PaperlayerView exact revision on arXiv

Source-linked AI summary

Chasing Carbon: The Elusive Environmental Footprint of Computing

Udit Gupta, Young Geun Kim, Sylvia Lee, Jordan Tse, Hsien-Hsin S. Lee, Gu-Yeon Wei, David Brooks, Carole-Jean Wu

arXiv:2011.02839v1cs.ARcs.CY

TL;DR

As computing becomes ubiquitous, its environmental impact grows despite efficiency gains and renewable energy. Using industry-reported characterization and carbon accounting across organizations and hardware life cycles, the paper finds that modern mobile and data-center emissions are dominated by hardware manufacturing and infrastructure. It therefore calls for carbon footprint to become a first-class design metric and outlines directions for reducing computing’s environmental impact.

  • Problem

    Computing’s expanding applications and ubiquity increase environmental impact, while the relative contributions of operational energy, hardware manufacturing, and infrastructure require characterization.

  • Method

    The paper uses industry-reported sustainability data, the GHG Protocol, and hardware life-cycle analysis to quantify carbon emissions for mobile and data-center systems.

  • Results

    Hardware manufacturing and infrastructure increasingly dominate carbon emissions; manufacturing accounts for over 74% of Apple’s 25 million metric tons of CO2 output, compared with 19% from product use.

  • Takeaways & Limitations

    Researchers and developers should treat carbon footprint as a first-order optimization target alongside workload and system characterization.

Abstract

from arXiv · show

Given recent algorithm, software, and hardware innovation, computing has enabled a plethora of new applications. As computing becomes increasingly ubiquitous, however, so does its environmental impact. This paper brings the issue to the attention of computer-systems researchers. Our analysis, built on industry-reported characterization, quantifies the environmental effects of computing in terms of carbon emissions. Broadly, carbon emissions have two sources: operational energy consumption, and hardware manufacturing and infrastructure. Although carbon emissions from the former are decreasing thanks to algorithmic, software, and hardware innovations that boost performance and power efficiency, the overall carbon footprint of computer systems continues to grow. This work quantifies the carbon output of computer systems to show that most emissions related to modern mobile and data-center equipment come from hardware manufacturing and infrastructure. We therefore outline future directions for minimizing the environmental impact of computing systems.

I. INTRODUCTION

Computing’s expanding energy demand and growing ubiquity have increased environmental concern, even as efficiency improvements and renewable energy reduce operational impacts. The paper argues that carbon emissions are increasingly dominated by hardware manufacturing and infrastructure rather than system operation.

  • Growing computing footprint: By 2030, ICT is projected to account for 7% of global energy demand.In 2015, even optimistic estimates placed ICT at up to 5% of global demand, with data centers alone at 1%.
  • Operational efficiency: Energy-efficiency improvements reduce operational energy consumption, while renewable energy further reduces operational carbon emissions.Specialized accelerators, warehouse-scale systems, and reduced cooling and facility overhead contribute to efficiency gains.
  • Operational efficiency: Facebook’s Prineville data center’s energy consumption increased from 2013 to 2019, but its operational carbon output reached nearly zero by 2019 after adopting renewable energy.The contrast demonstrates that energy consumption and carbon footprint can diverge.
  • Capex-related emissions: Between 2009 and 2019, mobile-device emissions shifted from primarily opex-related to primarily capex-related activities as efficiency and renewable-energy use increased.Opex covers hardware use and operational energy; capex covers infrastructure construction and chip manufacturing.
  • Future pressures: Growing application demand may widen the gap between capex- and opex-related carbon output as additional specialized hardware increases manufacturing footprints.Facebook hardware for AI training and inference grew by 4× and 3.5×, respectively, in less than two years.
  • Capex-related emissions: Hardware-manufacturing carbon emissions increased from 49% for the iPhone 3GS to 86% for the iPhone 11 over the past decade.The paper presents this shift as evidence that manufacturing and infrastructure now dominate computing’s carbon output.

II. QUANTIFYING ENVIRONMENTAL IMPACT

The paper uses established carbon-accounting practices to quantify emissions across technology organizations and hardware life cycles. It focuses on greenhouse-gas emissions and distinguishes direct, purchased-energy, and supply-chain sources.

  • Scope of analysis: The paper focuses on carbon emissions, representing total greenhouse-gas emissions, as one important environmental issue among broader environmental impacts.Other impacts mentioned include energy, water, and material consumption.
  • Industry-level accounting: The GHG Protocol categorizes organizational emissions into Scope 1 direct, Scope 2 purchased-energy, and Scope 3 upstream and downstream supply-chain emissions.The analysis builds on publicly available sustainability reports from technology companies.
  • Industry-level accounting: Scope 1 emissions include fuel combustion, refrigerants, transportation, and semiconductor-manufacturing chemicals and gases.Scope 1 is a small fraction for mobile-device vendors and data-center operators but exceeds half of operational carbon output for GlobalFoundries, Intel, and TSMC.
  • Industry-level accounting: Scope 2 emissions depend on operational energy consumption and the carbon intensity of the energy used by fabs, offices, and data centers.The passage notes that green energy can produce up to 30× fewer GHG emissions than coal or gas energy.
  • Industry-level accounting: Scope 3 includes supply-chain activities and capital goods, making hardware production and infrastructure especially important for technology companies.Accurate accounting requires considering construction, manufacturing, usage frequency, workload mix, and system lifetime.
  • Hardware life-cycle accounting: The hardware life cycle covers production, transport, use, and end-of-life processing, with opex based on use and capex aggregating the other stages.The paper applies this life-cycle framing to individual computer systems and components.

B. System-level carbon-output analysis

The paper models computer-system carbon emissions across life-cycle phases and separates operational use from infrastructure and hardware manufacturing. It uses accredited, publicly reported industry LCAs to quantify these emissions for systems and components.

  • Life-cycle accounting: Carbon emissions are organized into four life-cycle phases: production, transport, use, and end-of-life.Use includes static and dynamic power, data-center PUE overhead, and mobile battery-efficiency overhead.
  • Industry-reported LCAs: Over 98% of Apple’s total emissions come from the hardware life cycle, with manufacturing contributing 74% and hardware use 19%.Integrated-circuit manufacturing produces more carbon than hardware use.
  • Industry-reported LCAs: The analysis uses accredited, publicly reported LCAs from AMD, Apple, Google, Huawei, Intel, Microsoft, and TSMC.These sources support analysis of carbon output for computer systems.

III. ENVIRONMENTAL IMPACT OF PERSONAL COMPUTING

Using publicly reported industry data, the paper analyzes personal-computing carbon emissions and finds that hardware manufacturing dominates the environmental impact of companies such as Apple. Integrated circuits are a particularly large source of emissions.

  • Overall findings: Hardware manufacturing dominates the carbon output of personal-computing companies such as Apple.The analysis compares manufacturing with product use across personal-computing businesses.
  • Apple’s carbon footprint: Manufacturing accounts for over 74% of Apple’s 25 million metric tons of CO2 emissions, compared with 19% from product use.Manufacturing includes integrated circuits, boards and flexes, displays, electronics, steel, and assembly.
  • Apple’s carbon footprint: Integrated circuits contribute roughly 33% of Apple’s total carbon output and include CPUs, DRAMs, SoCs, and NAND flash storage.Their manufacturing emissions alone eclipse operational emissions from device energy consumption.

B. Personal-computing life-cycle analyses

Personal-computing carbon footprints vary by device platform and generation. Manufacturing dominates battery-powered devices and increasingly accounts for life-cycle emissions as performance and hardware capability improve.

  • Platform differences: Manufacturing dominates battery-powered devices, while operational energy consumption dominates always-connected devices.Manufacturing still contributes 40% for personal assistants and 50% for desktops.
  • Platform differences: Manufacturing accounts for roughly 75% of battery-powered-device emissions, while operational energy contributes approximately 20%.Always-connected devices derive most emissions from energy consumption.
  • Platform scale: An Apple MacBook’s total and manufacturing footprint is typically 3× that of an iPhone.The variation indicates that capex-related output depends on platform design and scale rather than a static overhead.
  • Generational trends: Manufacturing shares rose from 40% to 75% across iPhone 3GS-to-XR generations, from 60% to 75% for Apple Watch Series 1-to-5, and from 60% to 75% for iPad Gen 2-to-7.Across all three device families, the fraction devoted to manufacturing increased over time.
  • Generational trends: Despite declining operational carbon output across iPhone and Apple Watch generations, total carbon emissions grew steadily as manufacturing contributions increased.Higher hardware capability added flops, memory bandwidth, storage, application support, and sensors.

C. Performance and energy versus carbon footprint

The paper examines how software and hardware advances change the tradeoff between AI performance, operational energy, and manufacturing carbon. These advances improve efficiency and performance, but manufacturing emissions can dominate over realistic mobile-device lifetimes.

  • Performance and carbon tradeoffs: From 2017 to 2019, software and hardware optimizations primarily increased performance while overlooking the growth trend of carbon footprint.The performance/carbon Pareto frontier shifted right toward higher performance rather than lower carbon emissions.
  • Performance and carbon tradeoffs: The 2019 frontier includes iPhone 11 Pro at 75 images per second and 66 kg of CO2, versus Pixel 3a at 20 images per second and 45 kg of CO2.The iPhone 11 doubled iPhone X throughput from 35 to 70 images per second at slightly lower manufacturing output, from 63 to 60 kg of CO2.
  • Inference efficiency: MobileNet v2 is 17× faster than Inception v3 on a CPU, while a DSP provides an additional 3× speedup over the CPU.Algorithmic and hardware innovation increased energy efficiency by 36× and 2×, respectively.
  • Carbon amortization: MobileNet v3 requires five billion CPU inferences for operational carbon to equal manufacturing carbon, rising to 10 billion on a DSP.The DSP reduces operational footprint by 2×, while the ImageNet training set contains 14 million images.
  • Carbon amortization: MobileNet v3 requires nearly 1,200 days of continual DSP operation to match manufacturing carbon, approximately the device’s three-year expected lifetime.Algorithmic and architectural enhancements generally require inference beyond most mobile devices’ expected lifetimes to amortize manufacturing emissions.

IV. ENVIRONMENTAL IMPACT OF DATA CENTERS

As AI, autonomous driving, robotics, scientific computing, AR/VR, and other applications become ubiquitous, the paper examines data centers’ environmental impact. It highlights supply-chain emissions, including hardware manufacturing and construction, as increasingly important as renewable energy powers operations.

  • The section examines the environmental impact of data centers using industry-reported GHG Protocol data from Facebook and Google.The analysis also considers historical trends in data-center carbon emissions.
  • As data centers increasingly rely on renewable energy, carbon emissions increasingly originate from Scope 3 supply-chain sources such as hardware manufacturing and construction.

A. Breakdown of warehouse-scale data centers

For large data-center operators, supply-chain emissions dominate reported emissions, with construction and hardware manufacturing forming a substantial share of Scope 3. Evolving disclosure practices mean reported Scope 3 emissions should be treated as a lower bound.

  • Most emissions for data-center operators and cloud providers are capex-related, including construction, infrastructure, and hardware manufacturing.
  • Scope 3 emissions include supply-chain activities such as employee travel, construction, and hardware manufacturing, whereas Scope 1 and Scope 2 are opex emissions.
  • 21× higher Scope 3 emissions than Scope 2 were reported by Google in 2018: 14,000,000 metric tons of CO2 versus 684,000.
  • 23× higher Scope 3 emissions than Scope 2 were reported by Facebook in 2019: 5,800,000 metric tons of CO2 versus 252,000.
  • Up to 48% of Facebook’s 2019 Scope 3 emissions came from capital goods, including construction and hardware manufacturing.
  • Publicly reported Scope 3 carbon output should be interpreted as a lower bound because industry disclosure practices and guidelines are evolving.

B. Impact of renewable energy

Renewable energy lowers operational carbon emissions even as data-center energy consumption rises, shifting the hardware life-cycle breakdown toward manufacturing. Sustainable designs must therefore consider renewable energy alongside efficiency and capex-related choices.

  • Although overall data-center energy consumption has risen over the past five years, operational-energy carbon emissions have fallen as renewable-energy use increased.
  • Scope 2 emissions can be reported using location-based grid assumptions or market-based accounting for deliberately contracted energy sources.
  • Renewable energy reduces operational carbon emissions, causing hardware manufacturing to dominate the hardware carbon footprint.
  • For hardware using the US energy grid, hardware use contributes roughly 60% of Intel’s and 45% of AMD’s carbon emissions.
  • With solar or wind energy, over 80% of hardware life-cycle emissions come from hardware manufacturing.
  • Sustainable data-center design should consider renewable energy, efficiency-driven changes in opex emissions, and resource provisioning or leaner hardware affecting capex emissions.

V. ENVIRONMENTAL IMPACT FROM MANUFACTURING

Renewable energy can reduce carbon emissions from chip and wafer manufacturing, but manufacturing remains a major life-cycle contributor. The paper therefore keeps hardware-manufacturing emissions central to sustainable computer and workload design.

  • Hardware manufacturing comprises a large portion of emissions in both mobile and data-center systems.
  • Renewable energy will reduce hardware-manufacturing emissions, but manufacturing will remain a large portion of hardware life-cycle carbon footprints.
  • Renewable energy reduces fab carbon output by lowering the emissions associated with electricity consumption.
  • 64× lower electricity emissions from renewable energy produce only a roughly 2.7× reduction in TSMC wafer-manufacturing carbon output.
  • By 2025, TSMC estimates renewable energy will provide 20% of the electricity for forthcoming 3nm fabs.
  • Hardware manufacturing remains important to designing sustainable computers and workloads.

VI. ADDRESSING CARBON FOOTPRINT OF SYSTEMS

Reducing computing’s carbon footprint requires cross-layer optimization of both operational-energy and hardware-manufacturing emissions. The section identifies opportunities spanning algorithms, runtimes, systems, architecture, circuits, devices, manufacturing, and programming tools.

  • Operational-energy and hardware-manufacturing emissions both require optimization across the computing stack.The proposed research directions span multiple stack layers and address both opex- and capex-related carbon footprints.
  • Applications and algorithms: Reducing AI training resources lowers associated emissions by reducing processing footprint E, dataset size D, and hyperparameter search H.Reducing H also decreases the number of parallel training nodes required.
  • Systems: 2.7× higher manufacturing-related CO2 accompanies the high-performance Apple Mac Pro configuration.The configuration has more cores, memory, storage, GPU FLOPS, and GPU memory bandwidth.
  • Run-time systems: Schedulers can reduce carbon footprints by improving infrastructure efficiency and using periods when renewable energy is readily available.The analysis supports co-optimizing tail latency, throughput, power, and infrastructure-related emissions.
  • Compilers and programming languages: Compilers and programming languages can optimize directly for CO2 emissions rather than only indirectly reducing energy consumption.Existing work targets more energy-efficient code and compiler-level energy-efficiency optimizations.
  • Architecture: Architectural resource provisioning, scaling down hardware, and specialized circuits can reduce capex-related carbon emissions.Higher-performance hardware incurs higher manufacturing-related emissions, while low utilization creates a trade-off between dark silicon and manufacturing emissions.
  • Circuits; Semiconductor devices and manufacturing: Longer-lived devices, sustainable fabrication, memory reliability, and hardware consolidation can reduce production-related carbon emissions.Circuit- and device-level approaches include reliability improvements, low-carbon manufacturing, yield enhancement, and co-design for AI hardware.

VII. CONCLUSION AND FUTURE WORK

The paper concludes that hardware manufacturing increasingly dominates computing’s carbon footprint, while calling for broader collaboration and improved accounting. It also limits its focus to carbon emissions despite computing’s wider environmental effects.

  • Hardware manufacturing has increasingly dominated mobile systems’ carbon footprint over the last decade and dominates data-center footprints as renewable energy use grows.
  • Environmentally sustainable systems require collective industry and academic collaboration across design, building, and deployment.
  • Better accounting practices: Standardized accounting, disclosures, and broader organizational participation would provide further guidance for sustainable-system challenges.
  • Carbon footprint as a first-order optimization target: Researchers and developers should treat carbon footprint as a first-class design metric alongside workload and system characterization.
  • Beyond carbon emissions: The paper focuses on carbon emissions, while computing’s environmental impact also includes water consumption and use of materials such as cobalt, copper, and lithium.
Loading 2011.02839v1…