2000 installed base,
and that new units in 2004 use the same amount of power per unit as the sales-weighted average of shipments in 2005 of the top three units for each server class (Table 4).
The procedure is repeated for 2005, assuming that retirements in that year also use the same amount of power as the 2000 installed base.
(4) Direct power use is the product of the total installed base and the average power per unit. It does not include utility transmission and distribution losses.
(5) Total electricity use converts total power in million kW to billion kWh assuming 8760 hours per year for 2003/2005 and 8784 kWh/year for 2000/2004 (leap years).
(6) Servers are assumed to operate 100% of the year
(7) Total electricity consumption (including cooling and auxiliary equipment) is twice that of the direct server power consumption, based on typical industry practice.
(8) Electricity bills for U.S. and World servers calculated assuming average U.S. industrial electricity prices from the Energy Information Administration,
adjusted to 2006 dollars using the GDP deflator http://www.eia.doe.gov/cneaf/electricity/epa/epat7p4.html