I started with calculations based on our own power consumption (servers and in-datacentre network gear) and bandwidth output. We measure all power consumption and network activity, so all I had to do was design queries for our master data base.

Sampling one of our data centre suite’s energy consumption and bandwidth output, and dividing the former by the latter, gives a figure of 2.58 Watt-hours (Wh) per Megabyte (MB). That suite is our newest, and contains a mixture of Web hosting, application hosting, storage and other services. Some servers are used inefficiently at low utilisation, and some very efficiently, notable those which are part of our Miniserver VM^® grid.

For example, if we restrict the data to that from a selection of ten of our latest Miniserver VM^® host servers (quad-core 1U Dell Power Edge servers stuffed with RAM) the figure is only 0.61 Wh/MB, but I’m trying to be conservative here and make the calculations applicable to any Web-delivered service, not just the efficiently-run ones.

We must not forget the cooling and infrastructure losses (UPS, wiring etc). Our power usage efficacy (PUE) is about 1.4, but let’s err on the side of caution and add 50%, giving a total for electricity requirements at the data centre of 3.87 Wh/MB.

My educated guess of a typical 1U server’s embedded energy is 1,000 kWh (explanation here). If we assume a 3 year lifespan (it is actually more like 4-5 years in our case) then we get 333 kWh/year. Such a server uses about 100Watts continuously, or 1,100 kWh/year, so applying that ratio (1,100:333 = 3.3:1) we get an embedded energy cost of 0.78 Wh/MB.

The network and firewall boxes are probably in the same ball park in terms of embedded energy, however compared to servers there are very few of them: 1 switch per 30 servers, 1 router-firewall pair per ~200-500 servers, 1 border router pair per 2,000-20,000 servers. Therefore that equipment’s embedded energy is not significant in our sums. The same goes for the embedded energy of the data centre itself; when compared to the thousands of servers housed within, it rapidly becomes an insignificant factor. Rather than ignore it entirely, however, I shall add 10% to the embedded server energy factor (0.86 Wh/MB total) as an allowance for the embedded energy of the supporting infrastructure.

After some conversation with ISP friends we decided that the only really significant part of the network delivery energy will be the end points. Even though some big core routers consume kilo-Watts, they are shunting hundreds of gigabits per second, thus their power requirements are spread over a huge aggregation of data flows and rendered insignificant. The end points are the Digital Subscriber Line Access Multiplexers (DSLAMs – the big modems that send an ADSL signal down your phone line), and power for the home phone line itself.

According to the EU code of conduct for broadband equipment DSLAMs with more than 100 ports should use no more than about 1.3 Watts per port when active, so a fair high estimate of their average usage would be 1Watt. The phone line is a bit more difficult to estimate, but with my rudimentary knowledge of electrical systems and telecommunications equipment I think an upper estimate of 2 Watts is reasonable. Counting the phone line, and even the DSLAM, may seem a bit pointless since they are likely to be there anyway, but we are aiming to illuminate a worst-case scenario (ie. someone who otherwise would not have the phone line).

Anyway, if counting them, then we have, say, 3 Watts continuous (2,190 Wh) which gets used for about 2GBytes per month (a low average usage estimate, but in line with many ISP’s fair usage policies), which gives another 1.09 Wh/MB.

So, to sum up:

According to Google the average Web page is about 320 KBytes these days. Therefore, downloading a single Web page uses about 1.9 Wh. Boiling the water for a mug of tea in a kettle, by comparison, uses about 50 Wh (a 3KW kettle takes 90 secs to boil 2 mugs-worth – 37.5 Wh – plus a bit for the wasted water).

A music album is in the region of 100 MBytes (a bit more for iTunes, and varies depending on quality and number of tracks of course), which would be 590 Wh. A typical small electric car uses about 300 Wh per mile, so you could drive about 2 miles on the energy required to download and deliver a music album, or do the tea round for 12 people. The energy required to deliver music on CD is vastly more, of course, and I will look at that in another article, but it is interesting to see that downloading is certainly not without its carbon impact.

While the above calculations are fun it is important to point out that equating downloading to absolute energy requirements is extremely crude and should be treated with caution. A heavy Internet user would be much more efficient, for example, and the nature of the configuration of the servers in the data centre will also have a large effect, as seen with the differences if only our virtual machine grid is used in the assessment.

I also made a number of assumptions about the home end of the network, and it is useful to see the lower-end of the range. Therefore, if we use the Miniserver VM^® figure of 1.22 Wh/MB (including cooling and embedded energy), assume that the DSLAM ports are efficient when idle (most of the time) and call that 0.5 Watt and that the phone line itself is just 1 Watt giving a total of 0.55 Wh/MB (the same as if we assume the average utilisation is double my estimate, at 4BGytes/month), then we get a total figure of only 1.8 Wh/MB as the lower end of our download energy estimate.

For instance, if we assume that Apple run an efficient data centre infrastructure for serving their iTunes customers with similar properties to our virtual machine grid (which is very likely), and that the above lower estimate for the phone line is accurate, then a 100MByte album from them would only take 180 Wh – or about 3.5 mugs of tea worth of energy.

On the flip side, I have not included the energy to run the home network and PC(s)/laptop(s), which many papers do, but that greatly skews the figures and we are assuming here that those bits of equipment are used for other purposes than just downloading.

Decoding the academic paper

First I started with what appears to be the only paper on the subject; “Energy Intensity of Computer Manufacturing: Hybrid Assessment Combining Process and Economic Input-Output Methods” by Eric Williams of the United Nations University in Japan, and published in Environmental Science & Technology in 2004.

Unfortunately the paper bundles CRT (old-style monitor) production in with the figures which really muddies the waters, especially given that they are redundant technology) However, there seems to be one very nice bit of information embedded in the paper – a table listing the electricity, fossil, and total energy use in computer production. A quick bit of analysis: The total estimated cost of production is 6,400MJ, and if we remove the CRT-specific bits, we take off:

• CRT manufacture/assembly: 255MJ
• bulk materials – CRT 800MJ
• printed circuit boards: 20MJ (est)
• electronic chemicals: 200MJ (est)
• other processes: 400MJ (est)
• Total: 1,675MJ

So, from the paper a PC’s production is about 4,700MJ, which is 1,300kWh. At a green IT conference at Oxford University last year, Fujitsu gave a great presentation on their new super-green PC fabrication plant, and asserted that their range of green PCs took 730kWh to make (materials, production & distrubution). If his numbers are right that is an impressive improvement in 4 years, but Fujitsu have been working hard in the area. Of course, that does also depend on my estimates of what proportion are down the the CRT – I shouldn’t think I’m far off though (I’m good with numbers ).

As an aside, this is very interesting from a recycling point of view. Most PC manufacturers, be it Fujitsu, Dell or IBM will proudly telling us about less than 2% goes to landfill, but if you think about it surely the only energy that can be “reclaimed” from manufacture would be the bulk materials; all the energy of making chips, assembly, PCBs, transport etc is entirely lost. Therefore, in reality one could at most hope to recover perhaps 800-1,000MJ of the original energy-cost (ie. about 20%).

A server is just a PC with a slightly different set of components (an extra disk & more RAM, but less additional cards like graphics & audio), so I think it is reasonable to assume they are similar. Therefore, I pick a figure half way between what I have deduced from the paper (1,300 kWh) and the only convincing figure I have had from a vendor (730 kWh) and have gone for 1,000 kWh in my estimations.

Sweat the desktops

So what about the fabrication energy vs. utilisation? Well, I think the paper’s 81% fab, 19% use lifetime cost is probably no longer very accurate. First, he assumes 3 hours per day, which is far too low given the number of office PCs out there and the often intensive use of family PCs. Second, I think a 3 year lifetime is too low – most people I know use their PCs much longer (they get passed down / re-used rather than thrown away) – I believe the Fujitsu figure of 6.6 years for home users at least.

I would not, however, disagree totally with his figure of 128W for PC+screen – the gains we have made in LCD screen efficiency have been outweighed by power-hungry CPU-intensive machines in recent years, although that trend is reversing. Fujitsu’s figure was 80W for their “green” PC in full power mode, and an average LCD screen uses about 20W (about half a similar CRT).

So, a quick updated estimate (based on an average of PC & home use):

120W * 5 hours/day * 365 * 5 years ~= 1,100 kWh

If we assume LCD screens are as energy intensive as CRTs and go with Eric’s figure of 1,700 kWh for production then the ratio is 61% fab : 39% use.

If we assume that Fujitsu are telling the truth though then it is 730kWh in fabrication, plus ~300kWh for a screen (a guestimate – it is about 465 kWh for a CRT), giving about 1,000kWh fabrication then the embedded vs. use energies are almost equal.

If one then does the calculation based on an office PC usage pattern and a 6.6 year lifetime, then even with more energy efficient PCs the ratio is more like 35% fab : 65% use.

Therefore, I think that we can conclude that the ratio of production energy to usage energy for a PC (with or without screen – the proportions seem about the same) range widely from something like (35% fab : 65% use) to (70% fab : 30% use), and that the main determining factor is the usage pattern of the PC, which is also the one bit of data that we probably have the worst grasp on. Either way, though, you will use less energy overall if you sweat the desktop PCs, as we discussed in the recent BCS Green IT debate.

Replace the servers

The situation is very different for a server, however. A typical modern 1U pizza-box server will use 80W when idle and 140W when working hard. Most of the time they are not straining, so call it 100W:

100W * 24 hours/day * 365 * 1.25 PUE ~= 1,100 kWh per year

In other words, a server uses about the same amount of energy as was required to create it every single year, and the same amount that a PC with a fairly average usage pattern uses in 5 years.

Because of this it is worth while to replace servers with more efficient models on a fairly regular basis. Moore’s Law (that transistor density doubles every 18 months) means that server work capacity per Watt is increasing by a factor of 4 every 3 years. This means, that provided you are using the servers properly (virtualisation etc) and consolidating onto a smaller number of newer machines, if you replace a 3 year old server its 1,000 kWh embedded energy cost will be saved by the 3 you are turning off (4:1 consolidation) in only 4 months.