You are here

Power Distribution within Six PCs

August 30, 2005 by Devon
Cooke
and Mike Chin

How much electrical power does a computer system draw in real use?

This deceptively simple question is often critical in silent
computing. Why? Because power consumption and thermal dissipation are essentially
one and the same. Silent computing is largely the art of intelligent power management,
and thermal management with minimal airflow.

It can be broken down into several different questions:

  1. How much AC power does a computer and all its related peripherals (monitors,
    speakers, and printers, for example) draw?
  2. How much AC power does a computer draw apart from its peripherals?
  3. How much DC power does each individual component in the computer system
    draw?
  4. How much power does a computer system draw from each DC voltage rail on
    the power supply?

The first question is relevant when considering the electrical cost of a computer,
and is most likely to be asked by system integrators and IT departments with large numbers of systems where
even a small reduction in power per system may translate into a significant monetary
saving. It is useful to know how much each system as a whole costs to operate. However, this is not our focus.

The second question is of great interest to anyone interested in
assembling a quiet PC because it tells us the total heat produced
inside a computer case. Almost all power drawn by a computer ends up as
heat; AC power draw is the best measure of total heat
in the PC box. A system that draws 90W at full load is much easier to
keep cool (and do so quietly) than a system that draws 250W.



A Kill-a-Watt power meter used to measure AC power draw.

However, it's not just the total amount of heat that matters but where
it is being produced. In most systems there are two or three main sources of
heat: The CPU, the video card (if any) and the power supply. An effective means
of reducing heat is to replace any of these with a more efficient model, especially
if one of them is especially power hungry. It is helpful to know how much power
various components require, which is one reason for posing the third question
above.

Knowing the power requirements of each individual component is also useful
when "sizing" a power supply. Adding together the power required by each component
at full load can give a rough estimate of the maximum power that will be drawn
from the power supply. Just about every component inside a PC carries
some kind of rating for maximum power or current. But simply adding these numbers
together produces estimates for total system power demand that are always too
high, sometimes by as much as double. This is because no PC application stresses
all components simultaneously to maximum load.

This brings us to the last question, How much power does a computer system
draw from each DC voltage rail on the power supply?

We can answer this question and the question of how much power
is demanded by various components by measuring the actual current drawn on each
voltage line while the computer is in use. In fact, these are issues so often
discussed in the SPCR forums without any clear conclusions being reached that
we decided to measure the power distribution within half a dozen different PC
systems in an effort to shed a bit more light on the topic. The results should
bring some real empirical data to these discussions, and should help determine
what can — and what can't — be predicted about the power draw of a
system.

METHODOLOGY

The key to our testing is a Fluke 36 Clamp Meter. This is a device that
measures the electromagetic field around any wire that's carrying electricity
and translates it into a current readout in Amperes. If you've never seen one
before, it's almost magical how a clamp meter works. Playing with electricity,
one comes to assume that you have to expose bare wire and have good firm contact
before you can take any kind of measurement. Not so with the clamp meter. Typical
of such meters, the Fluke 36 is specified for about 1.9% accuracy, which is far less
precise than, say, voltage readings from a common digital multimeter. Our meter
has not been calibrated for over a year and is primarily an electrician's tool
for measuring AC current. It is sensitive to RF fields, which means its accuracy
may be adversely affected around lots of electronic gear. In short, it's not
exactly a precision lab tool. But, for characterizing general tendencies in
power distribution within a PC, our clamp meter is perfectly useful.



This Fluke 36 Clamp Meter was used to make current measurements.

Six different systems were available for testing in the lab:

P4 Socket 478 System

  • Intel Pentium 4 2.8 GHz (Northwood core)
  • AOpen AX4GE motherboard
  • 512 MB OCZ PC3200 RAM
  • ATI Radeon 9600XT AGP VGA card
  • 40 GB Seagate Barracuda IV HDD

P4 LGA775 System

  • Intel Pentium 670 (3.8 GHz)
  • Intel D915PBL motherboard
  • 512 MB Corsair DDR2 RAM
  • AOpen Aeolus 6800GT PCIe VGA card
  • 250 GB Western Digital Caviar SE HDD

Pentium D Dual Core System

  • Intel Pentium D 820 (2 x 2.8 GHz)
  • Intel D945GTP motherboard
  • 512 MB Corsair DDR2 RAM
  • 74 GB Western Digital Raptor HDD

AMD Socket A System

  • AMD Athlon 2500+ (Barton core)
  • MSI K7N2G motherboard (NForce 2)
  • 512 MB OCZ PC3500 RAM
  • 20 GB Seagate Barracuda IV HDD

AMD Socket 754 System

  • AMD Athlon 64 3200+ (Newcastle core)
  • Epox EP-8KDA3+ motherboard
  • 512 MB OCZ PC3500 RAM
  • Matrox MX440 AGP VGA card
  • 80 GB Samsung Spinpoint P80 HDD

AMD Socket 939 System

  • AMD Athlon 64 3500+ (Venice core)
  • DFI NF4 LanParty motherboard (NForce 4)
  • 2 x 512 MB OCZ PC4000 RAM
  • AOpen Aeolus 6800GT PCIe VGA card
  • 2 x 300 GB Maxtor DiamondMax 10 HDDs (in RAID)

The systems represent a decent cross-section of PCs in use today. None of them
can be directly compared to the others; what we are looking for is not the differences
between systems, but general tendancies of power usage that are true for almost
any configuration. A wide variety of CPUs and chipsets were tested, as well
as several different VGA cards. Each system was powered by its own PSU; there
were a variety of PSU models. All the PSUs were capable of delivering more than
15A on the 12V lines (combined).

For each system, the measurement procedure was the same: The individual wires

in the various cable sets from the power supply were separated and then recombined
according to the voltage they carried. In this way the total current from each
individual voltage line could be measured separately. The two +12V lines were
also separated and measured independently. The -12V and +5VSB lines were not
measured, as they carry so little current that they are insignificant (typically
well under 5W).

Each system was measured at idle and then under load using CPUBurn
(two instances were run for multi-threaded and dual core processors).

Ambient temperature at the time of testing was 26°C.

IDLE POWER TEST RESULTS

From a cooling perspective, the power draw at idle is largely irrelevant. A
good cooling system must provide adequate cooling when the system is under sustained
high load. Any system that can handle this will automatically be cool enough
at idle. The same thing applies when sizing a power supply: If it can handle
the peaks under heavy load, it should have no problems supplying the
power required by a system at idle.

From the standpoint of conserving power, however, the idle power draw is very
important because most systems spend a vast amount of time at or close to idle.
The total power draw at idle is largely determined by the load on the +12V lines.
In fact, with only one exception, the relative power draw on the +12V lines
was a good predictor of the total power draw.

POWER DISTRIBUTION WITHIN THE PC: IDLE
System

+12V (total)

+12V1

+12V2 (CPU)

+5V

+3.3V

Total DC Power
AMD Socket 754 Athlon 64 3200+ (Newcastle)

0.8A

0.4A

0.4A

2.5A

2.6A

31W
Intel Socket 478 P4 2.8 GHz (Northwood)

1.8A

0.9A

0.9A

1.1A

3.3A

38W
AMD Socket A Athlon 2500+ (Barton)

4.6A

2.0A

2.6A


2.3A

2.0A

73W
Intel LGA775 Pentium D 820 (2 x 2.8 GHz)

4.6A

0.5A

4.1A

3.6A

0.7A

76W
Intel LGA775 Pentium 670 (3.8 GHz)

5.1A

2.3A

2.8A

2.9A

1.5A

81W
AMD Socket 939 Athlon 64 3500+ (Venice)

4.6A

4.0A

0.6A

3.0A

3.9A

83W

Listed in order of increasing total DC Power

Note that the systems were listed in the table above in order of
increasing total power. That the A64-3200+ system should come in with
the lowest power draw at idle was no surprise, given what we know about
the relatively high efficiency of the A64. The modest idle draw of the
P4-2.8 is a bit of a surprise, as is the high power draw of the Athlon
2500+. The latter certainly dates it as a pre-Cool 'n' Quiet AMD
processor, but there may have been other factors, as discussed below.

The A64-3500+ system was arguably the most powerful system tested; it was the
only system with two sticks of RAM and the only one with two hard drives. This
is reflected in the relatively high current draw from both the +3.3V line (for
RAM) and the +5V line (for the HDD). In fact, the combined power draw on these
two lines totaled almost 30W, about 50% more than the other systems,
which drew 18~21W from these lines at idle. This is enough to make this system
the most power hungry at idle by a small margin.

So, why does an ostensibly power-efficient AMD-based system draw
the most idle power of any system tested? The extra RAM and hard drive obviously
contribute a little, but not enough to explain the 50W gap between this system
and the Socket 754 system. Most of the power is being drawn on the +12V1 line,
so it makes sense to figure out what is being drawn from this line. The
likely suspect is the high-powered 6800GT VGA card, which consumes a significant
amount of power.

However, this does not explain why the Pentium 670 system — with the same VGA card — draws so much less current from the +12V1 line.
We can only speculate. Perhaps the power regulation circuitry on the LGA775
motherboard is more efficient, or maybe the nForce 4 chipset for the AMD system
is especially power hungry. Another possibility is that two motherboards may
divide up the power from the different voltage lines in slightly different ways.
Ultimately, it doesn't really matter which component required the additional
power. As far as the power supply is concerned, all that matters is the distribution
across the various voltage lines.

The vanilla Athlon system also draws a substantial amount of power from the
+12V1 line. This is unlikely to be needed by the CPU, which has its own voltage
line, and it does not have a video card. So, the power seems to be required
for the motherboard itself. The only other load on the +12V1 line is
the hard drive, which accounts for just 0.3A. But, why should the motherboard
require so much power? It's impossible to say for sure, but is it possible that
the onboard video, based on the GeForce 4 MX, is the culprit?

Despite all of these minor differences, the general trend was quite clear:
The +5V and +3.3V lines draw very little power. Furthermore, the total power required by these lines
does not vary much between systems. Building a system that draws less power
at idle seems to rely mainly on keeping the power draw on the +12V line as low
as possible.

CPU LOAD POWER TEST RESULTS

POWER DISTRIBUTION WITHIN THE PC: CPU LOAD (CPUBurn)
System
+12V (total)
+12V1
+12V2 (CPU)
+5V
+3.3V
Total DC Power
AMD Socket A Athlon 2500+ (Barton)
5.4A
2.3A
3.2A
2.3A
1.8A
82W
AMD Socket 754 Athlon 64 3200+ (Newcastle)
5.2A
1.8A
3.4A
2.5A
2.6A
83W
Intel Socket 478 P4 2.8 GHz (Northwood)
6.5A
0.9A
5.9A
1.2A
3.7A
96W
AMD Socket 939 Athlon 64 3500+ (Venice)
7.8A
3.9A
4.0A
3.8A
3.8A
125W
Intel LGA775 Pentium D 820 (2 x 2.8 GHz)
11.4A
0.5A
11.0A
3.5A
0.6A
155W
Intel LGA775 Pentium 670 (3.8 GHz)
13.7A
2.2A
10.7A
2.8A
1.6A
183W

Listed in order of increasing total DC Power

In every system, the CPU load resulted in increased current draw on the +12V2
line. For the most part, none of the other lines were affected by the CPU load.
Once again, exceptions were seen in the AMD based systems.

As expected, the Socket 754 system saw an increase of 3.0A on the +12V2 line.
However, the current on the +12V1 line also rose by 1.4A; this increase was
not seen on any other system. This suggests that the CPU in this system was
drawing current from both +12V lines.

Likewise, the Socket 939 system also drew primarily from the +12V2 line, but
power draw also increased by 0.8A on the +5V line. In the grand scheme of things,
this increase is barely significant, but it does show that a given load on one
system may not produce the same load pattern on a different system. We can only
guess as to why the nForce4 board requires power from the +5V line when other
boards do not.

Note how the total system power ranking order changed. There's no question
here that under high CPU load, the Intel systems are the power hogs. The lower
end AMD systems didn't even get close. Only the AMD A64-3200+ system exceeded
the power draw of any Intel system, in this case, the much less capable P4-2.8.

Power demand on the +5V and +3.3V lines varied widely between different motherboards
and chipsets. Only the combined power draw of the two lines stayed fairly consistent.
Regardless of platform, whether at idle or at full CPU load, the combined power
draw for these two lines almost always stayed in the 20-30W range. Once again,
the amount of power drawn on the +12V line almost invariably determines the
total system power draw.

POWER DISTRIBUTION BY PERCENTAGE: CPU LOAD (CPUBurn)
System
+12V (total)
+12V1
+12V2 (CPU)
+5V
+3.3V
Intel Socket 478 P4 2.8 GHz (Northwood)
81%
11%
73%
6%
13%
Intel LGA775 Pentium 670 (3.8 GHz)
90%
14%
70%
8%
3%
Intel LGA775 Pentium D 820 (2 x 2.8 GHz)
88%
4%
85%
11%
1%
AMD Socket A Athlon 2500+ (Barton)
79%
34%
47%
14%
7%
AMD Socket 754 Athlon 64 3200+ (Newcastle)
75%
26%
49%
15%
10%
AMD Socket 939 Athlon 64 3500+ (Venice)
75%
37%
38%
15%
10%

Based on power (watts) drawn from each line.

It is instructive to look at the load on each individual rail as a percentage
of the total load. The most evident trend is that the vast majority of the power
is drawn from +12V2 line. This is not surprising. After all, CPUBurn stresses
the CPU almost exclusively, so it makes sense that the CPU voltage line dominates
power needs.

It is also interesting to note the differences between the AMD- and Intel-based
systems. The AMD systems drew about 75% of their power from the +12V
lines. This is a lot of power, but not as much as the Intel systems, which drew
as much as 90% of their total power from these lines.
This difference reflects the higher power demand of Intel CPUs, especially
the newer ones. The older Northwood-core P4-2.8 drew
much less power than Intel's newer CPUs, and consumed proportionally
less power. It behaved more like the AMD CPUs than the newer Intel processors.

POWER DRAW OF INDIVIDUAL COMPONENTS

In addition to looking at differences in power consumption between idle and
load, we also looked at how idle power consumption changed when individual components
were added or removed from a system. This allowed us to judge roughly
the power overhead for various components. Several different parts were
tested: Two different kinds of RAM, two different video cards, a PCI Ethernet
card, and an optical drive. Hard drives were not tested, as we
already have a reliable method of determining their power consumption
.

For each component, the power consumption on each voltage line was measured
with the system at idle. These measurements were then compared to the relevant
idle measurements without the component installed. The results below represent
the net change between these two measurements. Differences of 0.2A or less were
assumed to be within the margin of error for our testing equipment, and are
therefore not included in the results below.

RAM

Two types of RAM were tested: Regular PC3200 DDR and 533 MHz DDR2. In each
case, a single 512 MB stick of RAM was added to an existing system configuration.
The P4 Socket 478 system was used to test the PC3200, and the Pentium D Dual
Core system was used to test the DDR2.

To stress the RAM modules while we measured them, we used Memtest86
to drive the power consumption up. Because Memtest86 runs under its own OS, not Windows, a new idle baseline was needed for accurate comparison.
Measurements for each type of RAM, at idle and under load, are given below.
The total power draw of the system is only given for the 512 MB configuration
at idle. All other configurations are given as differences in power relative
to this baseline. Note that the power consumption for the +12V2 line was not
listed; it was assumed that the CPU was the only load on the +12V2 line and
therefore not used by the RAM.

POWER CONSUMPTION: RAM
RAM Type
Size
Load
+12V1
+5V
+3.3V
Rise from Baseline
PC3200 DDR
512 MB
Idle
0.5A
0.6A
3.0A
n/a
Memtest86
No Change
No Change
+0.7A
+2.3W
1 GB
Idle
No Change
No Change
+0.6A
+2.0W
Memtest86
No Change
No Change
+1.0A
+3.3W
533 MHz DDR2
512 MB
Idle
0.5A
3.6A
0.5A
n/a
Memtest86
No Change
+0.4A
No Change
+2.0W
1 GB
Idle
No Change
No Change
No Change
No Change
Memtest86
No Change
+0.9A
No Change
+4.5W

The power consumption of a single 512MB stick of SDRAM, either regular
DDR or DDR2, was too small for the coarse resolution of our measuring
equipment. The largest measured increase was only marginally larger than the
potential error in the system: 0.2A * (12V + 5V + 3.3V) = 3.9W. Let me state
this in another way: It is statistically possible that the total power consumption
went down for almost every test case, which is a result that should
be impossible given proper test conditions.

That said, the very fact that our potential for error is so high implies
that the power required by RAM is basically insignificant. Even in the worst
possible case, DDR2 under load, the total increase over the baseline measurement
remains under 10W.

A tentative estimate of the power draw for a single 512 MB stick of SDRAM can
be gleaned by comparing the rise in load when a stick of RAM is added. For
ordinary DDR, this change was a paltry single watt, and the higher speed DDR2 was limited
to 2.5W. When idle power requirements were compared, the situation was reversed:
Ordinary DDR increased power draw by two watts, while DDR2 did not appear
to draw any power. As mentioned above, these numbers are too rough to draw
any firm conclusions, but the average power consumption appears to be in the
1-3W range per 512MB stick.

PCI Ethernet Card

The P4-2.8 GHz system was used as a baseline for comparison. There were no
identifying marks on the card itself, but the controller chip was marked "RTL8139B",
which identifies RealTek as the OEM.

POWER CONSUMPTION: PCI ETHERNET CARD
Load
+12V (total)
+5V
+3.3V
Net Change in Power Draw
Idle
No Change
No Change
+0.5A
1.6W
Network Data Transfer
See Text
See Text
+0.6-0.8A
2-2.6W

Like the RAM test, only a small change in the load on the +3.3V line was
observed after the Ethernet card was installed. The net increase was ~2W:
Barely significant.

An attempt was made to measure the power consumption during a sustained data
transfer. This produced a small increase on all voltage lines, but it is unlikely
that the increases on the +12V and +5V lines could be attributed to the PCI
card itself. Most likely this power was needed by the hard drive to copy the
data used during the transfer. Hard drives do not consume power from the +3.3V
line, so it seems safe to attribute this increase to the add-in card.

Optical Drive

The baseline system is completely irrelevant to the measurements of the optical
drive (Creative Labs 52x CD-ROM). Because the optical drive is powered by
an individual IDE power connector, the current through these wires could be
measured directly.

POWER CONSUMPTION: OPTICAL DRIVE
Load
+12V
+5V
Total Power Draw
Idle

0.0A

0.3A
1.5W
Typical Read
0.3A
0.4A
5.6W
Full Speed
1.1
0.5A
15.7W

At lower speeds, the power draw for the optical drive was fairly small, but
at full speed and spin-up the sustained power draw was about 15W, mostly from
the +12V line. This is enough to be worth considering when sizing a power
supply.

Video Cards

Three different video cards were also tested: A Matrox G550 and an
ATI Radeon 9600XT. AGP cards were tested on the P4-2.8 socket 478 system,
and the AOpen Aeolus 6800GT PCIe card was tested on the Intel Pentium
D dual core system. Only differences at idle were examined, as trying to gauge
differences at load proved to be near impossible because too other components
are brought into play (particularly the CPU).

POWER CONSUMPTION: VIDEO CARDS
Video Card
+12V (total)
+5V
+3.3V
Change in Power
Matrox G550
No Change
+0.3A
+1.7A
+7.1W
ATI Radeon 9600XT
+0.3A
+0.4A
+0.3A
+6.6W
Aeolus 6800GT
+3.0A
No Change
No Change
+36.0W

The Matrox and the ATI both drew less than 10W at idle. The Matrox card
drew most of its power from the +3.3V line, while the Radeon seemed to
draw power from all three main voltages.

The Aeolus 6800GT, on the other hand, drew much more power at idle —
about five times as much. All of the power came from the +12V
line; neither of the other two lines were affected. To put it in perspective,
the entire A64-3200+ socket 754 system used less power at idle than this video
card.

Although the 6800GT uses a PCIe connector to draw power directly from the
power supply, most of its power seemed to be coming through the PCIe slot
on the motherboard. Of the total 3.0A load on the +12V line, 2.1A was drawn
through the motherboard, and 0.9A came from the direct connection to the power supply. Note
that this is only at idle; it may change during high load.

From the measurements presented here, it is hard to generalize about
the power consumption of video cards. The power requirements for each card
are unique; there do not seem to be any similarities between the models we examined. That said,
it is well known that the most powerful (and recent) cards draw
from the +12V line. This can be seen by examining the external PCIe connector:
It has only +12V and ground wires. It is important to remember that, while
this generalization might be true for powerful cards, the more mainstream
cards do not necessarily echo this power profile.

THE HEAVIEST LOAD

As a matter of curiosity, a system of the most power hungry components
in our lab was put to the highest load we could devise. The AOpen Aeolus 6800GT
was installed in the Pentium D Dual Core system, and both the CPU and the
GPU were stressed by running two instances of CPUBurn simultaneously with
3DMark05. Peak power consumption was noted on each voltage line.

In addition to the usual measurements, the +12V wires in the ATX motherboard
connector were measured separately from the rest of the +12V wires, as were
the PCIe wires. This allowed us to measure the power drawn by the video card
from the power drawn by the CPU. The +12V ATX wires were also measured without
the video card installed so that a baseline could be established; no current
seemed to be drawn from these wires unless the PCIe slot was in use.

POWER CONSUMPTION: HEAVY LOAD TEST
Load
+12V (total)
+12V1
+12V2
+12V (ATX)
+12V (PCIe)
+5V
+3.3V
Total Power
2x CPUBurn +

3DMark05
16.9A
6.4A
10.4A
3.4A
2.5A
3.6A
0.7A
223W
91%
34%
56%
18%
13%
8%
1%

Under this extreme load, the total power draw peaked at about 220W. Neither
the +5V nor the +3.3V lines drew appreciably more power during the test. About
90% of the power was drawn though the +12V line.

Although the overall power draw was much higher than the original CPUBurn
test, it is interesting to note that the CPU seemed to consume slightly less
power when 3DMark05 was added to the mix. Since CPUBurn should have been using
all idle processor cycles, it is possible that the CPU operations demanded
by 3DMark05 do not require as much power as the ones used by CPUBurn.

The total power draw of the Aeolus 6800GT at load can be estimated at about
six amperes on the +12V line, or around 70W. At first glance, this seems considerably
higher than the
55W measured for the 6800GT by X-bit Labs
. However, our loads and measuring tools
are different from the ones used by X-bit Labs. In addition, our measurement
also includes efficiency losses that may occur in the power circuitry
in the motherboard.

CONCLUSIONS

It is important to keep in mind that the measurements presented here are
continuous loads. Our test equipment does not have the resolution
to measure peaks, which may last for 10 ms or less and may be much higher than
the continuous load. Most power supplies are rated for a continuous
load with allowances for higher peaks, but the internal protection circuits
may still be tripped by an exceptionally high peak. It is wise to leave perhaps 30% headroom
for peaks when sizing a power supply.

With these caveats, some broad, predictable conclusions can be drawn:

1) It seems to
be highly unlikely that a modern system will ever overload either the +5V or
+3.3V lines of a ATX12V 2.x compliant power supply.
In our systems, neither of these lines ever drew more
than 5A under any circumstance, and many power supplies rate them above
20A. The power draw on these lines tended to be quite stable and did not fluctuate
much with load.

2) The +12V lines, on the other hand, are very heavily used, especially under load. For Intel-based systems with no external
VGA card, this power comes almost exclusively from the +12V2 line. Adding a
high powered VGA card may add some load to the +12V1 line, although not all
cards use +12V. The systems with AMD CPUs tended to draw power more evenly across
the two +12V lines, mainly because they do not consume as much power as Intel
CPUs.

3) Looking at the total power draw alone, it would appear that all of our systems
could easily be handled by a 300W power supply.
Given that as much as 90%
of that power comes from the +12V lines, it is likely that that the ratings
for the +12V lines matters more than the total wattage. If these lines are inadequate,
the power supply may not provide enough power even if its "wattage rating"
exceeds the total power draw of the system. It would be wiser to qualify our
statement thus: When it comes to adequate power delivery, all of our test systems could easily be handled by a 300W power supply
that conforms to ATX12V v2.xx.
Conversely, an older PSU rated honestly for 300W output may not be adequate for the most powerful system examined here because of the much lower 12V current capacity on models that comply with v1.3 and earlier versions of the ATX12V spec.

NOTE: None of the above conclusions are meant to suggest that power delivery alone are the only criteria by which a PSU should be chosen. We are only considering adequate power delivery. We have not touched on noise, efficiency, cooling, voltage regulation — in short, all of the other relevant criteria we examine in our PSU reviews.

* * *

Discuss
this article in the SPCR Forums.

Google

www SPCR