生态效率
GES:绿色效率得分 --每瓦特智能。
](https://pypi.org/project/ecoai-efficiency/)  
生态效率 使用统一的指标衡量AI的能源效率。
GES = Accuracy (%) / Energy (kWh per 1000 inferences)from ai_efficiency import measure
score = measure(your_model, test_data)
print(score)
# → Efficiency: 47,000 (accuracy per kWh)
# → Grade: A
# → Carbon: 0.8g CO2 per 1000 queries为什么?
- 监管机构来了。 欧盟人工智能法案将要求能源披露。
- 投资者在问。 ESG基金想要人工智能的碳足迹。
- 成本正在爆炸。 GPU账单很重要。效率=节省的钱。
没有标准的方法来衡量人工智能的效率。现在有了。
安装
pip install ecoai-efficiency快速开始
测量任何模型
from ai_efficiency import measure
# Works with any model that has a predict() or __call__() method
score = measure(
model=your_model,
test_data=X_test,
n_samples=1000
)
print(f"Efficiency Score: {score.efficiency:,.0f}")
print(f"Energy per 1K queries: {score.kwh_per_1k:.4f} kWh")
print(f"Carbon per 1K queries: {score.co2_per_1k:.2f}g CO2")
print(f"Grade: {score.grade}")比较型号
from ai_efficiency import compare
results = compare([model_a, model_b, model_c], test_data)
# Output:
# Model | Accuracy | kWh/1K | Efficiency | Grade
# ----------|----------|---------|------------|------
# model_a | 94.2% | 0.0100 | 9,420 | B
# model_b | 91.8% | 0.0008 | 114,750 | A+
# model_c | 96.1% | 0.0450 | 2,136 | C生成报告
from ai_efficiency import report
r = report(model, test_data)
r.save("efficiency_report.pdf") # For regulators, investors
r.save("efficiency_report.json") # For CI/CD pipelines绿色效率评分(GES)
Accuracy (or Quality Metric)
GES = ──────────────────────────────────────────────
Energy Consumption (kWh per 1000 inferences)越高越好。使用0.001千瓦时的模型准确率为95%,得分为95000。
成绩
阈值来自74个模型基准人群(NeurIPS 2026 D&B):
| 等级 | GES分数 | 百分位数 | 含义 |
|---|---|---|---|
| A+ | ≥3265200 | 前10% | 卓越的效率 |
| A | ≥1306469 | 前25% | 非常高效 |
| B | ≥512892 | 前50% | 高于中位数 |
| C | ≥187135 | 前75% | 低于中位数 |
| D | \<187135 | 底部25% | 需要优化 |
碳计算
我们使用区域电网碳强度:
from ai_efficiency import measure
# Specify your region for accurate carbon calculation
score = measure(model, data, region="KR") # South Korea: 450g CO2/kWh
score = measure(model, data, region="EU") # Europe avg: 250g CO2/kWh
score = measure(model, data, region="US-CA") # California: 200g CO2/kWh实时碳强度
获取实时电网碳数据(需要 电力地图 API密钥):
from ai_efficiency import get_carbon_intensity
# Set API key (or use ELECTRICITY_MAPS_API_KEY env var)
intensity = get_carbon_intensity("KR", real_time=True, api_key="your-key")
print(f"Current grid intensity: {intensity} gCO2/kWh")云提供商碳因素
from ai_efficiency import get_cloud_carbon_factor
# AWS, GCP, Azure carbon factors
aws_carbon = get_cloud_carbon_factor("aws", "us-east-1") # 379 gCO2/kWh
gcp_carbon = get_cloud_carbon_factor("gcp", "us-west1") # 92 gCO2/kWh (low!)SCI for AI(绿色软件基金会)
生成符合以下要求的报告 人工智能SCI 标准:
from ai_efficiency import sci_report
sci = sci_report(model, test_data, region="KR")
print(sci)
# → SCI Score: 0.0234 gCO2eq/1000 inferences
# → E (Energy): 0.000052 kWh
# → I (Carbon Intensity): 450 gCO2eq/kWh
# → M (Embodied): 0.0001 gCO2eq
# Export SCI-compliant report
print(sci.to_sci_report(model_name="MyModel"))学术论文指标
使用LaTeX导出生成纸质指标:
from ai_efficiency import measure_academic
metrics = measure_academic(
model, test_data, test_labels,
region="KR",
flops_per_sample=1.2e9 # Optional: FLOPs if known
)
# LaTeX table for papers
print(metrics.to_latex_table())
# Markdown for README
print(metrics.to_markdown_table())
# Full metrics
print(f"GES: {metrics.ges:,.0f}") # Green Efficiency Score
print(f"Energy/FLOP: {metrics.energy_per_flop:.2e}") # If FLOPs provided比较模型(学术)
from ai_efficiency import compare_academic
latex_table = compare_academic(
[model_a, model_b],
["BERT-base", "DistilBERT"],
test_data, test_labels
)
# → Generates LaTeX table sorted by efficiency支持的模型
- PyTorch:任何
nn.Module - TensorFlow/Keras:任何型号
predict() - Scikit学习:任何估算值
- 拥抱脸:变压器、扩散器
- OpenAI/人类学:基于API的模型(估计)
- 自定义:任何可调用
.predict()或__call__()
CLI工具
# Measure a saved model
ecoai-efficiency measure model.pt --data test.csv
# Compare multiple models
ecoai-efficiency compare model_a.pt model_b.pt --data test.csv
# Generate compliance report
ecoai-efficiency report model.pt --data test.csv --output report.pdfCI/CD集成
# .github/workflows/efficiency.yml
- name: Check AI Efficiency
run: |
pip install ecoai-efficiency
ecoai-efficiency check model.pt --min-grade B --fail-below C研究
该项目基于韩巴特国立大学EcoAI实验室的研究:
- 具有能源意识的机器学习
- 人工智能系统的碳足迹
- 高效推理优化
贡献
我们需要帮助:
- GPU功率测量精度
- 更多硬件配置文件(TPU、苹果硅)
- 云提供商集成(AWS、GCP、Azure)
- 区域碳强度数据
看 贡献.md.
路线图
- \[x\] 基本测量(CPU)
- \[x\] GPU功率测量(NVIDIA)
- \[x\] 实时碳强度(电力地图API)
- \[x\] 云提供商碳因素(AWS、GCP、Azure)
- \[x\] 人工智能合规SCI(绿色软件基金会)
- \[x\] LaTeX导出的学术指标
- \[x\] 碳含量计算
- \[\]过程级功率测量
- \[\]优化建议
- \[\]API认证
引用
如果您在研究中使用ecoai效率,请引用:
@inproceedings{lee2026ges,
title = {GES: A Unified Metric and Benchmark for AI Energy Efficiency},
author = {Lee, Sangkeum},
booktitle = {NeurIPS 2026 Datasets and Benchmarks Track},
year = {2026},
url = {https://github.com/ecoailab/ecoai-efficiency}
}@software{ecoai_efficiency,
title = {ecoai-efficiency: Green Efficiency Score for AI},
author = {EcoAI Lab, Hanbat National University},
year = {2025},
version = {1.0.0},
url = {https://github.com/ecoailab/ecoai-efficiency}
}许可证
麻省理工学院
团队
汉巴特国立大学生态人工智能实验室
- 主任:李教授
- 网站: ecoai.hanbat.ac.kr
______________________________________________________________________
*“最环保的人工智能是用更少的资源做更多的事情。”*
