The cloud is inefficient, and it looks like we can improve a lot on this side. Computer Science with their systems have reached industrial scales and can be compared to build airports, highways and metro systems in terms of public infrastructure, yet, due to their immaterial and intangible nature, the perception of these systems do not match their perceived reality by the majority of the people.

While classical engineering designs physical objects, computer science designs virtual objects ~Gustavo Alonso CCA Lecture 14 May 2025 ETH Zürich

Introduction to Green Computing

Common problems in the field

• Energy consumption: the energy consumption of data-centers is huge.
  • Sometimes you cannot train a big model on a single cluster because the cable power is not strong enough to support that. A lot of the consumption nowadays is for training AI systems
    • Estimated training for Gemini Ultra is like 100 billions of petaflops (Yotta or Ronnaflops, see Introduction to Big Data) .
  • Many datacenters are going nuclear, literally: they are building datacenters powered by nuclear power (see Cumulus data atomic datacenters for a half a billion acquired by Amazon).
  • One nice thing is that IT itself will help with green transition, meaning investing in technology would help to find new ways to become more sustainable in our computations.
    • This motivates companies to invest far more in Capital Expenses (see Cloud Computing Services), in the order of billions of dollars for a single company (e.g. Amazon, Microsoft), or trillions per year globally.
• Emission Computation: we care about this value since it is a cost value not only a social concern.
  • Microsoft in 2020 or close to that, has reported to be zero emission by 2030, last year (2024) they said that was not possible anymore. This is mostly due to AI investments that made computation far more common.
• No Incentivization for green computing: For example security constraints require disks are shredded and not reused. Legacy architectures and software are not incentivized to be environmentally friendly

Overengineered systems for specific workloads

Suppose you need to transport 200 Kg of something for some distance $S$. You have a truck or a F1:

• Truck: one trip at 80 Km/h
• F1 needs 3 trips and carry only 100 Kg, to be more efficient in terms of time
  • This means that you need to be three times as fast to make the F1 worth it. This example is close to use too complicated systems to solve easy workloads. Like using hadoop to process a single small text file would be overkill, there has been actually a paper, see (Rowstron et al. 2012) (below 100GB it’s faster to use a single machine).

Scaling and Performance tradeoffs

There are some tradeoff between scaling and performance measurements. Frank Mcsherry showed that a single threaded well implemented algorithm was much faster for processing the same dataset with a same algorithm.

TODO: put frank image or GP - Speed up comparison image.

This raises questions in comparing complex systems with efficient ones, and possible trade-offs between these values. The problem is that to have a small gain in performance, it needs a lot of more resources in the order of like 4 times of speed up for 20 machines.

Externality

An externality is a side effect—positive or negative—that spills over onto third parties. The key idea is non-compensation: those affected aren’t paid for the cost they bear (in negative cases) or don’t pay for the benefit they receive (in positive cases).

These impose costs on others.

• Example 1: Pollution
  A factory emits pollution into the air. The local residents suffer health issues, but the factory doesn’t compensate them. The cost of pollution isn’t in the price of the factory’s products. One example is the carbon footprint for computing, which is a kind of externality in terms of pollution.
• Example 2: Traffic congestion
  One more driver on the road slows everyone else down. Each driver doesn’t pay for the added congestion they cause.

These create benefits for others.

• Example 1: Vaccination
  If you get vaccinated, you reduce disease spread, benefiting people around you—even those who didn’t vaccinate.
• Example 2: Education
  An educated population leads to higher civic engagement and innovation, benefiting society more broadly.

Markets fail to allocate resources efficiently when externalities exist, because the social cost or benefit differs from the private cost or benefit. This is what economists call a market failure.

• Taxes on negative externalities (e.g., carbon tax)
• Subsidies for positive externalities (e.g., grants for solar panels)
• Regulation (e.g., pollution limits)
• Tradable permits (e.g., cap-and-trade systems)

An externality is an unintended side effect of an action that affects others without being priced in the market, leading to potential inefficiencies or unfair outcomes.

Green IT

All big industries need to think about sustainability. IT as one of those big industries need to think about it too. One main problem to solve is to reduce the use of resources. As a society we need to find a new way to think about the computing, with great effects on how this industry operates.

Energy Emissions

There are three possible scopes for emissions:

• Direct emissions (Scope 1) because of economical activity (heating your house)
• Indirect emissions (energy): (Scope 2) Buying energy from suppliers (caused by electricity consumed)
• Indirect emissions (other): (Scope 3) related to whole activity (e.g. production of the hardware)

The above terms are defined. While we also use the term free emissions for emission allowances that are allocated for free.

Environmental impact of common production

In the case of Apple’s mobile phones manifacturing the integrated circuits et cetera is like 74% of the emission and 19% is the hardware use account. The emission of the use is very very small. Also the consume of water for chip production is high: Taiwan had a discussion between Chipmakers and farmers, something like 63m tons of water in 2019. And this water is highly polluted and contamined. This is a free emission which has a big impact on the environment.

References

[1] Rowstron et al. “Nobody Ever Got Fired for Using Hadoop on a Cluster” Association for Computing Machinery 2012