WEMO24-Compiled Report-Interactive compressed (002) 4154-7882-8116 v.1 - Flipbook - Page 36
Global
Outlook
• Strong forecasting ability: For example, weather forecasting:
Google DeepMind Graph Cast model delivers 10-day weather
predictions at unprecedented accuracy in under one
minute. This helps to better forecast sun and wind power
generation and thus to better balance the grid, with much
lower processing power than currently used by weather
super calculators.
A second example is provided by the city of Zurich. By
integrating AI into the management of its public lighting and
thanks to traffic forecasts, it managed to reduce its energy
consumption by 70%, by switching lightening on only when
it was needed.
• Huge consumptions: As AI makes it possible to analyse,
correlate, and generate a lot of data, it can improve complex
situations modelling.
Chat GPT 3 training would have emitted around 502 tonnes
of CO2, which is the equivalent of what 56 French people emit
for a year. Each single query generates, only for computation,
around 5g of CO2.
More generally, Microsoft increased its greenhouse gas
emissions by 30% between 2020 and 2023 due to 16
infrastructures for artificial intelligence (which were not
always installed near decarbonized electricity sources).
146 https://deepmind.google/discover/blog/
deepmind-ai-reduces-google-data-centre-cooling-bill-by-40/
If AI keeps its current trend, this consumption should
grow significantly.
The energy required to run AI tasks is already accelerating
with an annual growth rate between 26% and 36%. This
means that by 2028, AI could be using more power than the
entire country of Iceland used in 2021!
The AI lifecycle impacts the environment in two key stages:
the training phase and the inference phase. In the training
phase, models learn and develop by digesting vast amounts
of data. This training phase is very intensive, and usually
performed several times to adjust the meta-parameters of the
algorithms. Once trained, they step into the inference phase,
where those models are applied to solve real-world problems.
The consumption of each request is much smaller (equivalent
to a few grams of CO2), but the number of requests can be
massive.
At present, the environmental footprint is split, with training
responsible for about 20% and inference taking up the lion's
share at 80% for large public systems. As generic AI models
gain traction across diverse sectors, the need for inference
and its environmental footprint will escalate.
WEMO 2024
On the negative side:
Today data centres consume around 1% of global electricity,
or 194 TWh. So far, the increase in the number of calculations
has been nearly offset by gains in energy efficiency. As a
result, data centre’s consumption increased by only 6%
between 2010 and 2020.
36
• Energy consumption reduction: At Google, reducing energy
usage has been a major focus over the past 10 years. It has
built its own superefficient servers, invented more efficient
ways to cool its data centres and invested heavily in green
energy sources.
In five years, it improved by 3.5 times the computing power
out of the same amount of energy.
A breakthrough was achieved by applying DeepMind’s
machine learning to Google data centre. It resulted in lowering
the amount of energy used for cooling by up to 40% .146
HOME
NEXT
ARTICLE
