masterlion/wue-african-datacenters

Water Efficiency Dataset for African Data Centers

Description

This dataset accompanies the paper "A Water Efficiency Dataset for African Data Centers" presented at COMPASS '25 (ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies, July 2025, Toronto, Canada).

It provides nation-level estimates of water usage efficiency (WUE) for data centers across 46 African countries spanning five distinct climate regions. The dataset covers:

• Direct (onsite) water consumption from cooling towers, computed using wet-bulb temperature via the cold-water approach model.
• Indirect (offsite) water consumption from electricity generation, weighted by each country's energy fuel mix.
• Infrastructure leakage losses based on national non-revenue water fractions.
• LLM inference water costs for four real-world scenarios using Llama-3-70B and GPT-4.

Motivation

While water consumption of data centers has gained attention in developed countries, there is a critical gap for African nations — many of which face severe water stress. This dataset enables researchers, policymakers, and practitioners to quantify and compare the water footprint of data center operations across Africa.

Source

• Paper: A Water Efficiency Dataset for African Data Centers
• Authors: Noah Shumba, Opelo Tshekiso, Pengfei Li, Giulia Fanti, Shaolei Ren
• Venue: ACM COMPASS '25

Dataset Structure

A single flat table with 184 rows (46 countries × 4 scenarios):

Scenario Columns

Column	Type	Description
country	string	Country name
scenario	string	LLM inference task (see below)
energy_wh	float	Server energy consumption (Wh)
energy_kwh	float	Server energy consumption (kWh)
wue_onsite	float	Onsite WUE for this country (L/kWh)
water_onsite_L	float	Onsite (cooling) water consumption (L)
leak_fraction	float	Water leakage fraction for this country
water_leak_L	float	Water lost to infrastructure leakage (L)
wue_offsite	float	Offsite (electricity generation) WUE (L/kWh)
pue_scenario	float	Power Usage Effectiveness (PUE)
water_offsite_L	float	Offsite water consumption (L)
water_total_L	float	Total water consumption (L) = onsite + offsite + leakage

Country-Level Columns

Column	Type	Description
wet_bulb_C	float	Average wet-bulb temperature (°C)
wet_bulb_F	float	Average wet-bulb temperature (°F)
wue_onsite_L_per_kWh	float	Baseline onsite WUE (L/kWh)
share_other_renewables	float	Electricity share: other renewables
share_bioenergy	float	Electricity share: bioenergy
share_solar	float	Electricity share: solar
share_wind	float	Electricity share: wind
share_hydro	float	Electricity share: hydro
share_nuclear	float	Electricity share: nuclear
share_oil	float	Electricity share: oil
share_gas	float	Electricity share: gas
share_coal	float	Electricity share: coal
wue_offsite_L_per_kWh	float	Baseline offsite WUE (L/kWh)
leakage_fraction	float	Country-level non-revenue water fraction
total_fuel_twh	float	Total national electricity generation (TWh)
climate_region	string	Climate region (see below)

Scenarios

Each country has 4 rows, one per LLM inference scenario:

Scenario	Model	Task	Energy (Wh)
Llama email	Llama-3-70B	Generate a short email (~100 tokens)	10.0
GPT-4 email	GPT-4	Generate a short email (~100 tokens)	232.0
Llama 10-page	Llama-3-70B	Generate a 10-page report (~5,225 tokens)	52.25
GPT-4 10-page	GPT-4	Generate a 10-page report (~5,225 tokens)	4,660.0

Climate Regions

The 46 countries span five African climate regions: Rainforest, Savanna, Desert, Steppe, and Mediterranean.

Usage

from datasets import load_dataset

ds = load_dataset("masterlion/wue-african-datacenters")
df = ds["train"].to_pandas()

# Total water per country for the Llama 10-page report
llama_report = df[df["scenario"] == "Llama 10-page"]
print(llama_report[["country", "water_total_L", "climate_region"]].to_string(index=False))

# Compare GPT-4 vs Llama water footprint
import pandas as pd
pivot = df.pivot_table(values="water_total_L", index="country", columns="scenario")
pivot["gpt4_vs_llama_report"] = pivot["GPT-4 10-page"] / pivot["Llama 10-page"]
print(pivot["gpt4_vs_llama_report"].describe())

# Water consumption by climate region
region_avg = df.groupby(["climate_region", "scenario"])["water_total_L"].mean()
print(region_avg.unstack())

Citation

If you use this dataset, please cite:

@inproceedings{shumba2025water,
  title     = {A Water Efficiency Dataset for African Data Centers},
  author    = {Shumba, Noah and Tshekiso, Opelo and Li, Pengfei and Fanti, Giulia and Ren, Shaolei},
  booktitle = {Proceedings of the ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies (COMPASS '25)},
  year      = {2025},
  publisher = {ACM},
  address   = {Toronto, ON, Canada},
  doi       = {10.1145/3715335.3735483},
  isbn      = {9798400714849}
}

License

This dataset is released under CC BY-NC-ND 4.0.