Carbon footprint – laptops vs servers, Intel vs ARM
Published in Cloud, Data Center Energy, Environment, Hardware.
Apple is one of the few companies publishing the carbon footprint of their products. These reports go all the way back to 2009 and they publish reports for new products as soon as they are released. They’re based on the ISO 14040 and 14044 international standards for lifecycle assessment, include multiple configurations to match the options available for purchase, and assume lifetimes of 4 years for macOS/tvOS devices, and 3 years for iOS, iPadOS, and watchOS devices.
For example, the new MacBook Air 13″ with the Apple Silicon M1 chip with 256GB of SSD storage has a carbon footprint of 161 kgCO2. Microsoft does the same with its Surface lineup, which I compared with Apple devices in a previous post.
Even with the new MacBooks, only 15% of the carbon footprint is in using it. Everything else is manufacturing and transport, at least until Apple reaches its net-zero goal by 2030.
For consumer devices, my conclusion is still: don’t buy a new laptop.
Apple produces the most environmentally friendly laptops from both a carbon footprint and energy efficiency perspective, but it depends on which model you compare. Comparing the devices is interesting but if there is one thing to take away from this analysis it is: don’t buy a new laptop. The majority of the environmental footprint is in the production. If your laptop is working fine then why buy a new one?The environmental impact of buying a new laptop – comparing the MacBook Air and Surface Laptop
Carbon footprint of servers
Servers are different.
Everything we do online uses servers inside data centres which use a lot of energy and water. Servers are just like laptops – they have processors, memory, disks, circuit boards and power supplies. They just don’t have a keyboard, display, or radio components like wi-fi.
Companies like Google and Amazon build their own custom servers, usually with the designs kept secret. Microsoft and Facebook are more open. At their scale it makes sense to create custom designs, but everyone else buys them from server vendors like Dell, Lenovo, and SuperMicro.
Server power consumption is very important because when you rack your own servers, you have to pay for power. As a result, power specs for each server are easy to find. However, finding carbon footprint reporting that includes the manufacturing is more difficult.
Apple is the standard to beat, but something is better than nothing. Dell is one of the few companies that publishes something. Although reporting is missing for many of its products, and the reporting is for a single product spec using a non-standards-based tool, it is still useful as an illustration.
Let’s take a Dell PowerEdge R240 1U rackmount server as an example:
The estimated carbon footprint single CPU model R240 with 16GB RAM and x2 1TB disk drives is 5,260 kgCO2e, of which 22% is in the manufacturing and 77% in use. This assumes a 4 year lifecycle.
However, there is significant uncertainty in that calculation. The standard deviation is +/- 4,860 kgCO2e, the majority of which is in the use stage.
This variation is not explained by Dell, but we can speculate that it is to do with factors such as server load, time spent idle, and maybe even the carbon intensity of the electricity grid (the report assumes the EU as the server location).
There is also a big difference across product class. The R240 is a small, cheap, 1U rackmount. If we compare it to a larger 4U 4-CPU, storage and memory dense Dell R930, the carbon footprint hits 13,300 kgCO2e +/- 14,200 kgCO2e. 85% of that is in the manufacturing but the variance is so large that this breakdown can’t be trusted.
ARM the future of low power server chips?
The R930 maximum power is reported at 1,668 W and 906 W idle. The R240 maximum power is 174 W and 64 W idle. These Dell servers all use Intel chips, like most laptops. However, things are starting to get interesting with ARM chips.
Due to their reputation for energy efficiency, ARM chips have been in mobile devices like phones for over a decade. They are now starting to appear in laptops – Apple just launched what will probably be the first mainstream device with their M1 “Apple Silicon” chip, but Microsoft’s Surface Pro X is another consumer-like device.
ARM chips are starting to come to servers, too. Amazon has its custom ARM-based Graviton chip and it claims 20% lower cost and up to 40% higher performance for M6g, C6g, and R6g instances over M5, C5, and R5 instances i.e. Graviton instances are faster and cheaper than Intel or AMD equivalents.
And for those running your own hardware, companies like Ampere are already shipping rackmount servers with custom ARM chips, such as the Ampere Altra which has a TDP of 45W for its baseline model. Unfortunately, they don’t yet publish a carbon footprint report but compare the maximum power of 45W for a 32 core ARM-based server to the 174 W of a single CPU (up to 8 cores) Intel-based server. The clock speed is lower, but if you can parallelize your workloads then ARM chips offer both performance and power advantages.
For consumers, the same rule applies: don’t buy anything new unless you have to.
For cloud users, see if you can make use of ARM-based instance types such as the AWS Graviton instances. Most applications will have no issue compiling for ARM and many dependencies probably already have ARM binaries. This will only accelerate as developers adopt the new Apple devices.
For those racking their own servers, the availability of server-class ARM rackmounts is exciting, but likely to be held back by software performance. But given usual hardware lifecycles, by the time refreshes come around the software support – accelerated by the availability of cloud instances and developer laptops running ARM – will likely have improved (see these 2017 benchmarks, likely improved by now).
With such large variances for servers, understanding true power draw is going to be key to accurately estimating carbon footprint. Manufacturing carbon intensity is a fixed, one-time cost. It might come down over time, but that only affects newly manufactured devices. In contrast, use-stage carbon footprint changes as the device continues to be used and the grid decarbonises. As usual, we need more and better data.