Score:3

Is there a way to calculate the global energy footprint of AES?

cn flag

Something that's puzzling me for a while is how we could calculate approximately the energy cost of using AES, for a company like Amazon or Google. How likely it is due to the nature of their business model that its an important part of their fixed cost? Considering key generation, encryption and decryption etc of so much data on a daily basis.

kelalaka avatar
in flag
If you really know what CPU and networking capabilities then you can model, otherwise this is an opinion-based question.
Tunnel_Vision avatar
cn flag
Yes indeed! Thank you for your input. Any case studies or research regarding this? That would be extremely helpful. I've been searching a lot but I'm still unable to find something useful.
Score:2
es flag

Back of the envelope calculation:

According to the World Bank, 2020 internet traffic was 100,000 GB/s. According to this analysis, an Intel i5-8250U CPU uses 2.2 J/GB when using AES-NI to encrypt. Since traffic needs to be both encrypted and decrypted, let's double this to get an energy requirement of 4.4 J/GB.

This works out to approximately 14 terajoules of energy per year.

The US BLS says that February 2022 US electricity costs were $0.137 per kWh.

Therefore, if all internet traffic is encrypted with hardware similar to that referenced above, the total electricity cost to encrypt and decrypt all internet traffic for a year is approximately only $530,000.

Tunnel_Vision avatar
cn flag
Hey! That's really helpful. I was wondering though. Can we really base the whole internet consumption based on an Intel i5-8250U CPU? I mean the average for 3 trillion GBs has to be more than 4.4J/GB right? Also the half-million $ electricity cost sounds a bit too small don't you think?
knaccc avatar
es flag
@Tunnel_Vision on the one hand, not all encryption will use AES-NI and be power efficient, so it will be an underestimate. On the other hand, the calculation assumes that all traffic will be encrypted, causing an over-estimate. The AES paper says that "AES consumes between 20 to 30 joule per Gigabyte, AES-NI consumes between 2.2 to 2.7 joule per Gigabyte". I'd assume data centres almost exclusively use CPUs with AES acceleration. I was surprised too that the cost is so low, but I checked the math and I can't see a mistake. The conclusion is that AES is not power-hungry compared to other tasks.
Tunnel_Vision avatar
cn flag
I see. Thank you very much. I agree with you on the math side. Now question is, who in the world would use AES without acceleration since it seems to be so slow and power-consuming. I wonder if our smartphones have that feature?
knaccc avatar
es flag
@Tunnel_Vision not all CPUs have AES-NI. However, it's only slow in comparison, and is still not a huge burden. A 45W CPU running at full power will use 162,000 joules per hour, and if part of what it was doing was encrypting 10 GB in that hour (without AES-NI), that would only be 300 joules.
Tunnel_Vision avatar
cn flag
Hey! I found this article: https://www.seagate.com/files/www-content/our-story/trends/files/idc-seagate-dataage-whitepaper.pdf Essentially it says that the storage demands and generally the Global Datasphere would hit 175 Zettabytes by 2025. Would that give us any additional clues? Is traffic the only metric to calculate the energy cost?
knaccc avatar
es flag
@Tunnel_Vision The energy cost isn't directly affected by much data is stored, only by how much of that data is decrypted and encrypted. You could include in the estimate the amount of data that is written/read from encrypted disks per day. However, this would be incredibly speculative, as there is no way to know how much a server in a data center is reading/writing to disk in encrypted form. Overall, I think it's safe to say that the cost of energy spent on encryption/decryption is a tiny rounding error on the cost of the equipment and the cost of performing other tasks on that equipment.
mangohost

Post an answer

Most people don’t grasp that asking a lot of questions unlocks learning and improves interpersonal bonding. In Alison’s studies, for example, though people could accurately recall how many questions had been asked in their conversations, they didn’t intuit the link between questions and liking. Across four studies, in which participants were engaged in conversations themselves or read transcripts of others’ conversations, people tended not to realize that question asking would influence—or had influenced—the level of amity between the conversationalists.