Climate Impact Of AI: Difference between revisions

From Climate Change AI Wiki
 
(9 intermediate revisions by the same user not shown)
Line 4: Line 4:
   
 
AI and machine learning has a climate impact from the energy used at all stages from development to deployment, as well as considerations around embodied energy in hardware from GPUs to data centres. The majority of emissions come from CO2 released from electricity generation powering AI processes. The carbon intensity of any particular project is highly dependent on its location and the electricity mix that it consumes.
 
AI and machine learning has a climate impact from the energy used at all stages from development to deployment, as well as considerations around embodied energy in hardware from GPUs to data centres. The majority of emissions come from CO2 released from electricity generation powering AI processes. The carbon intensity of any particular project is highly dependent on its location and the electricity mix that it consumes.
 
Efforts to quantify the CO2 impact of ML have been undertaken<ref name=":0">{{Cite journal|last=Lacoste|first=Alexandre|last2=Luccioni|first2=Alexandra|last3=Schmidt|first3=Victor |last4=Dandres|first4=Thomas|date=2019-10-21|title=Quantifying the Carbon Emissions of Machine Learning|url=https://arxiv.org/abs/1910.09700|journal=arXiv:1910.09700 [cs, stat]}}</ref><ref>{{Cite journal|last=Henderson|first=Peter|last2=Hu|first2=Jieru|last3=Romoff|first3=Joshua|last4=Brunskill|first4=Emma|last5=Jurafsky|first5=Dan|last6=Pineau|first6=Joelle|date=2020-01-31|title=Towards the Systematic Reporting of the Energy and Carbon Footprints of Machine Learning|url=http://arxiv.org/abs/2002.05651|journal=arXiv:2002.05651 [cs]}}</ref>, and tools and recommendations for best practice have been created<ref name=":1">{{Cite web|last=|first=|date=|title=ML CO2 Impact|url=https://mlco2.github.io/impact/|url-status=live|archive-url=|archive-date=|access-date=2021-03-27|website=}}</ref><ref>{{Cite web|url=https://github.com/Breakend/experiment-impact-tracker|title=Experiment Impact Tracker|date=2021-03-27|website=|url-status=live|archive-url=|archive-date=|access-date=}}</ref>.
 
   
 
== Machine Learning Application Areas ==
 
== Machine Learning Application Areas ==
   
 
===Deep Learning===
 
===Deep Learning===
 
* '''Models and training''': Many recent projects in deep learning have used increasingly large computational resources to create models. The compute resources used for the largest model training runs was shown to have doubled in size every 3.4 months from 2012-2018<ref>{{Cite web|url=https://openai.com/blog/ai-and-compute/|title=AI and Compute|date=2018-05-16|website=OpenAI|language=en|access-date=2021-03-27}}</ref>.
====Models and training====
 
 
* '''Inference''': Some estimates show the vast majority of energy (80-90%) used by deep learning models is spent on inference<ref>{{Cite web|url=https://www.forbes.com/sites/moorinsights/2019/05/09/google-cloud-doubles-down-on-nvidia-gpus-for-inference/|title=Google Cloud Doubles Down On NVIDIA GPUs For Inference|last=Strategy|first=Moor Insights and|website=Forbes|language=en|access-date=2021-03-27}}</ref><ref>{{Cite web|url=https://youtu.be/ZOIkOnW640A?t=5327|title=AWS re:Invent 2018 - Keynote with Andy Jassy|last=Jassy|first=Andy|date=|website=YouTube|url-status=live|archive-url=|archive-date=|access-date=2021-03-27}}</ref>.
Many recent projects in deep learning have used increasingly large computational resources to create models.
 
 
* '''Efficiency gains'''. Redundant training. Experiment tracking<ref>{{Cite web|url=https://towardsdatascience.com/deep-learning-and-carbon-emissions-79723d5bc86e|title=Deep Learning and Carbon Emissions|last=Biewald|first=Lukas|date=2019-06-24|website=Medium|language=en|access-date=2021-03-27}}</ref>. Smart hyperparameter selection
   
 
=== Data Centres ===
The compute resources used for the largest model training runs was shown to have doubled in size every 3.4 months from 2012-2018<ref>{{Cite web|url=https://openai.com/blog/ai-and-compute/|title=AI and Compute|date=2018-05-16|website=OpenAI|language=en|access-date=2021-03-27}}</ref>.
 
 
====Inference====
 
Some estimates show the vast majority of energy (80-90%) used by deep learning models is spent on inference<ref>{{Cite web|url=https://www.forbes.com/sites/moorinsights/2019/05/09/google-cloud-doubles-down-on-nvidia-gpus-for-inference/|title=Google Cloud Doubles Down On NVIDIA GPUs For Inference|last=Strategy|first=Moor Insights and|website=Forbes|language=en|access-date=2021-03-27}}</ref><ref>{{Cite web|url=https://youtu.be/ZOIkOnW640A?t=5327|title=AWS re:Invent 2018 - Keynote with Andy Jassy|last=Jassy|first=Andy|date=|website=YouTube|url-status=live|archive-url=|archive-date=|access-date=2021-03-27}}</ref>.
 
 
==== Efficiency gains ====
 
Redundant training
 
 
Experiment tracking<ref>{{Cite web|url=https://towardsdatascience.com/deep-learning-and-carbon-emissions-79723d5bc86e|title=Deep Learning and Carbon Emissions|last=Biewald|first=Lukas|date=2019-06-24|website=Medium|language=en|access-date=2021-03-27}}</ref>
 
 
Smart hyperparameter selection
 
 
== Data Centres ==
 
 
Data centres accounted for almost 1% of global energy demand in 2019<ref>{{Cite web|url=https://www.iea.org/reports/data-centres-and-data-transmission-networks|title=Data Centres and Data Transmission Networks – Analysis|website=IEA|language=en-GB|access-date=2021-03-27}}</ref>, at around 200TWh, and while demand increases, efficiency gains mean this may stay flat for now<ref>{{Cite journal|last=Masanet|first=Eric|last2=Shehabi|first2=Arman|last3=Lei|first3=Nuoa|last4=Smith|first4=Sarah|last5=Koomey|first5=Jonathan|date=2020-02-28|title=Recalibrating global data center energy-use estimates|url=https://science.sciencemag.org/content/367/6481/984|journal=Science|language=en|volume=367|issue=6481|pages=984–986|doi=10.1126/science.aba3758|issn=0036-8075|pmid=32108103}}</ref>. AI's current total impact can be estimated as a fraction of this, though growing extremely quickly. AI may itself offer efficiency gains for data centres by optimising control systems<ref>{{Cite web|url=https://blog.google/outreach-initiatives/environment/deepmind-ai-reduces-energy-used-for/|title=DeepMind AI reduces energy used for cooling Google data centers by 40%|date=2016-07-20|website=Google|language=en|access-date=2021-03-27}}</ref>.
 
Data centres accounted for almost 1% of global energy demand in 2019<ref>{{Cite web|url=https://www.iea.org/reports/data-centres-and-data-transmission-networks|title=Data Centres and Data Transmission Networks – Analysis|website=IEA|language=en-GB|access-date=2021-03-27}}</ref>, at around 200TWh, and while demand increases, efficiency gains mean this may stay flat for now<ref>{{Cite journal|last=Masanet|first=Eric|last2=Shehabi|first2=Arman|last3=Lei|first3=Nuoa|last4=Smith|first4=Sarah|last5=Koomey|first5=Jonathan|date=2020-02-28|title=Recalibrating global data center energy-use estimates|url=https://science.sciencemag.org/content/367/6481/984|journal=Science|language=en|volume=367|issue=6481|pages=984–986|doi=10.1126/science.aba3758|issn=0036-8075|pmid=32108103}}</ref>. AI's current total impact can be estimated as a fraction of this, though growing extremely quickly. AI may itself offer efficiency gains for data centres by optimising control systems<ref>{{Cite web|url=https://blog.google/outreach-initiatives/environment/deepmind-ai-reduces-energy-used-for/|title=DeepMind AI reduces energy used for cooling Google data centers by 40%|date=2016-07-20|website=Google|language=en|access-date=2021-03-27}}</ref>.
   
 
Estimates of electricity use of AI
 
Estimates of electricity use of AI
   
 
* '''Cloud comparison''': The major cloud computing providers, Amazon, Google and Microsoft, have varying targets and carbon intensities for their services. Google now publishes hourly estimates of the proportion of carbon-free energy (CFE) and the carbon intensity for all its cloud regions<ref>{{Cite web|url=https://cloud.google.com/sustainability/region-carbon|title=Carbon free energy for Google Cloud regions|website=Google Cloud|language=en|access-date=2021-03-27}}</ref>.
=== Cloud comparison ===
 
 
* '''Energy buying'''
The major cloud computing providers, Amazon, Google and Microsoft, have varying targets and carbon intensities for their services.
 
  +
* '''Targets''': Several Cloud Providers have published CO2 emission targets:
 
** AWS, 100% renewable energy for its data centers by 2030, on track for 2025<ref>{{Cite web|url=https://blog.aboutamazon.co.uk/sustainability/amazon-becomes-the-worlds-largest-corporate-purchaser-of-renewable-energy|title=Amazon becomes the world’s largest corporate purchaser of renewable energy|date=2020-12-10|website=UK Day One Blog|language=en|access-date=2021-03-27}}</ref>
 
** Google, 24/7 carbon-free energy (real-time matching of supply and demand, without buying renewable generation certificates) by 2030<ref>{{Cite web|url=https://cloud.google.com/blog/topics/inside-google-cloud/announcing-round-the-clock-clean-energy-for-cloud/|title=Google Cloud aims for carbon-free energy for its data centers|website=Google Cloud Blog|language=en|access-date=2021-03-27}}</ref>.
 
** Microsoft, carbon negative by 2030<ref>{{Cite web|url=https://blogs.microsoft.com/blog/2020/01/16/microsoft-will-be-carbon-negative-by-2030/|title=Microsoft will be carbon negative by 2030|date=2020-01-16|website=The Official Microsoft Blog|language=en-US|access-date=2021-03-27}}</ref>.
   
 
=== Processing Units ===
Google now publishes hourly estimates of the proportion of carbon-free energy (CFE) and the carbon intensity for all its cloud regions<ref>{{Cite web|url=https://cloud.google.com/sustainability/region-carbon|title=Carbon free energy for Google Cloud regions|website=Google Cloud|language=en|access-date=2021-03-27}}</ref>.
 
 
==== Energy buying ====
 
 
==== Targets ====
 
 
AWS, 100% renewable energy for its data centers by 2030, on track for 2025<ref>{{Cite web|url=https://blog.aboutamazon.co.uk/sustainability/amazon-becomes-the-worlds-largest-corporate-purchaser-of-renewable-energy|title=Amazon becomes the world’s largest corporate purchaser of renewable energy|date=2020-12-10|website=UK Day One Blog|language=en|access-date=2021-03-27}}</ref>
 
 
Google, 24/7 carbon-free energy (real-time matching of supply and demand, without buying renewable generation certificates) by 2030<ref>{{Cite web|url=https://cloud.google.com/blog/topics/inside-google-cloud/announcing-round-the-clock-clean-energy-for-cloud/|title=Google Cloud aims for carbon-free energy for its data centers|website=Google Cloud Blog|language=en|access-date=2021-03-27}}</ref>.
 
 
Microsoft, carbon negative by 2030<ref>{{Cite web|url=https://blogs.microsoft.com/blog/2020/01/16/microsoft-will-be-carbon-negative-by-2030/|title=Microsoft will be carbon negative by 2030|date=2020-01-16|website=The Official Microsoft Blog|language=en-US|access-date=2021-03-27}}</ref>.
 
== Processing Units ==
 
 
GPUs
 
GPUs
   
Line 50: Line 30:
   
 
== Background Readings ==
 
== Background Readings ==
 
*''' Efforts to quantify the CO2 impact of ML''' <ref name=":0">{{Cite journal|last=Lacoste|first=Alexandre|last2=Luccioni|first2=Alexandra|last3=Schmidt|first3=Victor |last4=Dandres|first4=Thomas|date=2019-10-21|title=Quantifying the Carbon Emissions of Machine Learning|url=https://arxiv.org/abs/1910.09700|journal=arXiv:1910.09700 [cs, stat]}}</ref><ref>{{Cite journal|last=Henderson|first=Peter|last2=Hu|first2=Jieru|last3=Romoff|first3=Joshua|last4=Brunskill|first4=Emma|last5=Jurafsky|first5=Dan|last6=Pineau|first6=Joelle|date=2020-01-31|title=Towards the Systematic Reporting of the Energy and Carbon Footprints of Machine Learning|url=http://arxiv.org/abs/2002.05651|journal=arXiv:2002.05651 [cs]}}</ref>.
   
 
== Online Courses and Course Materials ==
 
== Online Courses and Course Materials ==
Line 56: Line 37:
   
 
== Libraries and Tools ==
 
== Libraries and Tools ==
  +
*'''ML CO2 Impact''': A tool to calculate Machine Learning CO2 emissions, available [https://mlco2.github.io/impact/ here].
  +
*'''Experiment Impact Tracker''': Anther CO2 emissions calculator providing information about power draw from CPU and GPU, hardware information, python package versions, estimated carbon emissions information, and in California realtime carbon emission information, available [https://github.com/Breakend/experiment-impact-tracker here].
  +
*'''Microsoft Emissions Impact Dashboard''': A tool by Microsoft to track carbon emissions related to Microsoft cloud services usage, [https://www.microsoft.com/en-us/sustainability/emissions-impact-dashboard here]
  +
* '''Codecarbon''': A software package that integrates into Python codebase to estimate the amount of carbon dioxide produced by the cloud or personal computing resources used to execute the code, [https://codecarbon.io/ here]
   
 
== Data ==
 
== Data ==
  +
  +
* '''Hugging Face Model Cards''': Several model cards of trained Hugging Face models report the amount of CO2 it took to train them, [https://huggingface.co/models?other=co2_eq_emissions here].
   
 
== References ==
 
== References ==

Latest revision as of 12:44, 25 December 2021

This page is part of the Climate Change AI Wiki, which aims provide resources at the intersection of climate change and machine learning.

🌎 This article is a stub, and is currently under construction. You can help by adding to it!

AI and machine learning has a climate impact from the energy used at all stages from development to deployment, as well as considerations around embodied energy in hardware from GPUs to data centres. The majority of emissions come from CO2 released from electricity generation powering AI processes. The carbon intensity of any particular project is highly dependent on its location and the electricity mix that it consumes.

Machine Learning Application Areas[edit | edit source]

Deep Learning[edit | edit source]

  • Models and training: Many recent projects in deep learning have used increasingly large computational resources to create models. The compute resources used for the largest model training runs was shown to have doubled in size every 3.4 months from 2012-2018[1].
  • Inference: Some estimates show the vast majority of energy (80-90%) used by deep learning models is spent on inference[2][3].
  • Efficiency gains. Redundant training. Experiment tracking[4]. Smart hyperparameter selection

Data Centres[edit | edit source]

Data centres accounted for almost 1% of global energy demand in 2019[5], at around 200TWh, and while demand increases, efficiency gains mean this may stay flat for now[6]. AI's current total impact can be estimated as a fraction of this, though growing extremely quickly. AI may itself offer efficiency gains for data centres by optimising control systems[7].

Estimates of electricity use of AI

  • Cloud comparison: The major cloud computing providers, Amazon, Google and Microsoft, have varying targets and carbon intensities for their services. Google now publishes hourly estimates of the proportion of carbon-free energy (CFE) and the carbon intensity for all its cloud regions[8].
  • Energy buying
  • Targets: Several Cloud Providers have published CO2 emission targets:
    • AWS, 100% renewable energy for its data centers by 2030, on track for 2025[9]
    • Google, 24/7 carbon-free energy (real-time matching of supply and demand, without buying renewable generation certificates) by 2030[10].
    • Microsoft, carbon negative by 2030[11].

Processing Units[edit | edit source]

GPUs

Development of other chip types - TPUs, IPUs etc

Background Readings[edit | edit source]

  • Efforts to quantify the CO2 impact of ML [12][13].

Online Courses and Course Materials[edit | edit source]

Community[edit | edit source]

Libraries and Tools[edit | edit source]

  • ML CO2 Impact: A tool to calculate Machine Learning CO2 emissions, available here.
  • Experiment Impact Tracker: Anther CO2 emissions calculator providing information about power draw from CPU and GPU, hardware information, python package versions, estimated carbon emissions information, and in California realtime carbon emission information, available here.
  • Microsoft Emissions Impact Dashboard: A tool by Microsoft to track carbon emissions related to Microsoft cloud services usage, here
  • Codecarbon: A software package that integrates into Python codebase to estimate the amount of carbon dioxide produced by the cloud or personal computing resources used to execute the code, here

Data[edit | edit source]

  • Hugging Face Model Cards: Several model cards of trained Hugging Face models report the amount of CO2 it took to train them, here.

References[edit | edit source]

  1. "AI and Compute". OpenAI. 2018-05-16. Retrieved 2021-03-27.
  2. Strategy, Moor Insights and. "Google Cloud Doubles Down On NVIDIA GPUs For Inference". Forbes. Retrieved 2021-03-27.
  3. Jassy, Andy. "AWS re:Invent 2018 - Keynote with Andy Jassy". YouTube. Retrieved 2021-03-27.
  4. Biewald, Lukas (2019-06-24). "Deep Learning and Carbon Emissions". Medium. Retrieved 2021-03-27.
  5. "Data Centres and Data Transmission Networks – Analysis". IEA. Retrieved 2021-03-27.
  6. Masanet, Eric; Shehabi, Arman; Lei, Nuoa; Smith, Sarah; Koomey, Jonathan (2020-02-28). "Recalibrating global data center energy-use estimates". Science. 367 (6481): 984–986. doi:10.1126/science.aba3758. ISSN 0036-8075. PMID 32108103.
  7. "DeepMind AI reduces energy used for cooling Google data centers by 40%". Google. 2016-07-20. Retrieved 2021-03-27.
  8. "Carbon free energy for Google Cloud regions". Google Cloud. Retrieved 2021-03-27.
  9. "Amazon becomes the world's largest corporate purchaser of renewable energy". UK Day One Blog. 2020-12-10. Retrieved 2021-03-27.
  10. "Google Cloud aims for carbon-free energy for its data centers". Google Cloud Blog. Retrieved 2021-03-27.
  11. "Microsoft will be carbon negative by 2030". The Official Microsoft Blog. 2020-01-16. Retrieved 2021-03-27.
  12. Lacoste, Alexandre; Luccioni, Alexandra; Schmidt, Victor; Dandres, Thomas (2019-10-21). "Quantifying the Carbon Emissions of Machine Learning". arXiv:1910.09700 [cs, stat].
  13. Henderson, Peter; Hu, Jieru; Romoff, Joshua; Brunskill, Emma; Jurafsky, Dan; Pineau, Joelle (2020-01-31). "Towards the Systematic Reporting of the Energy and Carbon Footprints of Machine Learning". arXiv:2002.05651 [cs].