WEMO24-Compiled Report-Interactive compressed (002) 4154-7882-8116 v.1 - Flipbook - Page 37
Global
Outlook
• More parsimonious, eventually targeted use of generative
AI is needed as well as the optimization of the type of
model used.
• It is already possible to optimize some core elements in the
models. With the re-usability of neural network infrastructure
choice, and the data centre location (where the energy mix is
less carbon-intensive), GHG emissions can be divided between
100 and 1,000 times.
In data centres, the objective is to reduce the PUE (Power Usage
Effectiveness) through various technologies (see WEMO 2023147)
through the development at scale of innovative cooling methods
such as immersing servers in an oil bath or direct cooling of the
chips themselves. Microsoft Azure and Google Cloud aim to
achieve a PUE below 1.125 or 1.120, which means that their
energy efficiency will be greater than 89%.
• Improvements can also be made to microprocessors and
cooling so that they consume less energy and to algorithms
so that they become more efficient and require less data and
computing power for training.
HOME
• How to master this consumption evolution?
NEXT
ARTICLE
Conclusion: Adaptation
measures
From the previous analysis one can conclude that the
world is not on the right trajectory to meet the Paris
agreement objectives. Besides mitigation measures
that were outlined above, politicians, public services
and Companies must also demonstrate realism
by launching adaptation measures to combat the
negative e昀昀ects of climate change which will be
inevitable (see WEMO 2023148).
147 https://www.capgemini.com/insights/research-library/world-energy-markets-observatory/
148 https://www.capgemini.com/insights/research-library/world-energy-markets-observatory/
37
WEMO 2024
• Research is emerging about other actionable steps that can
be taken today to align AI progress with sustainability.
For example, capping power usage during the training and
inference phases of AI models presents a promising avenue
for reducing AI energy consumption by 12% to 15%, with a
small trade-off on time to finish tasks with GPUs expected
to take around 3% longer.
