• Home
  • blog
  • Desktop CPU Power Survey, April 2006

Desktop CPU Power Survey, April 2006

blog image

It began as a follow-up to Turion 64 on the Desktop article, posted in mid-February. But then, the piece ballooned into something more ambitious, into an attempt to answer the question,

“What is the best power efficiency achievable with current AMD and Intel processors that can be used on a desktop PC?”

In the process, we ended up examining more than 15 processors in half a dozen family groups on six different test platforms, and took a stream of power measurements that kept our heads spinning for a couple of weeks. We have some answers. They comprise a snapshot of the fast-changing processor scene, taken from a particular angle, at this point in time.

March 6, 2006 by Mike Chin

This article began as a relatively simple follow-up to Turion 64 on the Desktop, posted in mid-February. It was meant to answer the question most often asked by readers of the Turion 64 article,

"So how does the Turion 64 fare against the latest Athlon 64s, in terms of power efficiency, performance, ease of use on the desktop and price?"

It’s a pertinent question for many reasons:

  1. The same basic core is at the heart of the Turion 64 and the Athlon 64.
  2. The power efficiency of the Athlon 64 has been improved by AMD, quietly but steadily. The "E" steppings and beyond are particularly cool-running, and may actually approach Turion 64 levels.
  3. If an Athlon 64-939 is almost as cool as a Turion 64, it is a better choice to run on a desktop system for most people, due to ease of implementation, lower cost for similar clock/cache, and the superior performance of socket 939 motherboards — due mostly to dual channel memory support but also because the most ambitious motherboards are socket 939.

Somehow, during the research to compose a reasonably complete answer to the above question, the piece ballooned into something more ambitious. We ended up examining almost every recent CPU in the SPCR lab on several different platforms to answer not only the question above, but also the question,

"What is the best power efficiency achievable with currently available AMD and Intel processors that can be used on a desktop PC?"

It is not possible to answer this question comprehensively with precision without access to dozens of processors and test beds, and an enormous amount of time, manpower and resources. Still, we managed to get a good sampling of the processor families that represent the vast majority of CPUs actually being bought and used for desktop PCs today. We used the test methods that worked well for the Turion 64 article.


Some of the gear assembled for this article.

WHY ASK THE QUESTION?

In the above question, the phrase "best power efficiency" means lowest power consumption, which relates directly to the amount of power demanded and generated into heat by the CPU, and to a lesser degree, the motherboard. (More on the latter later.) The relevance of this question is obvious for those who seek to minimize PC noise: The cooler the PC, the less airflow required to cool it, which means fewer and/or slow spinning fans that make less noise.

We won’t go into PC silencing in any detail here. Suffice it to say that reducing heat
means cutting down on power consumption. A number of methods for doing so have
become popular at SPCR, including:

  • Undervolting and underclocking: Lower voltage chips running at slower clock speed consume less power.
  • Deliberately choosing low-end parts, as they are often slower and
    lower voltage than the latest and greatest.
  • Choosing more efficient parts, with high performance-per-watt ratios.
  • Using parts designed for the laptops (such as HDDs and CPUs), which consume minimal
    power to conserve battery life.

The power and heat issue is relevant to everyone else in a variety of ways:

  • Environmental: High power efficiency means lower energy consumption becomes a significant environmental issue. Computers do represent a significant percentage of electricity consumption not only directly but indirectly with increased air-conditioning cost in enterprise applications due to their added heat.
  • Performance: High power efficiency means that the electrical and thermal stress in high performance computers can be reduced, and make even higher performance achievable. When the high expense and complexity of delivering high power into a PC and then evacuation the ensuing heat is considered, a cooler, more efficient CPU makes perfectly good sense for dedicated gamers and others who need or seek extreme performance. Achieving extreme clock speeds requires the best cooling; less heat is a good starting point.
  • Financial: Both of the above points touch upon the significant expense, both short and long term, directly associated with low power efficiency. High power efficiency means lower costs all around. The costs of other components drop when the CPU consumes less power, from the power supply to the heatsink/fan to the enclosure cooling requirements. Other cost factors external to the PC — including Uninterruptible Power Supplies, air-conditioning, and direct electricity costs — all drop when less power and heat are in the PC. The savings become very serious when large numbers of computers are considered, as in big enterprises like government agencies and corporations.

As we noted in the Turion 64 article, hardware manufacturers have recognized the importance of reducing power. High efficiency power supplies are increasingly more common, and Intel and AMD both identify performance-per-watt as a key benchmark.

TESTED PROCESSORS

The table below lists all the processors we tested, ordered and roughly matched by price. All the prices those officially cited for large corporate customers, in US$, as listed here: AMD Processor Pricing and Intel Processor Pricing. This may not be the best way to compare, as prices change so quickly, but given the huge range of processing performance represented here, it is probably about as good as any. There is a rough performance match. The A64-3000+ and P4-630 are relatively close, as are the A64 X2 3800+ and Pentium D 930. The A64 X2 4800+ and Pentium D 950 might also be a reasonable performance contest as well. The Core Duo T2600 probably won’t stand up to the desktop top dogs in performance, but the price matches. In any case, the table is simply a way to list all 15 processors tested in one simple table. (There are only 14 listed; the missing one is a second A64-3500+ with a different core stepping.)



AMD



PRICE



INTEL



Sempron 3400+

754, 2GHz / 256KB, 64-bit, Venice

$120



Athlon 64 3000+

939, 1.8GHz / 512KB, 64-bit, Venice

$140

$178

Pentium 4 630

775, 3.0GHz / 2MB, 64-bit, Prescott

Athlon 64 3500+

939, 2.2GHz / 512KB, 64-bit, Venice & Winchester


$205

$210

Pentium D 820

775, 2.8GHz / 2MB, 64-bit, Smithfield


Turion 64 MT40

754, 2.2GHz / 1MB, 64-bit, Lancaster

$268




Athlon 64 X2 3800+ Processor
939, 2.0GHz / 2x512KB, 64-bit, Toledo

$301

$316

Pentium D 930
775,
3.0GHz / 2x2MB, 64-bit, Presler 65nm

Athlon 64 4000+ Processor
939, 2.4GHz / 1MB, 64-bit, San Diego

$341
  


$424

Pentium M 770
479, 2.13GHz / 2MB, Dothan
  

$605

Pentium 4 670

775, 3.8GHz / 2MB, 64-bit, Prescott

Athlon 64 X2 4800+ Processor
939, 2GHz / 2x1MB, 64-bit, Toledo

$643

$637

Core Duo T2600

479, 2.16GHz / 2MB, Yonah


$637


Pentium D 950
775,
3.4GHz / 2x2MB, 64-bit, Presler 65nm

An obvious but important fact to point out is that an articles like this is a snapshot of moving targets. It’s a survey that’s relevant today. Within six months, especially in this year of major processor line changes by both Intel and AMD, the whole processor scene may be changed, and we will probably be scrambling to put together another survey relevant for that time.

CPU SAMPLE VARIANCE

The power consumption characteristics of processors varies from sample to sample within the same model and stepping. It can vary as much as >10%, according to some sources, although it is probably considerably less than that on average. It is exactly the same variance that makes some individual processor samples easily overclocked (or undervolted) and others not overclockable at all. We have a mix of samples, mostly provided directly by Intel and AMD, mostly only one sample of each model. Our results are probably more or less repeatable but it would be a surprise if anyone obtained identical results with a similar collection of samples. Our results are good general indicators, but please don’t assume that because our sample managed to run stable at 1.15V on our particular motherboard that all samples of the same CPU model can do the same. Also, don’t base your buying decision between two closely ranked processors on the basis of any single result we report. Price, suitability, availability, peripherals, ease of implementation — these are all important buying considerations.

In recognition of these sample variances, which have always existed, AMD began embedding sample-specific information about Thermal Design Power (TDP) and TCaseMax (maximum casing temperature) in E-Revision A64 processors some time last year. These are the chips that contain this information include:

* Athlon 64 (Venice, San Diego)
* Athlon 64 FX (San Diego)
* Athlon 64 X2 (Manchester, Toledo)
* All Rev E Opterons and Dual Core Opterons

TDP and TCaseMax are closely related to each other, moving higher and lower together. A processor cannot have low TDP and high TCaseMax or vice versa. For each model, there is a broad enough range of TDP/TCaseMax so that some people are very interested in cherry picking. Unfortunately, this information is not visible on any marking on the outside of the processor or the packaging. The processor has to be plugged into a motherboard and run before the information can be accessed. There is a little utility written just for this purpose: AMD64 TCaseMax v1.18. Here are some screenshots from this program.

The processor on the left is a Winchester core (D0 revision) and lacks the sample-specific TDP info. The Venice core (E3 revision) on the right has it: 67W TDP, which is the stock spec for this processor. It suggests this sample is run-of-the-mill for overclocking or undervolting.

The A64 X2 3800+ on the left is a more recent E6 revision (also called stepping) with a TDP of 65.6W, which is considerably lower than the standard 89W spec for this part. The same is true of the X2 4800+ on the right; 85W is much lower than the standard 110W TDP spec for this model.
The 44.1W TDP 3000+ E3 stepping on the left has the lowest TDP of any A64 in this roundup. The standard spec is 67W. The 4000+ E4 on the right is also a cool operator, coming in 39W below its standard 89W spec. Both of these samples should be good undervolters.

AMD says that the average TDP of most of their processor models has been dropping steadily over time. This is the result of a conscious effort, of continuous refinement in the manufacturing process. It means that while you might still find 89W 4000+ samples, the number of such parts coming off the line is much lower today than when the part was first introduced. The 50W samples, on the other hand, may not even have existed when the model was first introduced.

Intel does not embed individualized TDP information with its current line of Prescott, Smithfield, and Presler core processors. Variances certainly exist among Intel processors, but we have never had many of the same model to get any sense of how much they vary in thermally. The Prescott P4 is infamously hot, and while there have been improvements to help manage the heat better since their initial release, increasing clock speeds and memory cache sizes have not helped much. Enhanced Intel Speedstep (EIST) dynamic clock/voltage adjustments similar to AMD’s Cool’n’Quiet, has helped, along with C1E halt state support. Both help reduce idle power but do nothing for power draw at load. While idle power consumption is more important for overall energy efficiency of the PC, it’s the maximum CPU power peaks that the cooling system must be capable of handling. The tendency toward very high maximum power demand of the Intel desktop processor line has not changed, with the exception of the mobile offshoot Core Duo / Solo processors. We will have to wait for Conroe later in the year for an Intel desktop processor with TDPs boasted by AMD processors today.

TEST PLATFORMS

Six test platforms were used. One each for socket 754 AMD processors, for socket 939 AMD processors, for Pentium
M, for Intel Core Solo / Duo and two for Intel socket 775. Each motherboard was flashed to the latest BIOS. Each system was run with the same amount of RAM, and a notebook drive, which all have idle power consumption of ~1W or less, in order to keep the system power as consistent as possible. The same power supply unit was used for each and every platform to eliminate any differences in power
supply efficiency. This entailed a lot of plugging and unplugging, as you might guess.

We were able to use the onboard graphics card for all but one of the motherboard. Onboard graphics was one of the criteria used to choose the motherboards for this project. The main reason was to minimize total system power consumption and to reduce differences in peripheral power differences among the systems. (The one exception was the Core Solo / Duo motherboard, which is a pre-production test lab sample. When the testing was being done, there were no retail Core Solo / Duo boards available anywhere.)


CPU Socket

Motherboard

RAM

HDD

754

MSI RS482M-IL

2 x 512 MB OCZ Gold PC3200 DDR SDRAM


Seagate Momentus 5400.3 ST9160821A


939


ASUS A8N-VM CSM


Seagate Momentus 7200.1 ST910021AS



479-1
(Pentium M)

AOpen i915Ga-HFS

2 x 512 MB Corsair DDR2 SDRAM



Hitachi E7K100


479-2
(Core Solo/Duo)

AOpen 975Xa-YDG*

Samsung MP0402H



775-1


AOpen i945Ga-PHS

Seagate Momentus 5400.2 ST9120821AS


775-2*


Intel D945GTP

*NOTES:

A. Each and every platform is a minimalist system. There are no extras of any kind. This is realistic for the low end or low power processors, but not so realistic for the higher power / performance ones. Few users would buy a high performance processor and only use onboard graphics or a 5400 rpm notebook drive. We know that. The objective of this article required the most minimalist setups, however, and consistency in the test platform configurations.

B. The AOpen 975Xa-YDG Lab Sample preproduction board is not equipped with onboard graphics. There were no other options for Core Solo / Duo boards at time of writing, certainly none on the retail market. An outboard PCIe graphics card was used, the AOpen Aeolus PCX6600-DV128LP. Based on X-bit Labs’ recent roundup of graphics card power, we’d estimate that this card has a minimum (idle) power draw of 10~14W and a maximum of 25~30W. For our testing, only the idle power matters.

A quick glance at this new AOpen board reveals its high end aspirations clearly. Rest assured it will have a hefty price tag. (NOTE: This board was marked "Lab Sample" with a BIOS that was not quite ready for prime time; a review will have to wait for a production sample.)


The AOpen 975Xa-YDG board for Core Solo/Duo sports many features, including Dual PCIe x16 video card slots, a full compliment of I/O for 7.1 ch audio, S/PDIF in and out, gigabit LAN, and finally, a socket 478 heatsink mounting bracket that allows the use of large, high performance coolers that can run quietly with big fans.

C. There are two socket 775 platforms for the simple reason that the Intel 930 and 950 processors, which arrived late in the testing process, would not run with the AOpen i945Ga-PHS board on which other 775 socket processors had been tested. This is despite running the latest available BIOS that was supposed to expressly support the 900 series. The Intel D945GTP did support the 900 series processors, but it was not an ideal test bed because of a very limited BIOS and lack of support for Windows-based clock-speed and Vcore utilities. This is why the 930 and 950 were tested only with default voltage settings; no other setting could be applied, either from the BIOS or via Windows utilities.

TEST METHODOLOGY

Let’s review the question we’re trying to answer here:

"What is the best power efficiency achievable with currently available AMD and Intel processors that can be used on a desktop PC?"

In keeping with the spirit of the question and its focus on minimizing power consumption, the approach we took was to optimize each CPU for the lowest stable Vcore setting. This gave us the lowest CPU power consumption. We then took power measurements. The procedure was as follows:

Step One: The CPU was installed and tested at totally stock setting on the test platform. The CPU heatsink/fan was set to ensure that overheating could not be a source of instability; we generally disregarded noise for this project.

Step Two: We then used CrystalCPUID to reduce the core voltage of the CPU (Vcore) step by step while running the CPUBurn stress utility (one or two instances to match the number of cores or hyperthreading) at medium priority. With each Vcore setting, the system was run for a minimum of 10 minutes to watch for instability. We’d run other programs simultaneously to check for instability; if there was any, it showed up fairly quickly.

Step Three: When instability was encountered, the system would often need to be rebooted. We then moved the Vcore incrementally above the setting where instability occurred, seeking the lowest Vcore at which stability under load could be reached.

Step Four: Once the minimum stable Vcore was found, we confirmed it by running CPUBurn for 20 minutes before taking measurements.

NOTE: You might criticize that CPUBurn for 20 minutes is not a great test of system stability, and we would agree, but we think it’s good enough. A system that goes 20 minutes of CPUBurn without instability is reasonably stable. In real world conditions, it’s rare for a desktop system to undergo this high a load for this long. It may be that for these processors to survive a 24 hour test with Prime95 — an often used standard for system stability — the Vcore would have to raised slightly higher. But that would have taken way too long, and this project was tedious enough already.

While we took care to obtain accurate details about CPU power consumption, in the real world, power efficiency is best measured at the AC socket. Drops in efficiency of the PSU and the VRM at low and high power levels tends to compress CPU power differences. The concept of Average Power as a measure of real world computer efficiency is also introduced and discussed.

ACTUAL MEASUREMENTS

All processors were measured in three states to establish the full range of
power consumption:

  • Idle, with Cool’n’Quiet, PowerNow!, or Enhanced Intel Speed Step (EIST) enabled when available.
  • Under heavy load using CPUBurn
    to stress the processor at the default Vcore for the CPU.
  • Under heavy load using CPUBurn
    at the minimum stable Vcore for the CPU.

Two power measurements were taken in each state:

  • DC power at the 2x12V (AUX12V) connector on the motherboard.
  • Total AC power consumed by the system as a whole.

DC power measurements involved a high precision current sensor plugged directly into the 2x12V connector on
the motherboard, so that all power through this connection passed through
the power meter. The line voltage (nominally +12V) and the current are measured
with multimeters, and multiplied together to get the total power running through
the connection. Because the CPU only draws power through the 2x12V connection
and nothing else does, this tells us the amount of power consumed by the
CPU and the voltage regulators on the
motherboard.

The DC power measurements do not take the efficiency of the voltage regular
module (VRM) on the motherboard into account. VRM efficiency does vary somewhat from board to board, and also with power level. The average VRM efficiency is not much higher than 80%, but not likely to be lower than 75%. So the actual CPU power draw is probably around 20% lower than the 2x12V current we’re reporting. Little is known about VRM efficiency at very low power levels, like <10W in idle. We suspect VRM efficiency could drop substantially below 75%..

FYI, if we could measure power right at the CPU socket, we could characterize not only CPU power demand but also the VRM efficiency of motherboards at different power levels. However, taking voltage / power measurements directly at the CPU socket requires an investment we cannot justify. We know of a system developed by a power engineering team at Intel; the components would cost US $3,000 for basic equipment plus ~$3,000 per CPU socket type.


An custom-built shunt featuring a LTS 25-NP current sensor allowed us…


…to measure voltage and current on the +12V AUX connector that powers
the CPU.

The final accuracy for this power calculation is better than ±1W, maybe as good as ±0.1W.

AC power was measured to obtain a power profile of each system as a whole.
By design, this includes power lost in the power supply itself during conversion
from AC to DC. Most power supplies become less efficient as they approach zero
output. At the low power loads of these systems, the power conversion loss may
account for as much as 50% of the total system power. Measurements for AC power
were read off of the digital display on the Extech power meter.


The Extech
Power Analyzer / Data Logger 380803
power meter kept track of AC power.

Complete List of Test Tools:

  • AC power was measured with an Extech
    Power Analyzer / Data Logger 380803
    power meter.
  • High accuracy Extech MM560 True RMS multimeter.
  • Two other multimeters of good precision.
  • High precision LTS 25-NP Current Sensor (to read the AUX12V current), courtesy of Intel.
  • A Fluke 36 Clamp Meter
  • Processor voltage was monitored using SpeedFan
    4.27
    or alternative motherboard utilities where SpeedFan was not supported.
  • Athlon 64 and Turion 64 specifics were documented using the utility A64 TCaseMax v1.18
  • Other processor details were checked with CPU-Z.
  • CrystalCPUID,
    a utility for setting and modifying clock speed and voltage.

TEST RESULTS

A. AT FULL LOAD

Prevailing conditions during testing were 20~21°C and 117~119VAC, measured with
the Extech power meter.

Power Consumption at Full CPU Load
Processor


Platform


Clock Speed
Vcore
CPU Power* (DC)
System Power† (AC)



Stock



Min


Stock



Min



Stock



Min


Intel P-M 770
Dothan


479-1


2.13 GHz

1.30V



1.12V


23.3W



13.8W


65W



53W


AMD Turion 64 ML-40
Lancaster


754


2.2 GHz

1.22V



1.05V


26.4W



18.1W


64W



54W


Intel Core Duo T2600
Yonah


479-2


2.16 GHz


1.26V


1.15V


25.4W


19.5W


78W‡


75W


AMD A64 3000+
Venice


939


1.8 GHz


1.39V


1.18V


31.9W


20.5W


74W


61W


AMD Sempron 3400+
Venice


754


2.0 GHz


1.39V



1.19V


38.9W



26.4W


82W



66W


AMD A64 3500+
Winchester


939


2.2 GHz


1.41V


1.24V


47.5W



31.6W


98W


80W


AMD A64 4000+
San Diego


939


2.4 GHz


1.39V


1.29V


42.8W


34.3W


88W


76W


AMD A64 3500+

Venice


939


2.2 GHz


1.41V


1.22V


53.8W



34.6W


104W


83W


AMD A64 X2 3800+
Toledo


939


2.0 GHz


1.39V


1.22V


58.0W


41.4W


109W


90W


AMD A64 X2 4800+
Toledo


939


2.4 GHz


1.37V


1.22V


87.3W


61.9W


144W


115W


Intel P4 630
Prescott


775-1


3.0 GHz


1.23V


1.23V


77.7W


77.7W


128W


128W


Intel P-D 820
Smithfield


775-1


2.8 GHz


1.31V


1.14V


124.0W


90.5W


181W


142W


Intel PD 930 Presler


775-2


3.0 GHz


1.22V


1.21V


93.6W


93.6W


146W


146W


Intel PD 950
Presler


775-2


3.4 GHz


1.22V


1.22V


105.1W


105.1W


160W


160W


Intel P4 670
Prescott


775-1


3.8 GHz


1.34V


1.34V


128.9W


127.7W


197W


195W

Notes for above table:
*CPU power includes losses in the motherboard voltage regulators. The power drawn by the CPU alone is probably ~20% lower.
The system power includes losses in PSU AC/DC conversion.

At >100W AC, it is probably not much more than 20%. At <70W AC, the loss could be >30%.


System power on the Yonah platform is higher because it required an outboard PCIe graphics card.
With an onboard graphics motherboard, it would be 10~15W lower.

The table above is ranked by CPU Power, with the lowest at the top. The top two-thirds of the table (<60W) contains only AMD processors and two mobile Intel processors. There is only one AMD processor, the X2-4800+, arguably the most high performance CPU in the entire group, that reaches 60W. Five Intel processors make the bottom ranks; the 670, a single core P4-3.8, requires more than double the power of the dual core X2 4800+.

This is not the result of us choosing processors to bias the test. The Intel 630, just about the slowest desktop P4 Intel offers at this time, draws more power than the AMD A64 X2 4800+, the second most powerful desktop processor AMD offers currently. To top it off, the 630 is a single core processor while the X2 4800+ is a dual core.

It is true that Celerons are missing while there is a Sempron in the mix. However, this would have made little difference. All the current 2.4~2.8 GHz Celerons have 60~68W TDP, while our Sempron 3400+ sample measured 31.9W including VRM losses. Also, retailers FrontierPC and NCIX in the Vancouver area reported that there was no demand for Celerons — and hardly any supply at this time.

B. AT IDLE

Power consumption at idle is similarly one-sided. At the top with 1~2.2W power consumption are the mobile processors. At the bottom are the Intel desktops at 21.6~33.6W. AMD desktop processor systems ranged from a low of 4.1W to a high of 9.8W, which is pretty narrow, given the performance range represented — Sempron 3400+ to X2 4800+.

Power Consumption at Idle
Processor


Platform


Clock Speed

CNQ / EIST
Vcore
CPU Power*
System Power†


Intel P-M 770
Dothan


479-1


800 MHz


yes


0.734V



1.0W



40W


Intel Core Duo T2600
Yonah


479-2


1.0 GHz


yes


0.95V



1.5W



53W‡


AMD T64 ML-40
Lancaster


754


800 MHz


yes


0.92V



2.2W



40W


AMD Sempron 3400+
Venice


754


1.0 GHz


yes


1.10V



4.1W



44W


AMD A64 3000+
Venice


939


1.0 GHz


yes


1.18V



4.8W



48W


AMD A64 X2 3800+
Toledo


939


1.0 GHz


yes


1.18V



5.6W



51W


AMD A64 4000+
San Diego


939


1.0 GHz


yes


1.18V



5.6W



51W


AMD A64 3500+
Winchester


939


1.0 GHz


yes


1.19V



8.1W



53W


AMD A64 X2 4800+
Toledo


939


1.0 GHz


yes


1.18V



8.8W



53W


AMD A64 3500+

Venice


939


1.0 GHz


yes


1.18V



9.8W



55W


Intel P4 630
Prescott


775


2.8 GHz


yes


1.23V



20.7W



64W


Intel PD 820
Smithfield


775


2.8 GHz


yes


1.17V



26.7W



71W


Intel PD 950
Presler**


775


2.4~3.4 GHz


yes


1.19V



6.4 / 31W



74W


Intel PD 930
Presler**


775


2.4~3.0 GHz


yes


1.22V



7.8 / 32W



75W


Intel P4 670
Prescott


775


3.8 GHz


yes


1.24V



33.6W



84W

Notes for above table:
*2x12V power includes losses in the motherboard voltage regulators. The power drawn by the CPU alone is probably ~20% lower.
† The system power includes losses in PSU AC/DC conversion. At >100W AC, it is probably not much more than 20%. At <70W AC, the loss could be >30%.

System power on the Yonah platform is higher because it required an outboard PCIe graphics card.
With an onboard graphics motherboard, it would probably be just over 40W.
** The Intel Pentium D 930 and 950 idle power measurements
require some explanation. When our standard high precision LTS 25-NP Current Sensor was used to read the AUX12V current, very low <10W values were obtained. Using our old Fluke 36 Clamp Meter on the 12V leads gave us the second set of 31~32W readings, which are much more believable. It’s an odd testing anomaly isolated only to these processors.

Idle CPU Voltage for the socket 939 processors could not be set below 1.18V on the ASUS A8N-VM CSM
motherboard used for thiis platform, neither in the BIOS nor with any software utilities like Crystal CPU ID. Chances are very good that the idle Vcore could be considerably lower, and this would mean lower system power as well. Some SPCR forum members have reported fine operation with <1.0V idle C’n’Q Vcore on A64 processors.

D. ENERGY CONSUMPTION

Here, we need to look at system power consumption at the AC socket. Realistically speaking, neither the full load nor the idle system power truly reflect energy consumption. In order to get a truer picture of energy consumption, we need to consider how much the system is in idle or at full load during the course of a typical day. Research by the US Environment Protection Agency for the Energy Star program shows that computers are in idle 95% of the time that they are powered on.

But that raises our hackles. Most of us are enthusiasts here; surely we must stress our PCs a little harder than that (we all say with a little machismo and swagger). So let’s say 90% idle and 10% full load for PC enthusiasts like us.

Then we can calculate the average electricity consumption of the systems with this simple formula: (0.9 x Idle power) + (0.1 x Load Power) = Average Power. The EPA’s draft computer specification scheduled for implementation next year states that a PC will be an Energy Star when it consumes no more than 49W in idle. So for argument, let’s use 49W as the reference and calculate how much more (or less) energy each system consumes, expressed as a percentage. Having gone this far, we might as well calculate the annual energy consumption in kWh assuming that each computer is turned on an average of eight hours a day.

Processor



Idle
Power (W)



Load
Power (W)



Avg. Power (W)


re: Energy Star
kWh/yr


Intel P-M 770
Dothan

40

53



41.3



-16.2%


120.6


AMD Turion 64 ML-40
Lancaster

40

54



41.4



-16.2%


120.9


AMD Sempron 3400+
Venice

44

66



46.2



-5.7%


134.9


AMD A64 3000+
Venice


44


61


49.3



+0.6%


144.0


AMD A64 4000+
San Diego


51


76


53.5



+9.2%


156.2


AMD A64 X2 3800+
Toledo


51


90


54.9


+12.0%

160.3


Intel Core Duo T2600
Yonah


53


75


55.2


+12.6%

161.2


AMD A64 3500+
Winchester


54


80


56.6


+15.5%

165.3


AMD A64 3500+

Venice


55


83


57.8


+18.0%

168.8


AMD A64 X2 4800+
Toledo


53


115


59.2


+20.8%

172.9


Intel P4 630
Prescott


64


128


70.4


+43.7%

205.6


Intel PD 820
Smithfield


71


142


78.1


+59.4%

228.1


Intel PD 930 Presler


75


146


82.1


+67.6%

239.7


Intel PD 950
Presler


74


160


82.6


+68.6%

241.2


Intel P4 670
Prescott


84


195


95.1


+94.1%

277.7

It’s easy to see that the idle power is much more important than the peak. The idle power is a much better indicator of typical energy consumption than load power. This would not hold true if you are a gamer with a high performance rig that is turned on only when you play. Then, the load power consumption becomes much more indicative of total energy consumption. Of course, all your power numbers would go up as much as >200W depending on what graphic cards you’re using.

E. THE EFFECT OF PSU EFFICIENCY

Looking at all the data so far, it’s clear that small differences in CPU power, either at load or idle, don’t really change the total energy consumed. Especially when the higher cost of low power processors like the Turion 64, Pentium M or Core Duo are considered, this doesn’t seem a very cost effective way to improved power consumption. What about a more efficient power supply?

The FSP Green 400 PSU used for all the testing is no slouch for efficiency. But we also happen to have a Seasonic SS-400HT that is 80 PLUS certified. It has even better efficiency. We’ve tested both units in our labs.


FSP Green 400W PSU

SEASONIC SS-400HT, 80 PLUS

DC Output Load
(W)

AC Input (W)

Calculated Efficiency

AC Input (W)

DC Output
Load (W)
40.4
58
69.7%
76.6%

55
42.1
63.7
84
75.8%
81.5%

78
63.6
89.7
114
78.7%
82.8%

109
90.2
150.6
185
81.4%
85.3%

180
153.5
199.3
245
81.3%
85.3%

231
197.2

So what happens when we swap the FSP Green out for the Seasonic SS-400HT 80 PLUS? (Both PSUs can be found online for ~$70.) We chose just a few systems to try this, as time was running short. Only the data on system power is cited, as there is no change in the power delivered to the CPU / 2x12V connector.

Processor



Idle
Power (W)



Load
Power (W)



Avg. Power (W)



FSP Green



Seasonic 80 Plus



FSP Green



Seasonic 80 Plus



FSP Green



Seasonic 80 Plus


Intel P-M 770


40


37


53


49


41.3


38.2


AMD Turion 64 ML-40


40


37


54


50


41.4


38.2


AMD A64 4000+


51


47


76


72


53.5



49.5


AMD A64 X2 4800+


53



50


115



108


59.2



55.8


Intel PD 930


75


71


146


138


82.1


77.7


Intel P4 670


84


78


195


187


95.1


88.9

It’s clear that power supply efficiency is a key determinant in system efficiency. With the higher power processors, the gains are most dramatic, but they are measurable even with the lowest power processors. Especially when there is no price difference between the two power supplies, the choice is obvious.

It should be noted that we are comparing two relatively high efficiency power supplies. If we were to compare the Seasonic to a typical or generic PSU which has 70~72% efficiency at best and <65% efficiency at low loads, the real differences in energy consumption would exceed 15W average power in many systems. Over the course of a year or for an enterprise running hundreds of PCs in a building, this difference is very significant.

NOTE: SLEEP MODE VS. HIGH EFFICIENCY PSU – The jury is still out on whether high efficiency PSUs or full implementation of effective sleep and advanced idle modes is a more effective means of reducing energy consumption by computers. For various points of view on this matter, please see the relevant PDF presentations at the Energy Star Computer Specification web page. HP, for one, contends that effective, universal implementation of Sleep Mode is not only much cheaper than a more efficiency power supply, it actually leads to much greater energy savings all around.

F. WHAT ABOUT TURION 64 VS. ATHLON 64?

Our original article about the Turion 64 on the Desktop examined the this AMD mobile chip as a lower cost alternative to the Intel Pentium M. One of the often-repeated questions that came up after the Turion 64 article was,

"How does the Turion 64 fare against the latest Athlon 64s, in terms of power efficiency, computing power, ease of use on the desktop and price?"

This article was started originally to answer the above question. We have enough data to answer it. Please consider the summary table below. Note that all the figures are based on minimum stable CPU voltages; they are the lowest power numbers we could obtain in the systems we assembled.

Turion 64 versus Athlon 64
Processor

Price


Platform


Max Clock

CPU Power* (DC)
System Power† (AC)




Avg. Power


Idle



Load

Idle



Load


AMD Turion 64 ML-40
Lancaster


$268


754


2.2 GHz


2.2W

18.1W


40W

54W



41.4W


AMD A64 3000+
Venice


$140


939


1.8 GHz


4.8W


20.5W

48W


61W



49.3W


AMD A64 4000+
San Diego


$341


939


2.4 GHz


5.6W


34.3W

51W


76W



53.5W


AMD A64 X2 3800+

Toledo


$301


939


2.2 GHz


5.6W

34.6W

51W


83W



54.9W

1. NOISE

A major reason for seeking out low power CPUs is that they can be cooled more easily with less noise than the high power ones. The T64 ML40 has an idle power of just 2.2W and full load power of 18.1W. With an appropriately large heatsink optimized for low airflow, this CPU could probably be cooled just by a low speed fan in an efficient power supply — assuming a minimalist system with few heat producing parts, with the CPU placed close to the PSU, as is the norm in most case layouts.

Can a similar performance Athlon 64 be cooled passively? At idle, yes, but at extended full load, probably not except for the A64 3000+. The other A64 models exceed 30W at full load, and they would probably need a little more cooling, either an extra case exhaust fan on the back panel near the heatsink, or a fan directly on the CPU heatsink. But the airflow could be kept very low. The noise difference between a T64 system with one fan in the PSU versus an A64 system with an additional very quiet fan on the CPU might be inaudible in many environments for many users. We could be talking about as little as a 1~2 dBA@1m difference. When you consider that a dual core X2 3800+ is in the mix, the added performance might well be a worth the price of a 1~2 dBA noise increase. Pricewise, the cost of the X2 is just 10% more.

If the A64 3000+
Venice was overclocked to 2.0~2.2 GHz to match the performance of the T64 ML-40, we could expect the CPU power at load to go up to 25W. This might still be cool enough for passive cooling with a big heatsink like the Scythe Ninja. But the problem is that it’s unlikely that this clock speed would be reachable while the CPU was undervolted as it has been to reach the 20.5W mark at 1.8 GHz. Raise the Vcore back up to default while overclocking to 2.2 GHz, and we’re probably at 30W, which is not a realistic passive cooling target.

2. POWER

The relevant data to examine is the AC power consumption. The 54W full load power of the T64 system is impossible for the others to match. They don’t even come close. The A64 3000+ idle power is 8W higher, and the others are 11W higher. If you look at the big picture, however, the energy consumption difference between the T64 and the X2 3800+ is not 30W; it is 13.5W average power, which is very small. This is assuming you use your computer much like most users do, in idle mode 90% of the time.

Even looked at only in average power terms, there seems no way the A64s can match the miserly T64. But they are not that far off. Compared to the 41.4W average of the T64 system, the others consume 19%, 29% and 33% more energy. The decision about whether the extra energy savings or extra performance is preferable is ultimately a personal one: It’s your choice.

G. THE NEW PENTIUM D 930 & 950

These new dual cores may be the last of Intel’s Netburst technology that began with the original P4. They differ from the 800 series in a number ways. Here’s a three-way comparison.


Feature


Intel 900

Intel 800

AMD A64 X2


Core


Presler / 65nm

Smithfield / 90nm

Manchester, Toledo / 90nm


Clock


2.8 – 3.4 GHz

2.8 – 3.2 GHz

2.0 – 2.4 GHz


FSB


800 MHz

800 MHz

HyperTransport
1 GHz


Cache


L1: 24+32KB, L2: 2MB
per core

L1: 24+32KB, L2: 1MB
per core

L1: 128KB, L2: 512KB or 1MB per core


Memory Controller


n/a

n/a

Integrated on-die,
dual ch DDR-400


Dynamic clocking


EIST

EIST

CnQ


64-bit


yes

yes

yes


Other


Virtualization, SSE 2/3, NX Bit Support

SSE 2/3, NX Bit Support

SSE 2/3, NX Bit Support

A primary difference between the Pentium Ds and the A64 X2 is that the former has two separate cores in one package while the latter has a fully integrated die with two cores within it. The Pentium D cores use the shared FSB to communicate with the memory controller located on the motherboard, which means worse memory latency and inability for the cores to communicate directly. The on-die memory controller of the Athlon 64 X2 has none of these limitations.

The big news with the 900 series is that it’s on a smaller 65nm die with 376 million transistors compared to the 800 series’ 90nm and 230 million transistors. There’s also been extensive optimizations to reduce idle power consumption as much as possible. These refinements do help the 900s to outperform the 800s while drawing at bit less power. Unfortunately, the improvement is not enough to match to the energy efficiency of the AMD X2 processors.

Intel Pentium D
Processor

Price

Clock

CPU Power* (DC)
System Power† (AC)

Avg. Power

Idle



Load

Idle



Load


Intel D 820


$241

2.8 GHz


26.7W

90.5W


71W

142W



78.1W


Intel D 930


$316

3.0 GHz


31W


93.6W

75W


146W



82.1W


Intel D 950


$637

3.4 GHz


32W


105.1W

74W


160W



82.6W


AMD A64 X2 4800+


$643

2.4 GHz

8.8W

61.9W

53W


115W



59.2W

The 930 and 950 were somewhat hampered in this comparison by a motherboard that locked the Vcore. The only board on hand that supported these processors was the Intel D945GTP. It was not an ideal test bed because of a very limited BIOS and lack of support for Windows-based clock-speed and Vcore utilities. It is very possible that these processors could have run with a touch less Vcore, perhaps 0.1V, which would have been enough to drop the load CPU power by at least 5W and the system power by as much as 8~10W.

CONCLUSIONS

This article has been an exhausting exercise, despite it being still not an exhaustive sampling of current processors for desktop use. Nevertheless, most of the major processor families from Intel and AMD were covered.

Our focus on thermals, power and energy efficiency led to mostly predictable results: Mobile processors are best, followed by AMD desktop processor in general, and then Intel desktop processors. The power efficiency of AMD Athlon 64 single and dual core processors is excellent, even for their highest performance models. The Intel desktop processors suffer from inefficiency, even on the 65nm die. We are not sorry that the upcoming Intel Conroe will soon sweep the Netburst generation away.

It was disappointing that many of the Intel processors and boards seemed to disallow downward adjustments of the CPU core voltage. Small reductions in Vcore cannot improve power demand dramatically, but even small power reductions in power and heat would be welcome.

The Core Duo is a delightful exception in Intel’s camp, with probably the highest performance-per-watt ratio of all the processors in our survey. Our experience with the AOpen / Core Duo combination leaves us eager to see more of these products in the market, both components as well as complete systems. The Core-Duo powered Apple 17" iMac — a sample of which we’ve been assessing for almost two weeks — has been an eye-opener about the kind of power efficiency, acoustics and performance that can be achieved today at a moderate price with intelligent system design by a mainstream computer maker.

The one set of data that surprised us was the estimated watt-hour rating of the various test systems. There was a greater than 1:2 ratio between the lowest and highest power system in idle. The ratio jumped to 1:4 when the systems were in full load. But in estimated average power consumption, the ratio dropped back down to 1:2. It tells us that idle power consumption is the single most critical aspect of energy conservation in computers. Sleep mode may be even more important than improved power supply efficiency.

As stated earlier, this article is a snapshot of the current desktop processor scene taken from a particular angle. In just a few months, new processor lines from AMD and Intel will have transformed the scene. It will be time, then, for another processor survey.

Many thanks for the generous assistance of:

Intel for processor and motherboard samples
AMD for processor samples and the Extech 380803 AC power meter
AOpen for
the i915Ga-HFS, 975Xa-YDG and
i945Ga-PHS
motherboard and AeolusPCX6600-DV128LP graphics card samples
Asus for the A8N-VM CSM
sample
FrontierPC for the Intel 630 sample.
NCIX for
the MSI RS482M-IL motherboard sample
Corsair and OCZ for
the RAM in the test systems
Seagate, Samsung and Hitachi for the notebook hard drives used in the test systems

* * *

SPCR Articles of Related Interest

The State of the Industry, March 2006: Through Silent Eyes
AMD Turion 64 on the Desktop

Quiet OC’ed Pentium D 830 System
Cool’n’Quiet PC for Torrid Thailand
CrystalCPUID: User Configurable Cool ‘n’ Quiet

CPU Undervolting & Underclocking Primer

* * *

Discuss this article in the SPCR Forums.

Leave a Comment

Your email address will not be published. Required fields are marked *