What progress costs us › PAGE online

What is the amount of the CO₂-Footprint of AI, blockchains and smart home applications?

• AI, blockchains, smart home, automobility – is it really that easy to calculate the resulting energy consumption? Is there an exact total? No! Nevertheless, we have researched figures that provide a little information.

However, accurate estimates are complicated because complex service infrastructures always result in measurement variability. For example, the carbon footprint of video streaming varies depending on the providers involved (optical fiber is much more energy-efficient), data centers (some feed waste heat back into district heating) and end devices (the older and larger the screen, the more expensive).

Digitization per year per capita

According to Study by the KfW It is estimated that digital applications in Germany caused around 34 million tons of CO in 2020₂-Emission. The digital footprint per person and year was around 740 kilograms of carbon dioxide with average private use of digital devices. With intensive use, this value increases to around one tonne of CO₂ per year and head.

Streaming is better than Blu-ray

Video consumption made up a large part of traffic even before the corona pandemic, even if you look at the numbers (Netflix around 1 kilogram of CO₂ per hour) can argue in detail. Michigan University engineers concluded in 2019 that streaming a movie is still less polluting than buying and watching a Blu-ray disc.

Bitcoin, Blockchain and Co: Turbo Boost into the climate catastrophe?

According to a study by the Linux Foundation in cooperation with the Hyperledger Foundation, the 178,040 gigawatt hours of the Bitcoin blockchain are responsible for the emission of around 114 megatons of carbon dioxide per year. That is about as much as the entire Czech Republic consumes. Together with Ethereum (2021: 108.390 gigawatt hours), it was until recently responsible for around 0.36 percent of man-made CO₂-Emission responsible.

However, in August 2022, Ethereum switched from the computationally intensive proof-of-work to the proof-of-stake process and, according to its own statements, now consumes 99.95 percent less energy.

Ethereum last had its consumption estimated in September. The Germany-based Crypto Carbon Ratings Institute (CCRI) found that the network now consumes just 2.6 megawatt hours annually, equivalent to emissions of one megaton of CO₂ is equivalent to.

As I said, exact estimates are hardly possible, and the providers’ own statements should be treated with caution. Here is a comparison from the Deloitte Canada study: Solana 11.05, Polygon 0.79 and Flow 0.18 megawatt hours per year.

AI, Machine Learning and Big Data – Smart Pollution?

Humanity is already benefiting enormously from AI applications – for example in medical diagnostics, but also in climate protection and making energy supply more efficient. In addition to time, money and high-quality data, training an AI model requires a lot of energy.

The more parameters a model has, the more energy-intensive the calculations are – at least that’s what you can say, but the different models also use different amounts of electricity.

Thousands of Petaflops

For example, OpenAI’s GPT-3 natural language processing program is what is known as an autoregressive Transformer Language model. In 2020 it consisted of 175 billion parameters, the previous model GPT-2 had only 1.5 billion parameters when it was launched a year earlier.

While GPT-2 took a few tens of petaflop days to train, GPT-3 required several thousand — a massive amount of computation and power. The unit flops or floating point operations per second describes the processing speed of computers, one petaflop corresponds to one quadrillion floating point operations per second.

Long explanation for a total of 936 megawatt hours of electricity or 400 megatons of CO₂which only costs the training, not the operation of the AI.

Green Hope IoT – Smart Homes and Autonomous Vehicles

In order to calculate the additional costs for the smart home, you have to measure the power consumption of the devices in stand-by mode in addition to the new hardware to be procured. In fact, depending on the manufacturer – for example smart speakers – these often differ greatly from each other.

A modern Apple HomePod mini on stand-by only consumes 0.6 watts, 5.42 kilowatt hours per year, i.e. almost 2.3 kilograms of CO₂. An old Amazon Echo consumes around five times as much electricity per year and thus causes over 10 kilograms of greenhouse gases.

Users of a complete smart home get 250 to 500 kilowatt hours a year for stand-by operation; in CO₂ That’s around 100 to 200 kilograms. Of course, this can save energy on another level, such as heating or lighting.

Car refuels Internet

Many people also hope that autonomous vehicles will lead to a green turnaround in traffic, but according to a current Fraunhofer study, this hope can quickly be reversed.

A critical factor is data exchange. In test vehicles, the volume was between 1.4 and 19 terabytes per hour. As soon as more than 0.8 TB/h were transferred, the efficiency gain compared to conventional vehicles would be used up.