Energy efficient server

demko · Post by **demko** » Mon Jun 04, 2007 10:39 am

Hello everyone,

I currently have a small server running. It's a small Cubid case with a VIA EPIA (Nehemia) 1GHz CPU with 2x512MB of PC133 SDRAM and an old WD 80GB hard disk. It has an external (notebook-like) power supply with no fan.
In terms of energy efficiency I'm very pleased with it. It uses about 20W in idle mode and a tiny bit more when actually doing something.

For a simple webserver this small computer is sufficient, but I'm also using it as a database (MySQL) server. For this, it's not really that good. It's actually quite sluggish if queries become more complicated.

As you might have guessed I want to replace it with a more recent (actually more powerfull) 'server'. Since I don't want to spend too much money on it and do want a rather powerfull solution I was thinking of a dual core platform like the AMD Athlon64 X2 AM2 or an Intel Core 2 Duo platform. But it also shouldn't draw too much juice since it's supposed to be always running.

I read the Intel Core 2 Duo E4300 has a reasonable power consumption both in idle mode and in stressed mode. I also read it can be undervolted at a little below 1.0 Volts.
So I suppose the E2140 (branded 'Pentium') which runs at a little lower pace and has half the cache is probably at least as energy efficient.
Since I'm going to use it as a server the mainboard can do with integrated graphics. As I found out the chipsets for the Intel platform are not as energy efficient as those for the AMD platform. How is this for the 945G and 946GZ chipsets? anyone knows?

As for the AMD AM2 platform, would the Athlon64 X2 3600+ in combination with an AMD690 chipset based mainboard be a reasonable choice (in terms of performance and power consumption)?

For power supply I prefer to use a normal PSU (with fan). The 80+ power supplies look interesting, but usually they are only 80+ with a load of at least 20%. So my best guess is I should look for an 80+ PSU with a low power rating (since 20% of 380W is less than 20% of 600W)?

Now, the big question is (of course) which platform should I choose? Intel E2140, or AMD X2 3600+ (or something else maybe? Wait for Penryn?).

I know this site is called silentpcreview, but for this one I'm more interested in the power consumption and the performance of the pc I want to build. Now I don't expect it to be as low-power as my current Via EPIA, but let's say 40 Watts in idle would be very nice.

Any information which can help me choose is welcome of course. If someone actually measured the power usage of such a system I'm of course very interested.

All tips welcome

jaganath · Post by **jaganath** » Mon Jun 04, 2007 11:06 am

Now I don't expect it to be as low-power as my current Via EPIA

actually it could be very close. You should contact a user called "Palindroman" via PM (private message), low-power stuff is his area of expertise, I believe he has an X2 3600-based setup which only draws 23 watts!

Beyonder · Post by **Beyonder** » Mon Jun 04, 2007 11:09 am

I've been looking at similar configurations. Based on numbers I've seen floating about the web, < 40 watts probably isn't going to happen. I think a minimal C2D or A64 system is going to draw 60-90 watts at idle, although admittedly the information is a bit lacking. The 690 chipset seems to be the most efficient AMD platform; the Intel chipsets seem less efficient in general.

I've been leaning towards the A64/690 since the idle draw seems slightly lower, and the 3600+ is dirt cheap at the moment.

jaganath · Post by **jaganath** » Mon Jun 04, 2007 11:12 am

< 40 watts probably isn't going to happen.

as said above, Palindroman has a rig that draws 23 watts from the wall. Now, if you want to use several 3.5" HDD then <40W is likely not possible, but I don't know if that's what you meant.

Beyonder · Post by **Beyonder** » Mon Jun 04, 2007 11:25 am

I tried to find the information where Palindroman elaborated on the details of that power consumption (I'd be very interested), but I couldn't.

I have yet to see anyone who's had that much success with undervolting/underclocking. Most people seem to report better results with CNQ.

BillyBuerger · Post by **BillyBuerger** » Mon Jun 04, 2007 12:51 pm

I wish I had a Seasonic power angle or something around. I recently put together some PCs for my workplace. AMD X2 3600+ (Brisbane) with a Gigabyte 690V motherboard in an Antec NSK-3300 case (Seasonic SS-300SFD PSU) These things definitely run cool. I'm running the CPU at 0.9V/1GHz idle and 1.0V/1.9GHz load. The case exhaust doesn't feel warm at all. The PSU exhaust is just a little warm. The only thing that isn't helping the power is the additional ATI X300SE video card that I added for dual monitor support. That thing runs pretty warm although probably doesn't draw too much power still. I would love to know what the total power draw is. I'm guessing it's less than 50W idle.

demko · Post by **demko** » Mon Jun 04, 2007 12:52 pm

Palindroman indeed has a special interest in low power PC's. He actually has a website were he offers low power PC's.
Interestingly enough one of his business partners is a computer dealer near were I live. These websites are in Dutch (sorry).

I recently ordered the parts for an AMD system there (not being for this server, but for a business PC). It might be a good idea to play a bit with this system and measure the power consumption before I order anything for my server.

I realise that undervolting cannot do wonders (and I might have to lower the clock frequency aswell), but Cool'n'Quiet is indeed an interesting alternative feature.

jaganath · Post by **jaganath** » Mon Jun 04, 2007 2:00 pm

I tried to find the information where Palindroman elaborated on the details of that power consumption (I'd be very interested), but I couldn't.

this was communicated via PM,the system is still undergoing testing. I'm sure he will make a new thread giving details soon. the bare specs are as follows:

80W PicoPSU
84W Edac AC/DC adapter
Gigabyte GA-M61PM-S2
1 x 1024 MB Ram, 533 mhz
Samsung DVD-RW
Samsung 2.5" 60 GB HD
AM2 X2 3600/Sempron 3200

I have yet to see anyone who's had that much success with undervolting/underclocking.

one example: passive AC64pro on A64 3800+ G1 @ 1900mhz 0.95volt = 55*C

there are literally hundreds more scattered over the SPCR forums. Smilingcrow in particular did a lot of work in that area.

I realise that undervolting cannot do wonders

certainly it can do wonders. because CMOS power consumption scales with Vcc squared,a small reduction in Vcore can give a large reduction in power consumption.

johno · Post by **johno** » Mon Jun 04, 2007 3:36 pm

There are times where undervolting doesn't do a lot for you. Most PCs are idle-bound and the standard athlons idle at 1.1V / 1GHz under cool n quiet. There's a thread active at the moment that's saying the CPU itself draws aroud 4W at this point. (.95/1.1)^2=75% - meaning you save 1W - maybe 1.5 once you take the marginal power supply efficiency into account.

But then there are some people who are turning off the dynamic CPU scaling in favour of static undervolting, because they don't know how to do dynamic undervolting. In that case I can see you being worse off than with the default CnQ.

jaganath · Post by **jaganath** » Mon Jun 04, 2007 3:56 pm

fine, undervolting has the most pronounced effect on power consumption at full load; this is mainly because current idle power is so low, not because undervolting doesn't work or doesn't ALWAYS reduce power consumption.

there are some people who are turning off the dynamic CPU scaling in favour of static undervolting, because they don't know how to do dynamic undervolting. In that case I can see you being worse off than with the default CnQ.

I run a S754 Semp 3000 at 1.5Ghz,1.1V (ie static undervolting); this averages out to use less energy than the standard default CnQ mode over an average day's usage, according to my kill-a-watt-type power meter.

fri2219 · Post by **fri2219** » Mon Jun 04, 2007 4:02 pm

Many uses of computers are obscenely bloated wastes of energy as far as I'm concerned, but some applications really justify sucking watts.

I'm not sure what constitutes a heavyweight query in your eyes, so ignore me if I'm out of line- I develop some pretty demanding database applications for a living.

If the queries are getiing bogged down by taking advantage of MySQL 5.x features like correlated sub-queries or grabbing millions of tuples in one shot, you want memory bandwith and I/O bandwidth more than anything. MIPS are almost a secondary concern. The best chip using your constraints would probably be a 2G Opteron.

Jeremy Zawodny has an astounding amount of material on the subject floating on the web, you might want to look into what he does in the course of his job- he's Yahoo's MySQL guy. (They use MySQL for downloading their read only stuff into). Another good person to query for discussing what your circumstances might dictate would be Paul DuBois.

If you had not already purchased some hardware, I would have suggested trying colocation- it can be surprisingly cost effective, and probably more energy friendly in the long run.

johno · Post by **johno** » Mon Jun 04, 2007 4:12 pm

For power supply I prefer to use a normal PSU (with fan). The 80+ power supplies look interesting, but usually they are only 80+ with a load of at least 20%. So my best guess is I should look for an 80+ PSU with a low power rating (since 20% of 380W is less than 20% of 600W)?

The PSU is a big area of energy use in a lot of systems. I'm assuming this is a fairly low traffic server, and so largely idle. In that case, working off my estimated idle power usage in my PC,

Motherboard: 18W (asus M2A-VM)
Power Supply: 15W (SU-380)
Hard Disk: 10W (WD3200AAKS)
CPU: 6W (Athlon 4800+ 65nm)
DVD-RW: 1.5W
Fans: .5W
total: 49W

So the motherboard is key. The trick there is to find one that someone has reported good energy use out of. Just the chipset is not enough - efficiency has a lot to do with the power converters on specific motherboards.

Power supply is also a choice which will give widely varying results. That SU380 is fairly efficient as PSUs go - some will consume up to 25W to drive that load, and a more efficient one, like a pico PSU, would be down to 8W or less.

With the hard disk, you make a big difference by going to a 2.5" disk. After taking PSU inefficiency into account, you can save 10W in a 2.5" over 3.5" disk.

Then you can save a bit of power in unplugging the optical drive.

But really, for a web server, the easiest way to save power, is to put the sites on a shared webhost

Beyonder · Post by **Beyonder** » Tue Jun 05, 2007 8:03 am

The machine I'm building is going to be for Media Center development, so web hosting isn't an option for me. Mostly, my biggest interest is whether or not I'm going to get lower idle power consumption with the Brisbane 3600+ or an L2 stepping C2D. Idle seems particularly important, because I think the machine isn't going to be doing much actual work for 90% of the time it's sitting there.

Although the discussion of motherboard efficiency is really interesting, it's difficult to get reliable information on how much power any given motherboard consumes. For example, I was considering the ASUS M2A-VM, but I really have no idea how efficient it is in comparison to other 690G motherboards, or other chipsets.

Right now I'm leaning towards:

ASUS M2A-VM
AMD Athlon 64 X2 3600+ Brisbane
A-DATA 2GB (2 x 1GB) DDR2 800

Total: $212

...I already have a power supply (Seasonic Tornado; not great, but probably not bad either), DVD drive, case, etc. I figure if I'm unhappy with the integrated video, I can replace it later, but what I've seen online regarding the 690G seems generally positive.

One other thing I worked out--energy costs about ~7.5 cents per kilowatt hour in my area (range was ~6.8 to ~8.5 or so over the last year), so a server running 24-7 that consumed ~60 watts would consume about 1.44 kwh per day, or 526 kwh per year. This works out to about $39 a year.

Then I have a cable modem and WRT54G router. I don't know the power consumption for the two, but they're both warm to touch, so I'll assume ten watts per device (? dang I need to get a kill-a-watt). That'd be another $13 a year, although admittedly I'm paying these costs without an always-on web server. So roughly ~$52 to run a server over the duration of a year?

The cost of running router/cable modem sucks, but it'd be nice if I could get the server down to < 40 watts at idle.

johno · Post by **johno** » Tue Jun 05, 2007 6:51 pm

The Brisbane and L2S C2D desktop chips themselves seem to draw about the same at idle (ie, virtually nothing) - but the issue really comes down to the motherboard consumption - and all evidence I've seen puts the C2D chipset consumption well above the athlon one.

The M2A-VM is a fairly safe choice. Quite a few people here have them and are reporting reasonable power use, but I'm fairly sure it isn't the best available. It may be close to lowest out of the set of boards with DVI output though. We really need more people with the kill-a-watt or similar and reporting their configuration and power use.

Your power consumption calculations demonstrate how energy conservation and economic rationalism are in competition with each other. Its hard to recover the cost of energy efficient components though reduced electricity bills. It shows that electricity is too cheap - and that's not going to change anytime soon whilever low energy costs are so important to each country for their own domestic industry protection.

demko · Post by **demko** » Wed Jun 06, 2007 1:33 am

First, let me say thank you to all of you for posting all this interesting reading.

Correct me if I'm wrong, but if I understand correctly undervolting the CPU might cause a drop in the CPU power usage, but since the CPU isn't by far the only component that uses a significant amount of power I might reduce power usage more by using an efficient PSU or mainboard.

A shared webhost is not really an option for me since I run applications that use up quite a lot of CPU time. The EPIA I currently use is sometimes fully occupied with MySQL, which causes a bad responsiveness on the webserver.
I don't consider my queries really heavyweight (I'm guessing they're not by far as complex as the ones fri2219 develops...) but I indeed use MySQL 5.0 and correlated sub-queries. But the largest table in the DB I mainly use has a little over 5 million records. I don't think it justifies the power consumption of a big Opteron server.
In terms of memory and I/O bandwidth I think almost anything is a big improvement over the current EPIA with PC133 SDRAM and ATA100 harddisk.

Using a notebook HDD might indeed lower power consumption, but I don't want to kill the database performance by using a slow HDD. How do notebook HDD's compare to desktop ones actually?
Another solution could be the use of a solid state harddrive. Expensive and lower capacity, but energy efficient and fast (low access times anyway).

The power consumption calculations of Beyonder are also interesting. I guess he's from the States? That means I have to adjust some factors, since I live in The Netherlands.
Well, 526kWh would cost me about 125 euro's ($166). Mainly taxes.
The mainboard, CPU and memory would cost me about 220 euro's ($293).
For me this is also a good motivation to get an energy efficient server. It's also the reason I'm not too keen on colocation. All colocators here calculate relatively high prices because of their electricity costs.

Of course I will let you know what I'm going to buy. But I will first build the business PC I've already ordered. It will have the Asus M2A-VM mainboard, an Athlon64 3800+ 45W CPU and a 380W Antec PSU. I'm quite interested to see how much power it will draw from the wall, so I'll measure that first.

Palindroman · Post by **Palindroman** » Wed Jun 06, 2007 1:50 am

demko wrote: I realise that undervolting cannot do wonders (and I might have to lower the clock frequency aswell), but Cool'n'Quiet is indeed an interesting alternative feature.

Actually, undervolting does do wonders. I recently wrote an article about it and found that especially under load there's a lot to be gained. By undervolting a Sempron 3400+ I gained 4 watts (of the 40W with C 'n Q enabled). Under load the power consumption (measured at the wall) went down from 77W to 50W. That's 35% less with a technique for which you don't have to be a computer expert. I know I am not.

We really need more people with the kill-a-watt or similar and reporting their configuration and power use.

Up till now I've tested about 10 different setups with small variations. I will report my findings some of these days, as soon as my company has the position I want it to have (ie be recognized as the first company who started doing this). I'm all for transparency and think it's one of the key issues to promote innovation and conscious consumer choices.

edit: Demko, I would love to know what that 45W Athlon does powerwise, so please keep us informed! If you need more info or advice, you can PM me or mail me through my website. If you want to know more about undervolting buy the latest release of the Hardware.info magazine.

Beyonder · Post by **Beyonder** » Wed Jun 06, 2007 8:26 am

johno wrote: Your power consumption calculations demonstrate how energy conservation and economic rationalism are in competition with each other. Its hard to recover the cost of energy efficient components though reduced electricity bills. It shows that electricity is too cheap - and that's not going to change anytime soon whilever low energy costs are so important to each country for their own domestic industry protection.

At the risk of digressing, I'm interested in not paying an extra $40 a year just to run the equipment I purchased. If I can save $20 a year just by part selection and configuration, I'm totally game.

Also, my girlfriend and I decided that we're going to order green credits, which are more expensive but also focus more on renewable, clean energy sources. So I think it's likely that my energy costs are going to go up anywhere from 1-5 cents/kwh.

That's it. I'm ordering a kill-a-watt today.

edit: nuts. I just realized that the Seasonic Tornado has a 20-pin ATX connector. Does this mean I need a new PS for the Asus mb (which has 24-pin)?

Palindroman · Post by **Palindroman** » Wed Jun 06, 2007 12:06 pm

Beyonder wrote: That's it. I'm ordering a kill-a-watt today.

Congratulations. I had a lot of fun with mine so far.

edit: nuts. I just realized that the Seasonic Tornado has a 20-pin ATX connector. Does this mean I need a new PS for the Asus mb (which has 24-pin)?

I think you can just stick it in the mobo (there's only one way to do that). I forgot what those four extra pins were for but they're not that important.

matt_garman · Post by **matt_garman** » Wed Jun 06, 2007 12:46 pm

Beyonder wrote:edit: nuts. I just realized that the Seasonic Tornado has a 20-pin ATX connector. Does this mean I need a new PS for the Asus mb (which has 24-pin)?

Generally, no. The extra four pins are just redundant (ground, 12, 5, and 3.3 volt I think). Assuming you don't have an obscenely power hungry graphics card and/or a wildly overclocked CPU, you should be okay.

johno · Post by **johno** » Wed Jun 06, 2007 1:56 pm

The extra 4 pins are for voltages already on the 20pins for extra current capability. They are usually connected together together with the others, and so should work fine with a 20-pin supply. There is a possibility some boards may not, but I do have a M2A-VM running on a 20-pin supply.

demko · Post by **demko** » Thu Jun 07, 2007 4:35 am

Palindroman wrote:Actually, undervolting does do wonders. I recently wrote an article about it and found that especially under load there's a lot to be gained. By undervolting a Sempron 3400+ I gained 4 watts (of the 40W with C 'n Q enabled). Under load the power consumption (measured at the wall) went down from 77W to 50W. That's 35% less with a technique for which you don't have to be a computer expert. I know I am not.

This is very interesting. I guess you could consider it a wonder. Power-saving wise that is. I'm gonna get me that Hardware.info magazine and read some more about undervolting.

I suppose a "kill-a-Watt" meter is the US equivalent of the device I have, called "Energie Monitor" (German device).

I noticed some of the cheaper power measuring devices do not measure Watts, but VA. This means they do not take the power factor into account. These devices work OK when measuring a good-old light bulb, but give 'false' results when measuring for example fluorescent lights. For computers it shouldn't matter to much since PSU's of nowadays all have PFC.
But if you measure a 65W fluorescent light and wonder why the device measures 90W (actually 90VA); you probably have a VA-meter and not a Watt-meter.

The hardware I ordered (not for the server, but the business PC) should arrive here somewhere next week. I'll keep you informed.

nicke2323 · Post by **nicke2323** » Thu Jun 07, 2007 6:54 am

demko wrote: I suppose a "kill-a-Watt" meter is the US equivalent of the device I have, called "Energie Monitor" (German device).

I noticed some of the cheaper power measuring devices do not measure Watts, but VA. This means they do not take the power factor into account. These devices work OK when measuring a good-old light bulb, but give 'false' results when measuring for example fluorescent lights. For computers it shouldn't matter to much since PSU's of nowadays all have PFC.
But if you measure a 65W fluorescent light and wonder why the device measures 90W (actually 90VA); you probably have a VA-meter and not a Watt-meter.

I spent some time researching this a few weeks ago (but didn't save the links unfortunately). There seems to be a definite consensus among electronics experts that the overwhelmingly most accurate consumer devices available in Europe are the Energy Monitor and its little brother Energy Check by Voltcraft, sold by Conrad:

http://www1.conrad.de/scripts/wgate/zco ... 000A01022B
http://www1.conrad.de/scripts/wgate/zco ... 000A010221

Conrad probably has a local site for your country, so try www.conrad.xx.

demko · Post by **demko** » Sun Jun 10, 2007 6:13 am

Those two Voltcraft devices both take the Power Factor into account, which is good. But I'm not looking for such a device since I already own one that works fine. But your tip could be very usefull to others nonetheless.

Now about the undervolting again. I was wondering how much power my 'good old' (ahem) Pentium 4 3GHz (Northwood) PC was using. The PC is quite loaded with (mostly old) hardware. I don't want to bother you with summing it all up. But it used a good 182 Watts in idle mode.

Excluding the monitor that is. (An extra 52 Watts for the 20" TFT).

Of course I was wondering if I could undervolt this CPU just for the fun of it. It ran on 1.4750 Volts by default. I got the Vcore down to 1.1500 Volts without any stability problems and without adjusting the clock frequency.
The PC's power usage has dropped to 142 Watts now. I think it's quite a staggering difference (40 Watts!) for a small change which has no influence on the speed of the PC.
According to this site the maximum power usage of an Intel P4 3GHz Northwood CPU should be 81.9 Watts, so I guess 40 Watts less is a rather large saving on CPU power usage (maybe some of these savings are also caused by the mainboard power conversion circuitry).
To bad I didn't discover this earlier. It also makes you wonder why Intel didn't lower this voltage by default... they actually cost people a lot of money by not doing so. Not environmentally friendly aswell.

mcoleg · Post by **mcoleg** » Sun Jun 10, 2007 7:46 pm

johno - L2S C2D, what are they exactly?

frostedflakes · Post by **frostedflakes** » Sun Jun 10, 2007 8:28 pm

Just wanted to throw in my experiences.

The system in my sig minus the 7600GT (i.e. with integrated 6100-405 graphics) consumes 41w AC idle w/CnQ enabled. So yeah, 40w-50w idle shouldn't be very difficult at all with a decent power supply (I believe the one in my NSK-3300 is the 80Plus Seasonic SFX reviewed here on SPCR).

And Palindroman's 23w idle system seems reasonable. Assume that my system uses 33w DC idle. A 2.5" drive would probably save roughly 8w, making it 25w. Minus one stick of RAM is probably a few watts, say 3w, so 22w. Undervolting the processor further at idle would probably save a few watts, let's assume it's down to 20w. SPCR tested the PicoPSU and 80w Edac brick to be 80% efficient at 20w, which would bring AC power consumption up to 24w. I'm assuming he was testing in Europe with 240V input, which would probably slightly increase power supply efficiency over SPCR's 120V measurements.

demko · Post by **demko** » Sun Jul 01, 2007 5:41 am

I have build the business PC with the AMD Athlon64 3800+ 45W (65nm) CPU. I had some difficulties with the SATA DVD-burner, so I replaced it with an ATAPI one.
I installed Vista on it and discovered Vista has a nice option somewhere in the configuration menu that allows you to set the PC to 'low power', 'balanced' or 'performance'.
The PC consumes about 45W in balanced mode and a few less in low power mode. In performance mode the power consumption becomes somewhere around 65W.
So I didn't even need to use undervolting to get the power consumption in the region of 40 Watts.
The PC has 2GB memory and a normal 7200rpm Seagate 3,5" drive and is very quiet indeed. The complete PC uses half the power of the PC it replaces (was a 900MHz T-bird based PC).

Since I'm now convinced that building a server with a power consumption of <40W using an AMD Athlon64 CPU is very well possible I'm going to give it a try.

I want a dual-core CPU, so I was thinking about using an AMD Athlon X2 BE-2300 CPU. Also a 45W (65nm) model, but at 1900MHz and with two cores. Anyone any experience with this CPU over here?