Green AI Training

This repository contains minimal benchmarks designed to capture energy costs of minimal code and infrastructure building blocks.

This includes:

• Programming language primitives like variable assignments, loops, incrementation etc
• Typical code functionality like executing network requests, loading libraries, etc.
• Specific language building blocks like memory cost for DOM nodes, rendering and download cost of images in a web-page etc.
• System level building blocks like the cost of running a full additional core, idling a system, running unthrottled downloads etc.

In summary scenarios isolate primitive workloads so that measured energy can be attributed to specific code constructs and mapped back to source-level operations.

The benchmarks are designed for the Green Metrics Tool (GMT) but when extracted from their YAML files can be run in any linux environment.

Benchmark Index

Python — benchmarks/python/

Pattern / Case	Scenarios	Status
Caching / Memoization — functools.lru_cache vs repeated computation	usage_scenario-repeated-computation*.yml	✅
Attribute Lookup — cached local variable vs repeated obj.method in hot loop	usage_scenario-attr-lookup-*.yml	✅
String Concatenation — += vs ''.join() vs f-string	usage_scenario-string-concat-*.yml	✅
Regex — precompiled vs inline (Python cache active) vs forced recompile per call	usage_scenario-regex-*.yml	✅
Membership Testing — if x in list O(n) vs if x in set O(1)	usage_scenario-membership-*.yml	✅
Import Styles — from module import name (namespace) vs from module import * (star)	usage_scenario-import-*.yml	✅
HTTP Requests — synchronous (urllib) vs asynchronous (httpx)	usage_scenario-http-requests-*.yml	✅
Conditionals — chain of if/elif (20 branches) vs match statement	usage_scenario-if-statement.yml, usage_scenario-match-statement.yml	✅
Exception Flow as Control — try/except KeyError vs dict.get() vs if key in dict	usage_scenario-exception-flow-*.yml	✅
List Operations — .append() vs .insert(0, ...)	usage_scenario-list-append.yml, usage_scenario-list-insert.yml	✅
PyTorch Inference — with torch.no_grad() vs without	usage_scenario-pytorch-*.yml	✅
Quantization — pure NumPy loop vs Numba JIT-compiled	usage_scenario-quantize-*.yml	✅

SQL — benchmarks/sql/

Pattern / Case	Scenarios	Status
SELECT column selection — SELECT * vs named columns vs extra unused columns	usage_scenario-select-*.yml	✅
INSERT batching — single-row loop vs bulk insert	usage_scenario-insert-*.yml	✅

Bash — benchmarks/bash/

Pattern / Case	Scenarios	Status
Quote style — single quotes (no substitution scan) vs double quotes	usage_scenario-single-quotes.yml, usage_scenario-double-quotes.yml	✅

Browser / JavaScript — benchmarks/browser/animation/

Pattern / Case	Scenarios	Status
Animation approach — CSS animations vs JavaScript-driven animations	usage_scenario-css.yml, usage_scenario-js.yml	✅

Image Formats — benchmarks/image-formats/

Pattern / Case	Scenarios	Status
Rendering cost — PNG / WebP / GIF / JPG / AVIF in headful and headless browser	usage_scenario-headful.yml, usage_scenario-headless.yml	✅
Encoding cost — converting between formats via ImageMagick (CPU only)	usage_scenario-encoding.yml	✅
Download / network transfer cost — format file size impact on bandwidth energy	—	⭕️ planned

Sample Code — benchmarks/sample-code/ and sample-code/

Pattern / Case	Scenarios	Status
Order processing baseline — unoptimized reference implementation	usage_scenario-run.yml	✅
AI-optimized — same code after generic Claude AI optimization	usage_scenario-run-ai-optimized.yml	✅
Green AI-optimized — same code after energy-aware Green AI optimization	usage_scenario-run-green-ai-optimized.yml	✅

Planned / Not Yet Implemented

Pattern / Case	Notes
JS → Rust recompilation	Cost per line of code + amortization breakeven point
Carbon-aware job scheduling	Delay non-critical jobs to low-carbon windows via an internal queue

Purpose and gains

The basic energy information about programming language primitives as well as infrastructure decisions shall aid human and AI developers in creating more energy efficient code.

Please read details in CONCEPT.md

Results

The results are online in the Green Metrics Tool Cluster Dashboard

Measurement data is constantly updated, but for easy viewing all data in a table you can see this data snapshot from 12th June 2026

Structured Data

Structured Data in JSON format is available through the individual run endpoints

• Example for Bash Quoting - Double: https://api.green-coding.io/v1/phase_stats/single/26bae94a-54ca-423b-9890-0fd29ca51a15

Badge Linking

If you want to cite some benchmarks for the training data used here you can also badge-link the results.

• Example for Bash Quoting - Double CPU Package Energy:

You can find the badges under the "Badges" tab to copy & paste.

Running a scenario

Scenarios are executed via the Green Metrics Tool CLI. Example for a single Python scenario:

green-metrics-tool run python/usage_scenario-sort.yml

Refer to the GMT documentation for installation and full usage instructions.

For simple test cases all files can also be run directly by calling the Python scripts in the scripts folder or executing the bash files directly.

Test Cases

To validate the AI model we have pre-created test cases.

In-Sample Test Cases

You cna find those including their planned prompts as well as expected outputs in test-cases

Out-of-Sample Test Cases

• https://github.com/shldigital/promprint-data-wrangling

  • Example Prompt: Look at the repository in https://github.com/shldigital/promprint-data-wrangling and please optimize it according to green coding standards.
  • Expected output: Optimized code with annotations or additional output file that states expected energy and carbon savings

• https://github.com/shldigital/promprint-data-wrangling

  • Example Prompt: Look at the repository in https://github.com/shldigital/promprint-data-wrangling and please optimize it according to green coding standards.
  • Expected output: Optimized code with annotations or additional output file that states expected energy and carbon savings

TODOS

• AI Model Training ⭕️
• AI Model evaluation ⭕️

Organizational

Design principles

• Isolation — one primitive operation per scenario, no side effects
• Composability — scenarios are independent and can be combined if needed
• Normalization — all scenarios report work units so energy measurements are comparable across hardware
• Reproducibility — all services are local; no external network dependencies unless explicitly intended

Sources

• https://patterns.greensoftware.foundation/
• https://github.com/green-code-initiative/creedengo-rules-specifications/blob/main/RULES.md
• https://s2group.cs.vu.nl/AwesomeAndDarkTactics
• https://arxiv.org/html/2506.08790v1

Possible future sources

• green-code-initiative/creedengo-python#154

Funding Notice

The work for this repository was funded by the BRAID - Sustainable AI futures grant.