Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Apidays London 2024 - How to Limit Your Cloud I...

Apidays London 2024 - How to Limit Your Cloud Impact, from Bare Metal to AI by James Martin, Scaleway

How to Limit Your Cloud Impact, from Bare Metal to AI
James Martin, Head of Content & Sustainability Communications - Scaleway

apidays London 2024 - APIs for Smarter Platforms and Business Processes
September 18 & 19, 2024

------

Check out our conferences at https://www.apidays.global/

Do you want to sponsor or talk at one of our conferences?
https://apidays.typeform.com/to/ILJeAaV8

Learn more on APIscene, the global media made by the community for the community:
https://www.apiscene.io

Explore the API ecosystem with the API Landscape:
https://apilandscape.apiscene.io/

apidays

October 10, 2024
Tweet

More Decks by apidays

Other Decks in Programming

Transcript

  1. How to limit your cloud impact, from bare metal to

    AI James Martin, Head of Content & Sustainability Communications, Scaleway
  2. 3 Cloud impact vs. on-premise Migrating to the public cloud

    can reduce global carbon emissions by 59 million tCo2; the equivalent of taking 22 million cars off the road, re. Accenture. If, and only if, you move to a cloud provider that works to reduce its impact across: - Data centers - Hardware - Software
  3. HARDWARE 3% of digital emissions SOFTWARE/CODING Omnipresent digital emissions Green

    IT must cover all 3 pillars: DATA CENTERS 1% of digital emissions
  4. 6 DATA CENTERS / CLOUD - Measurement KPI Meaning Benchmarks

    Scaleway PUE Power Usage Effectiveness 1.55 (global average, source) 1.37 (best: 1.16) (2023) WUE Water Usage Effectiveness 0.18 (Meta, 2023) 0.49 (Microsoft, 2022) 0.216 (2023) 0.205 (2022)
  5. 7 DATA CENTERS / CLOUD - Measurement KPI Meaning Benchmarks

    Scaleway PUE Power Usage Effectiveness 1.55 (global average, source) 1.37 (best: 1.16) (2023) WUE Water Usage Effectiveness 0.18 (Meta, 2023) 0.49 (Microsoft, 2022) 0.216 (2023) 0.205 (2022) Low PUEs can hide high WUEs/total water consumption: • Scaleway water/DC/year: 1 113k litres (source) • Google water/DC/year: 829 million litres (source)
  6. 8 DATA CENTERS / CLOUD - Mitigation ❏ Low PUE

    and WUE (any cooling towers? 🤔) ❏ A/C alternatives used? ❏ Located in low-carbon intensity countries ❏ Maximum renewable energy ❏ Transparent & location-based reporting
  7. 10 HARDWARE - Measurement 1. Embodied Carbon = emissions of

    a productʼs entire lifespan Life Cycle Analysis, or LCA → Measure with Boavizta Manufacturer Data Repository 2. Usage impact = energy used/emissions generated: → IPMI/DCMI: measure serversʼ key stats → lm-sensors: detects all a machineʼs available sensors (including power) → RAPL: Intel processor power consumption → Scaphandre: measures serversʼ sustainability stats → Powerstat: for Linux devices with RAPL CPUs → Energizsta: open source tool by Boavizta → Manufacturer tools, e.g. Microsoft Surface Emissions Calculator COMPONENTS TOOLS
  8. HARDWARE - Mitigation x2 reconditioned Scaleway servers by 2025 x2

    maximum lifespan of Scaleway servers vs. hyperscalers
  9. 13 “Green codingˮ largely means “efficientˮ or “clean codingˮ. But…

    ❏ Managers can give developers energy-efficiency objectives (if devs can measure! ❏ Developers can: ❏ Adopt Carbon-Aware Computing (time- & location-shifting) ❏ Take Green Software for Practitioners GSF course ❏ Monitor code efficiency with SonarQube & EcoCode plugin ❏ Minimize cyclomatic complexity ❏ Avoid bloatware (if 90% of features unused; needed? ❏ Reuse existing code as far as possible ❏ Make code last as long as possible (anti-obsolescence) ❏ Ensure apps only process data they need to - no more, no less! CODE & SOFTWARE - Mitigation
  10. 14 What does your cloud provider measure? Fossil Energy Electricity

    Hardware/LCA Datacenter Network Employee impact Water WUE Azure Included Set to 0 gCo2eq – Visible kWh 1/6 from commissioning Not included Included Not Included Not included AWS Included Included Not Included Not Included Not Included Not Included Not included GCP Included Included 1/4 from commissioning Included Not Included Included Not included OVHCloud Included Included 1/5 from commissioning Included Included Included Not included Scaleway Included Included Included Included Included Included Planned Scope 1 Scope 2 Scope 3 Other Scaleway calculator V1: Bare Metal products measured, other cloud types to be added progressively
  11. AI

  12. 16 Impact of AI Training: • 552 tCO2e - GPT3.5

    (source) • 284 tCO2e - a medium-size LLM (source) • 30 tCO2e - BLOOM, a frugal LLM (source) Inference: • 100,000 tCO2e/year - GPT3.5 (source) > GenAI usage can be 200x more impactful than training < • 0.5L water: ‘cooling costʼ of a ChatGPT chat 2050 questions) (source) • ChatGPT4o = 725x more emissions than 3.5… (re. EcoLogits.ai) Hardware: GPUs vs CPUs • consume c. 4x more energy • generate 2.5x more heat (source) Global electricity consumption could double by 2026, due to AI, data centers & crypto (IEA)
  13. 17 AI Measurement & Mitigation ❏ Do I really need

    GenAI? Symbolic AI can do a lot, and requires 1000x less energy (source) ❏ Can I use an existing model? Hugging Face has nearly 500,000, can be fine-tuned ❏ Can that model be open source? Theyʼre transparent & measurable ❏ Do I have the right cloud provider? Renewable energy only + Minimal A/C & water usage ❏ Are GPUs inevitable? Ampereʼs AI inference CPUs = 35x less energy than NVIDIA GPUs ❏ Can I measure my modelsʼ impact? GenAI Impact/Ecologits.ai; Green-Coding.ai; CodeCarbon (here)...
  14. IMPACT MEASUREMENT MITIGATION Overall 4% of emissions & growing fast

    - Scopes 1-3 - Impact types Only use what you need! Cloud/ Data centers 1% of emissions & electricity - PUE - WUE - Renewable energy - Minimal A/C - Minimal water Hardware 3% of emissions - PCV - Boavizta - Manufacturer tools - Reuse - Avoid Rebound effect Code/ Software/ Web 50% of data center emissions due to code & infra, re. Intel - SonarQube - SCI - GSF Impact Framework - Eco-conception - Carbon-aware computing - Clean coding, codebase reviews - No bloatware AI - Training: 500 tC02e - Inference: x200 - Water: 0.5L/convo - Ecologits.ai - GreenCoding.ai - CodeCarbon - GenAI not inevitable - GPUs not inevitable - More AI for good