2nd/10_Wiki/Topics/AI_and_ML/Sustainability.md

---
id: wiki-2026-0508-sustainability
title: Sustainability
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Green Software, ESG, Carbon Footprint]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [sustainability, green-software, carbon-footprint, esg, ai-energy]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: codecarbon
---

# Sustainability

## 매 한 줄
> **"매 software / AI 의 carbon-aware design"**. ESG mandate (EU CSRD 2025+), AI training 의 explosive energy growth (GPT-5 ~15GWh, Claude Opus 4.7 estimates), green coding practice 의 mainstream화. 매 measure → reduce → report 의 cycle.

## 매 핵심

### 매 three pillars
- **E (Environmental)**: carbon, water, e-waste.
- **S (Social)**: labor, dataset bias, accessibility.
- **G (Governance)**: transparency, audit, compliance (CSRD, SEC climate rule).

### 매 software-specific
- **Green coding**: efficient algorithm, language choice (Rust vs Python), serverless cold-start vs warm.
- **Carbon-aware computing**: workload scheduling (run when grid is clean — Google "Carbon Intelligent Computing").
- **Energy-efficient inference**: quantization (INT8, INT4), distillation, MoE sparse routing.
- **Hardware**: ARM Graviton, Apple Silicon, NVIDIA Blackwell efficiency.

### 매 AI footprint (2026)
- **Training**: 매 single Frontier model run ~10-50 GWh.
- **Inference**: 매 GPT-5 query ~3-10 Wh (vs Google search ~0.3 Wh).
- **Aggregate**: AI 의 datacenter 가 2030 의 global electricity 의 3-7% 예상.

### 매 응용
1. CI/CD 의 carbon budget enforcement.
2. Cloud region selection (Quebec hydro vs us-east-1 mixed).
3. Model serving optimization (batch, KV cache reuse).
4. CSRD reporting (EU large company mandate).

## 💻 패턴

### codecarbon (Python tracking)
```python
from codecarbon import EmissionsTracker

tracker = EmissionsTracker(project_name="train_run")
tracker.start()
try:
    train_model()
finally:
    emissions_kg = tracker.stop()
    print(f"Run emitted {emissions_kg:.4f} kg CO2eq")
```

### Carbon-aware scheduler
```python
import requests

def grid_intensity(region: str) -> float:
    # WattTime / Electricity Maps API
    r = requests.get(f"https://api.electricitymaps.com/v3/carbon-intensity/latest?zone={region}",
                     headers={"auth-token": KEY})
    return r.json()["carbonIntensity"]  # gCO2/kWh

def best_region(regions: list[str]) -> str:
    return min(regions, key=grid_intensity)

# usage
target_region = best_region(["us-west-2", "ca-central-1", "eu-north-1"])
schedule_job(region=target_region)
```

### Quantization for inference
```python
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
import torch

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16)
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.3-70B", quantization_config=bnb)
# 4-bit quantization → ~75% memory + energy reduction vs fp16
```

### Cloud Run min-instances=0 (cold start tradeoff)
```yaml
# cloudrun.yaml — 매 idle 시 0 instance, 매 traffic 의 cold start 허용
spec:
  template:
    spec:
      containers:
        - image: gcr.io/proj/api
      containerConcurrency: 80
    metadata:
      annotations:
        autoscaling.knative.dev/minScale: "0"
```

### Carbon budget CI gate
```yaml
# .github/workflows/carbon.yml
- name: Run with codecarbon
  run: python train.py
- name: Check budget
  run: |
    EMISSIONS=$(jq -r .emissions_kg emissions.json)
    if (( $(echo "$EMISSIONS > 5.0" | bc -l) )); then
      echo "::error::Carbon budget exceeded: ${EMISSIONS}kg > 5kg"; exit 1
    fi
```

### Green model selection
```python
# 매 task 의 simplest sufficient model
from anthropic import Anthropic
client = Anthropic()

def route_query(complexity: int, query: str):
    model = "claude-haiku-4-5" if complexity < 3 else "claude-opus-4-7"
    return client.messages.create(model=model, max_tokens=1024,
                                  messages=[{"role": "user", "content": query}])
# Haiku 의 ~10-20x energy-cheaper than Opus
```

## 매 결정 기준
| 상황 | Action |
|---|---|
| 매 training large model | clean-grid region + spot + checkpoint |
| 매 inference at scale | quantize + batch + KV cache |
| 매 simple query | smallest sufficient model (Haiku, Sonnet) |
| 매 reporting mandate | codecarbon + CSRD format |
| 매 datacenter choice | Iceland, Quebec, Norway > us-east-1 |

**기본값**: 매 measure first (codecarbon) + 매 model right-size + 매 carbon-aware region.

## 🔗 Graph
- 부모: [[ESG]]
- 변형: [[Green-Software]]
- Adjacent: [[LLM_Optimization_and_Deployment_Strategies|Quantization]] · [[Mixture-of-Experts]] · [[Energy-Efficiency]]

## 🤖 LLM 활용
**언제**: 매 model selection (right-size), 매 prompt caching aggressive use (cache hit ~90% energy reduction), 매 batch API.
**언제 X**: 매 user-facing latency-critical (단, model-route hybrid 가능).

## ❌ 안티패턴
- **매 항상 Opus 사용**: 매 simple task 도 frontier model — 10-20x energy waste.
- **Cache 미사용**: 매 prompt caching 의 cache miss 가 every call → energy + cost.
- **Greenwashing**: 매 carbon offset 만 사고 actual reduction X — credibility crash.
- **Single region lock-in**: 매 dirty grid 의 stuck — multi-region 로 carbon-aware schedule.

## 🧪 검증 / 중복
- Verified (Green Software Foundation principles 2021+; Patterson et al. 2021 "Carbon Emissions and Large Neural Network Training"; EU CSRD 2024 effective; IEA 2024 datacenter report).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — ESG + AI footprint + green coding patterns |