Desktop CPU Power Survey, April 2006

[Mats]

Wonderful article. I've been waiting for this one.

On page 7:
G. THE NEW PENTIUM D 930 & 950

It says the slowest 800 runs at 2.8 GHz (820). There's also a PD 805 running at 20x133 = 2.66 MHz.
It have become quite famous for it's low price. $122 for a dual core is just great. Inefficient, hot, low price.

[Mats]

The 3000+ in the article would produce 4 W more than the T64 if using the the stock settings used for the latter, if capable of it.
On the other hand, the 3000+E3 is most likely older than the T64, it's possible that a newer CPU would use less power, just like MikeC and AMD said. A new, undervolted 3700+ would maybe have the same power consumption as this T64, who knows. But then again we have the 1.1 VID limit, I'd really like to see mobos with undervolting options in [%].

I get this feeling that the individual TDP is individual for the actual core it represents, but that it's totally independent of what model it becomes.
So if two CPU's performs exactly identical in AMD's stress tests but one becomes a 3000+ and the other a 3800+ they would still have the same individual TDP, most likely made for the 3800+ or a faster not yet/never released.

I know it sounds strange, but I've seen lots of independent TDP's and it's quite difficult to see a connection between it and the actual speed/model. What do you think? It's easy to figure why it shouldn't be like this, but have anyone actually heard any words from AMD?

[merlin]

Mike, this is an awesome article and I love how much data is presented regarding the power consumption and large effects of undervolting and/or underclocking on power usage. I'm sure you spent far too much time on this and you have our thanks for doing so!

Me too!

...intel loves using performance per watt as a buzzword and it would be great to see how their current processors compared to what they're saying for merom/conroe.

That's a corporate-political issue, not an SPCR issue IMHO. It does not belong in Mike's article.


If I hadn't made the mistake of posting a coupla trivially simple fan "review"s, I would not believe some of the above criticism! I independently discovered that SPCR persons are highly demanding; they want what they want and they want it now!

I am going to have to disagree on the point about Performance per watt.. Performance per watt isn't just corporate-political, it's something fully important to all of us who care about energy consumption and noise. All of us want the fastest, cheapest and quietest stuff, don't you think?

And in regards to the 2nd comment, I'm just providing feedback on what would improve the article in my opinion. It's certainly not a demand and I don't expect any of it to be implemented anytime soon(if at all).

[smilingcrow]

Thanks Mike, I don’t remember reading a more stimulating article on this topic. Hats off to you for having the patience to pull it all together.

It’s left me with a number of ideas for future articles and one burning curiosity with regard the current test. I’m wondering what the A/C draw of the various systems would be if you were to install the Aopen 6600 graphics card into all the systems that used integrated graphics. That would give an indication of how power hungry the various IGPs are and give the opportunity to estimate more accurately a baseline figure for the Core Duo system when using an IGP.
I know that this article was looking at CPU power draw, but with all but one of the systems using an IGP it was screaming out to me to add a dedicated GPU to the mix for comparison. Of course the temptation then, at least for me, would be to add the entry level nVidia 6200 and ATI equivalent and compare those as well.
I say this because with CPU performance per watt and maximum power consumption both improving; it seems to me that overall system power draw is becoming more important especially if you’re looking to build smaller form factor systems a la Apple.

Towards the end of the year when the Conroe/Merom/AM2 and possibly Turion X2 desktop platforms are available, I’d be interested to look at these from an overall platform power consumption perspective. I’m not sure if Turion X2 will be available on the desktop due to using a unique socket design!
What I’m suggesting is to focus more on the different motherboard chipsets and IGP versus entry level GPUs for these different platforms, using the slowest and lowest power CPUs for each platform as a baseline for each of them. If you added one of the higher end CPUs for each platform to the mix, that would give the opportunity to compare those and also extrapolate the power consumption of mid range CPUs also.

With the variance in power consumption between even CPUs of the same model number and also I assume of motherboards from different manufactures using the same chipset, this can only be an indicative assessment; but it’s not one that I’ve seen anyone attempt. If there’s one place that I’d like to see this attempted it would be here, due to your attention to detail which I really respect. God is in the details as somebody said

This may seem a daunting project but I don’t think that it’s that much more time consuming than the one you’ve just completed. Comparing the graphic cards power consumptions at idle is fairly trivial for example.
It’s got me thinking that maybe SPCR readers might be able to contribute to such a large project if necessary; either by lending components or by sponsoring it via a small financial contribution. I say this because I get the impression that many people would like to see a particular combination of hardware included in a group test and being able to actively support that might be appreciated by those willing to walk their talk. Of course, this kind of communal interaction might not be something that appeals to you as it would undoubtedly take more of your time. It’s an idea that fascinates me and I’m curious to see whether people feel that this is practical at all.

Anomalies:

I noticed that the Asus S939 board shows all CPUs as idling at 1.18/1.19V as opposed to the 1.10V that I would expect. Is this due to the MB adding extra voltage by default?

In the Turion versus A64 table at the top of page 7, the last entry for the X2 3800+ seems off when you look at many of the figures.

[innuend0]

Hi mike,

I write from France, i'm was one of www.pcsilencieux.com (french website) administrator. (with "Kostik" and others)

On page 1 of the article, you say :

>> The power and heat issue is relevant to everyone else in a variety of ways:
>> * Environmental
>> * Performance
>> * Financial

But you don't talk about PFC (i mean Power Factor Correction) in the whole article !

As I understand, Active PFC (i mean Power Factor Correction) is a great benefit for reducing the power produced by "Electric power companies".
Yes, as I understand, all the power that is not consumed by a non-PFC (or a passive one) power supply is redirected on the network, but completely lost ... It can't be used again to be delivered to anybody else.

That's why I concluded, for an environmental point of view, that buying an "Active PFC" power supply is as important as buying a "high efficiency" one (and often cheaper !).

Am I wrong ?
Why don't you talked about PFC in this article ?

[qviri]

Why don't you talked about PFC in this article ?

Perhaps because this article is about CPUs only, and not the whole computers, and PFC applies to PSUs, not CPUs.

If you look through the PSU reviews on SPCR, you will find a lot of dedication to promoting the cause of Active PFC...

[jaganath]

But you don't talk about PFC (i mean Power Factor Correction) in the whole article !

Virtually all of the PSU's recommended by SPCR have active PFC, and none do not have it at all. Also, as Qviri has pointed out, this was not the focus of the article. Certainly SPCR has not neglected the issue of PFC.

[sane?]

I realise that you were focusing on CPU/Motherboard power efficiency in this article. However in a typical system there is much more going on - graphics card, drives, monitors, printers. It all adds up to overwhelm the CPU contribution.

For instance, my system is currently drawing something like 270W total
65W for CPU core 1
40W for CPU core 2 (which seems weird to me, but its measureable)
40-50W for Monitor 1
40-50W for Monitor 2
the rest for drives, motherboard, graphics, plugins, pump, printers, etc.

Even worse, with all those black blob power socket PSUs the thing draws 25W turned off!

I'd suggest that people get hold of one of the power socket energy monitors - its the quiet ones that suck all the power. You might like to examine the subject more widely and write an article on how the total system contributes, particularly monitors and drives. You might even like to create an online calculator to give it to people in heaters. For instance it costs me roughly the same to heat the place as it does to run this machine for 16 hours a day.


BTW the 90:10 ratio is probably not right, and its certainly not right if you are running something like BOINC in the background. 100% all the way, all the time. That's why I know that ~40W is right for the second core, I can see the load rise as BOINC starts up.

[dhanson865]

Correct me if I'm wrong but wasn't "system" power ignoring the monitor?

The quote was "Total AC power consumed by the system as a whole." but that leaves open the definition of "system"

As to avg CPU consumption Mine is at 10% right now and drops do 5% or less when I look away and watch TV...

[qviri]

I realise that you were focusing on CPU/Motherboard power efficiency in this article. However in a typical system there is much more going on - graphics card, drives, monitors, printers. It all adds up to overwhelm the CPU contribution.

And we are well aware of that.

However, the purpose of this article was not to describe an entire system. That would render it far too long. We are sorry if you came here looking for one article describing in depth all of the aspects of efficiency of today's computing. I guess you'll have to read three separate articles instead.

As for the 90/10 ratio, I think it is fair to say that an average user does not run distributed computing programs in the background. Unless you qualify trojan horses as "distributed computing", that is.

[MikeC]

Regarding question of why the Idle Vcore in CnQ was not set below 1.18V in the socket 939 processors, checking through my notes, I recalled that the Asus A8M-VM CSM motherboard simply did not allow the Vcore to be lowered below this. It's an unfortunate limitation.

Another socket 939 board w/ integrated graphics came into the lab late last week, a DFI RS482 based on the ATI chipset. Like other DFI boards in the past, this one has a huge number of BIOS options that look extremely useful. I have been able to set Vcore manually in the BIOS w/o losing CnQ functionality, which makes for very simple optimizations of Vcore for min power. Simply reduce the Vcore for stability at max load; the CnQ idle Vcore will automatically be lowered by the same amount. A followup piece examining idle CnQ power for socket 939 processors with this board is planned.

As for definitions of system power, yes, certainly most general users will have many other components in the PC than we had in our test platforms. And it's true that the general definition of a PC system should include things like monitors and printers. However, all that was beyond the scope of this article, which was focused specifically on CPU power.

Thank you for the comments and feedback. They'll be carefully considered for future articles along this theme.

[Mats]

sane?: Maybe you should read this. No monitors though.

[QuietOC]

Regarding question of why the Idle Vcore in CnQ was not set below 1.18V in the socket 939 processors, checking through my notes, I recalled that the Asus A8M-VM CSM motherboard simply did not allow the Vcore to be lowered below this. It's an unfortunate limitation.

I just used CrystalCPUID to setup some custom C'n'Q settings on a ASUS Terminator T2-AE1 with a D0 Sempron 64 3100+. This little ASUS barebones doesn't support tweaking of any voltage or setting in the BIOS. This is my first time using CrystalCPUID to replace the standard C'n'Q setting, but it seems to work great. I first tested each of the three states to make sure each was stable under stress. I got these three states:

Minimum: 4x200 MHz = 800 MHz @ 0.875V
Middle: 6.5x200 MHz = 1.3 GHz @ 1.000V
Maximum: 9x200 MHz = 1.8 GHz @ 1.175V

I have CrystalCPUID launch from the Startup folder with the /CQ and /HIDE command line options. This will hopefully save some energy over the standard C'n'Q values (1.8GHz @ 1.4V -> 1GHz @ 1.1V). Thanks!

[Filias Cupio]

Is there any chance of a graph showing power consumption, benchmark speed (taken from the standard hardware review sites) and possibly cost? Either one graph (speed/power on x/y, cost as text annotating each point) or two (speed vs power on one, computation-per-Watt vs cost on another.) These would make the trade-offs much more obvious.

[Hifriday]

1) Regarding thermals, there is no mention of CPU temps in the article. Is it safe to assume this is directly related to the CPU/system power consumption? (Noting that CPUs without the heat spreader - like the Turion probably cool better with their direct contact to the heatsink) If so, at what kind of wattage would one expect a difference in temps to start appearing in a fanned quiet rig?

2) Why is it that the different AMD CPUs have such different power consumption even when run at the same speed and voltages? Noted the sample variance and also the difference in TDP, but when we look at the idle draw with nearly all A64 CPUs at 1000mhz 1.18v, the 3500+ Venice stands out like a sore thumb with 10.4W. From your Tmax screenshots, it seems to have roughly the same TDP as the X2 4800+ which despite it's dualcore only consumes 6.6W or 63%.

Can anyone help answer these questions or provide some links to where more info might be found?

Also is it correct to assume that chips with lower TDP/TCaseMax will on the most part 1) run cooler, 2) undervolt better, and 3) consume less power (compared with a chip with higher TDP but at same voltage/frequency settings)? If this assumption is correct, and AMD's on chip TCaseMax ratings are correct, then it would seem recent s939 Opterons are excellent contenders to the Turions. On the TCaseMax screenshots forums (albeit limited samples) show Opt 144s with 38.2-44.1W TDP. Opt 165/170 with 5 samples at 35.0W TDP and one sample at 55.0W whereas the similar X2 3800+ are posting 58-84W! The 3000+/3200+ Venice also seem decent, posting down to 38W, but unlike the Opterons, they cannot be undervolted below 1.1v via software, whereas many Opterons seem to run fine at stock speed undervolted to 1.0v.
It would seem the Turions still run the absolute lowest watts, but maybe only by a few watts that aren't significant in practical terms?

[MikeC]

Hifriday --

We weren't looking at cooling systems at all, a variety of different HSFs were used, generally disregarding noise. The objective was to ensure that inadequate cooling would not affect the efficiency of any components including both CPU and VRMs. The amount of "cooling power" required to cool a CPU to a target temp -- assuming all other factors are the same -- depends entirely on the heat it is producing.

As to the power variances in CPU idle, in general, it seemed fairly closely tied to the individual rated TDP.

Whether lower TDP/TCaseMax make for better undervolters... it seems probable but there's really no way to determine without a large batch of samples. I don't think the number in our lab is enough to point to any kind of "trend".

Turions do run coolest, and the margin may not be "big". Whether the difference is significant has to depend on application, user, etc.

[QuietOC]

Regarding question of why the Idle Vcore in CnQ was not set below 1.18V in the socket 939 processors, checking through my notes, I recalled that the Asus A8M-VM CSM motherboard simply did not allow the Vcore to be lowered below this. It's an unfortunate limitation.

I wonder about this. It seems like this locked lower voltage range might have something to do with certain processors and not motherboards. My new E6 Sempron 64 3100+ will not undervolt under 1.06V (1.1V setting) using either the BIOS or CrystalCPUID, but the E6 Sempron 64 2500+ will undervolt down to 0.88V on the same board using the BIOS (and of course offers no C'n'Q.) My D0 Sempron 2800+ also had the 1.06V BIOS limit on this board, while a 130nm Athlon 64 2800+ could go down to 0.850V at 800MHz. It seems like certain CPUs might be locked from lower voltages while others are not.

[MikeC]

I wonder about this. It seems like this locked lower voltage range might have something to do with certain processors and not motherboards..... It seems like certain CPUs might be locked from lower voltages while others are not.

On the basis of my still early experiments with the DFI482 board mentioned a couple posts back, this does not seem to be the case. The board + 2 of the CPUs used in the article undervolted to <1V in CnQ.

[smilingcrow]

I’m also now not clear on why certain AMD CPUs are limited to 1.1V when under-volting via software. I thought that this was purely a CPU limitation that was introduced with Venice and later revision! So, is it a CPU, chipset or BIOS limitation?
My Asus A8V deluxe would happily under-volt below 1V when using a Winchester 3000 using CrystalCPUID. When I installed an X2 3800 Manchester in exactly the same setup, I was limited to 1.1V as a minimum voltage. CrystalCPUID would allow me to set the voltage to below 1.1V, but it and other tools showed that the voltage was 1.1V and power consumption measurements confirmed it.
The only thing that changed was that the O/S was upgraded for a multiprocessor system.

[fastturtle]

Great Article Mike and couldn't be any better timing as I've been forced to rethink some ideas in regards to a new system. Only thing that would complete this is a comparison/testing of some 2xx series opterons as to their basic TDP/TCaseMax ratings against Xeons as it should provide food for thought.

[Hifriday]

MikeC, it seems that the CPU power consumption figures have been revised for the idle chart. Although not significant and doesn't change the outcome of the article, I'm curious why this was revised (or am I mistaken)?

[MikeC]

A couple of tiny adjustments -- due to transcription error on my part. Amazed anyone saw it.

[Hifriday]

Only reason I noticed was the X2 4800+ now has a closer power draw to the 3500+ Venice, which is what would be expected if we assume similar TDP means similar power draw at same clock/volts.

Since you are on the forums, can you shed any light on how much difference say a 10W increase in CPU power draw will affect temps? I know there are a lot of variables, but if any info from your testing would give any rough ideas or even if you would hazard a guess?

[MikeC]

Since you are on the forums, can you shed any light on how much difference say a 10W increase in CPU power draw will affect temps? I know there are a lot of variables, but if any info from your testing would give any rough ideas or even if you would hazard a guess?

You seem to forget that power (heat) and temperature have a relationship that is shaped by many other factors. They two do NOT have a linear relationship, nowhere close.

You can have a 10W heat source in a 1cm cube with no cooling save convection... and maybe encounter >100°C on its surface. You can also have a 100W heat source spread across a 1meter cube and see temperatures on its surface only marginally above the ambient air, if at all.

I know you want me to say "in a typical low noise setup on one of the tested CPUs". I haven't paid much attention to this. It depends way too much on conditions to generalize.

To wit -- if you have an 80W@load CPU and a cooling system that keeps it down to 50°C, then adding 10W could make the temps jump 10°C... or 1°C. I've seen both. It just depends on the cooling system -- not just the HS & fan, but the fan's airflow, and the overall airflow around the CPU as well. Ditto a 20W@load CPU. Same comments apply.

[jaganath]

can you shed any light on how much difference say a 10W increase in CPU power draw will affect temps?

In my own experience, 10W can make the difference between being able to cool a CPU passively or not. 30W is doable (just), 40W is not without hitting worrying temps (with my heatsink, which is a CNPS7000B lookalike).

Remember, the specific heat capacity of dry air is roughly 1 joule per kilogram kelvin. 1 joule is equal to one watt-second, ie the work done to produce one watt continuously for one second. I'm sure there is some kind of physical equation that can be done to solve for the rise in temperature of a given mass of air for a given input of thermal energy.

[Mats]

On page 7:
G. WHAT ABOUT TURION 64 VS. ATHLON 64?

The data for the X2 3800+ seems to be totally mixed up with the 3500+ Venice (which is not present in that table).

[zenyatta]

I'm impressed with this article just as the next guy but... can you please drop a hint as to when a follow-up could be published? I'm about to build my next box and can't decide between Socket AM2 and Socket 775. Power consumption is one of the deciding factors for me and my patience is low (once you start browsing the hardware sites it's hard to stop drooling )

[Devonavar]

Sooner or later, once we work out all of the methodological bugs.

I can say that, unless you intend to leave your system fully loaded for long periods of time, the AMD chips are still more efficient because they do much better at idle. For some reason Core 2 Duo chips seem to consume more power than they need to at idle, and given that most systems spend 95% of their time idling, that makes AMD a better choice if power consumption is a concern. However, we still don't really know how much Core 2 Duo chips consume under load, since much seems to depend on exactly what load.

[jaganath]

the AMD chips are still more efficient because they do much better at idle. For some reason Core 2 Duo chips seem to consume more power than they need to at idle

Could be the relatively high idle clock (1.6GHz C2D vs 1GHz AMD), although it can't explain all of the discrepancy.

[flokel]

Hi,

as this one is imo the best cpu review in terms of silencing,
I was wondering if there's a new review planned?

Would be interesting to see Core (2) Duo and AM2 (EE/SFF/65nm) competing against each other.

Regards
Flo