← 論文一覧に戻る

ハイパースケールAI学習のためのFinOpsガバナンスと炭素認識型ワークロードスケジューリング

FinOps Governance and Carbon-Aware Workload Scheduling for Hyperscale AI Training (原題)

Singh, Sonu Kumar

Zenodoプレプリント2026-11-18#AI×ESGOrigin: US経営インパクト: コスト削減対象セクター: cross_sector
DOI: 10.5281/zenodo.24357477
原典: https://zenodo.org/records/24357477
📄 PDF

🤖 gxceed AI 要約

日本語

GPUクラウド上でのAI学習において、コスト管理(FinOps)と炭素削減(GreenOps)を統合したKubernetesスケジューリング基盤を提案。FOCUS 1.2やISO/IEC 21031 SCI-for-AI、EU CSRDのScope2/3開示に対応した計測を組み込み、WattTimeやElectricity Mapsの炭素強度データでジョブを低炭素時間帯・スポット価格帯へ振り分ける。90日・44万GPU時間の実証でGPU費用46.8%削減、Scope2炭素強度38.4%低下、利用率89.4%を報告。

English

Proposes a unified FinOps and carbon-aware (GreenOps) Kubernetes scheduling architecture for hyperscale AI training. It aligns unit-cost and carbon telemetry with FOCUS 1.2, ISO/IEC 21031 SCI-for-AI, and EU CSRD Scope 2/3 disclosure, using WattTime MOER and Electricity Maps forecasts to shift jobs to low-carbon, spot-priced windows. A 90-day, 441,600 GPU-hour study reports 46.8% lower GPU spend, 38.4% lower Scope 2 carbon intensity, and 89.4% utilization.

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

📝 gxceed 編集解説 — Why this matters

日本のGX文脈において

AI学習の電力消費と炭素排出を開示対象として扱う点は、SSBJや有報でのScope2/3開示、データセンターの脱炭素を迫られる日本企業にとって実務的示唆が大きい。GPU調達と炭素会計を同時に最適化する手法は、国内クラウド・製造業のGX投資判断に応用可能。

In the global GX context

Directly engages CSRD Scope 2/3 disclosure and the emerging SCI-for-AI (ISO/IEC 21031) standard, extending climate disclosure scholarship into AI compute infrastructure. It offers a concrete operational bridge between carbon accounting standards and cloud cost governance, a gap in current ISSB/CSRD implementation literature.

👥 読者別の含意

🔬研究者:AI計算の炭素会計と開示基準の接続を定量実証した事例として、Scope3・データセンター脱炭素研究の参照点になる。

🏢実務担当者:クラウド費用と炭素排出を同時に削減するスケジューリング設計と、CSRD対応の計測指標の実装例として活用できる。

🏛政策担当者:AIインフラの電力・炭素開示を制度設計に組み込む際、SCI-for-AIや炭素強度データ連携の実務的論点を提供する。

📄 Abstract(原文)

Version 2.0.0 (Updated November 2026 — Plain-English & Late-2026 Standards Edition) Training foundation models and enterprise adapters across hyperscale GPU clusters consumes megawatt-hours of electricity and millions in monthly cloud bills. Historically, MLOps teams treated cost governance and sustainability as separate post-hoc reports, launching delay-tolerant batch jobs immediately on expensive on-demand GPUs during high-carbon evening grid peaks rather than exploiting 60–72% spot discounts and low-carbon solar/hydro energy windows. This revised Version 2.0.0 architectural white paper presents a unified FinOps and Carbon-Aware (GreenOps) Kubernetes scheduling architecture updated with verified 2025–2026 sustainability and cloud-cost standards: FOCUS 1.2 & GSF SCI for AI (ISO/IEC 21031) + EU CSRD Compliance: Upgrades unit-cost telemetry to FinOps Foundation FOCUS 1.2 (tracking USD per billion training tokens and kgCO2e per run alongside OpenCost, Kepler eBPF, and NVIDIA DCGM) and aligns operational + embodied carbon accounting with ISO/IEC 21031 SCI-for-AI and mandatory EU Corporate Sustainability Reporting Directive (CSRD) Scope 2/3 disclosure audits. Kubernetes Kueue & NVIDIA Slinky Gang Scheduling + Grid APIs: Integrates 5-minute WattTime Marginal Operating Emissions Rate (MOER) telemetry, 24-hour Electricity Maps forecasts, and per-SKU Weibull spot preemption hazard models into Kueue AdmissionChecks classifying jobs into Strict-SLA (<1h Reserved), Elastic-Batch (1–12h temporal+spot), and Opportunistic (>12h cross-region gated by GDPR residency and USD 0.02/GB WAN break-even thresholds). 1.15-Second Two-Phase PyTorch DCP Copy-on-Write Checkpointing: Eliminates 45–120 s blocking torch.save stalls by snapshotting FSDP shards to pinned DDR5 host RAM in 1.15 s and flushing asynchronously to regional object storage in under 14 s on IMDS preemption warnings, governed by Young-Daly optimal checkpoint intervals (12 min, 1.3% overhead) and Zeus GPU DVFS power-capping (700 W to 500 W) for smooth carbon throttling. Controlled 90-Day, 441,600 GPU-Hour Multi-Region Study: Separates verified industrial patterns (Google Carbon-Intelligent Computing, Azure Carbon-Aware Kubernetes, Spotify Cost Insights) from our controlled 90-day study across US-East, US-West-2, and EU-North (1,240 jobs, 18,400 GPU-days): cuts total GPU spend 46.8% (USD 1,704,000 to USD 906,500; USD 797,500 saved), lowers Scope 2 carbon intensity 38.4% (385.0 to 237.2 gCO2eq/GPU-hr; 65.3 tCO2e avoided), raises effective GPU utilization from 61.5% to 89.4%, cuts spot eviction waste 88.5% (18.2% to 2.1%), and holds 99.4% on-time deadline completion. What Changed in Version 2.0.0: Rewrote v1.0.0 in crisp, plain-English Lead Architect prose; upgraded FOCUS 1.0 to FOCUS 1.2 and integrated GSF SCI for AI (ISO/IEC 21031), EU CSRD Scope 2/3 compliance, Kueue + NVIDIA Slinky gang scheduling, and WattTime/Electricity Maps telemetry; removed repeated copy-pasted template filler blocks across Pages 3–14 (including inside Conclusion), broken numbering, equation/code typos, and mismatched citations from v1.0.0 Sections 1.3 and 8; verified 26 1-to-1 APA 7th Edition citations (cross-citing companion Zenodo FinOps and sovereign IaC papers); added formal citation guidance; and linked author ORCID (0009-0002-9180-4946) and affiliation (Capgemini US LLC). Recommended Citation (APA 7th Edition): Singh, S. K. (2026). FinOps Governance and Carbon-Aware Workload Scheduling for Hyperscale AI Training (Version 2.0.0). Zenodo. https://doi.org/10.5281/zenodo.23513209 Author: Sonu Kumar Singh, Member, IEEE (#102728576) — Senior Consultant & Lead Cloud/AI Architect, Capgemini US LLC, Chicago, IL, USA | ORCID: 0009-0002-9180-4946. Independent technical publication; views expressed are solely the author's.

🔗 Provenance — このレコードを発見したソース

🔔 こうした論文の新着を逃したくない方は キーワードアラート に登録(無料・3キーワードまで)。

gxceed は公開メタデータに基づく研究支援データセットです。要約・翻訳・解説は AI 支援で生成されています。 最終的な解釈・検証は利用者が原典資料に基づいて行うことを前提とします。