On First of April, 2009, Google's decided to unveil information concerning its servers,
including its uncommon technical approach to UPS design.
Based on the released reports, Google's server is a 2U thick Rackmount unit, comprising a
Power Supply/UPS fed from 208V/60Hz or 230V/50Hz, A motherboard and an internal 12V
The Power Supply works directly from the utility. It provides 12V nominal DC Buss voltage
with 92% efficiency, for powering the motherboard and charging the battery.
The Motherboard features 8 DIMM slots, 2 Intel or AMD x86 Processors and two Hard Drives.
The battery is a 12 V sealed lead acid VRLA type, providing ride through power for about 1
minute, until generator kicks in, in case utility is not restored.
The 12 V DC buss feeds directly the motherboard. Dedicated converters on the motherboard
supply 5V DC to the Discs and 3.3V DC (or 1.2 -1.8V) to the processors.
Google says that cost was the main factor dictating the design approach. No external UPS,
means lower investment in hardware and no running UPS losses.
The design seems at first glance trivial, even too simple to be taken seriously. Some considered
it to be Google's 1st of April fools joke. It indicates Google's decisiveness to dare and
implement uncommon solutions to achieve its main objective.
Let's, consider Google's approach:
Google uses practically a DC UPS, which serves also as the server's power supply.
It is apparent that elimination of external UPS eliminates also its cost, as well as the associated
power loss. An external On-Line UPS, if added, even if one considers it to have the best
available efficiency of 95-96%, would cause total efficiency, from utility to the DC Buss to
drop, from 92% to 88%.
A standby (Interactive) UPS with 98% efficiency, would decrease total efficiency to 90%.
Google uses a single 12V DC rail.
All 12V loads are directly supplied from the DC buss. Efficiency is further improved with the
necessary conversion to lower voltage consumers taking place at the motherboard. The
single12V buss approach eliminates 83% to 95% copper losses associated with transmitting
power to feed the low voltage 5V or 3.3 V circuits.
Google achieves 92% power supply efficiency. This is a remarkable figure, particularly taking
in account that the design is already 5 years old.
Each server has an internal 12V short time backup battery.
Google uses 1 minute battery. Smaller battery means lower cost. One minute backup time
provides sufficient time to cover 90- 95% of utility dips and short time outages, as well as the
override time needed for starting up a standby generator, which generally kicks in within 10 to
Locating the battery within the server, eliminates the need to add DC filters, as well as the
added cost of routed wires and work, required by externally located batteries.
No external UPS , nor redundancy is used in Google's design.
Compromising on reliability and increased maintenance seem to be the costs that Google has to
pay for pursuing "low cost" oriented approach. Server redundancy is the only way to supply
service in case of server or battery malfunction.
True, external UPS can not help when the server or its internal power supply fail. And servers
do fail. Generally however, an external UPS isolates the server from utility disturbances. It
provides an ideal environment to keep the server running, decreasing dramatically the
probability of server's hardware or data failures.
UPS redundancy is needed to eliminate exposing the server to unconditioned utility, in case of
UPS failure. Redundant batteries are used to avoid the possibility of server's failure, during
mains outage, because of a failed battery.
Google does not provide battery redundancy. The worst place to locate a battery is in the
Server. Google locates the batteries outside the mother board, relatively far from the Power
Supply and the processors which form the hottest point. But, battery life degrades by about
50% per each 10 degrees temperature rise. Thus the batteries, and particularly the ones located
in servers placed at to the top of the rack, where the heat piles up may require frequent
All in all, the total cost saving should take in consideration additional factors, such as the added
cost of maintenance and service due to server's and battery failures and wear, the added costs
of server redundancy, as well as rack space utilization, and other factors.
Lessons from Google
Google's approach bridges above the gap of two separate industries, the Computer industry,
and the Power protection industry, each one seeking the best performance at the lowest
possible cost. Recent technical developments, and legislation, result in using similar power
solutions in the UPS and the Server. Thus, duplicating total power supply hardware.
An On-Line double conversion UPS, converts the incoming AC voltage by means of a Power
Factor conversion circuit from AC to High Voltage DC, which is then converted from DC to
AC by means of an Inverter.
Similar circuits exist in computer power supply. Here, an additional AC to DC stage is added to
generate the DC voltages required by the server and its components. This additional stage
decreases available server's Power Supply efficiency today from 96% to 92%.
Introducing Blade Servers was a giant step towards increasing compactness and efficiency in
data centers. This is still a young technology, prime for evaluation which considers the entire
picture, examining end to end solutions from utility to the processor.
Breaking the mental gap between Computer and the UPS industries, may be the best lesson
from Google's philosophy. UPS and Computer manufacturers should perhaps re-examine the
current technology, with a more holistic attitude, considering technical issues, as well as
business venture approaches, aimed to eliminate unnecessary hardware duplication, in order to
evaluate simpler, more customer oriented solutions.
Source: April 2009 UPSonNet Newsletter
|Learning from Google's Server UPS Design
|Spread Your Word On the Globe
UPSonNet Most Viewed Articles