callensxavier/runux-tpu-v5e-benchmarks

<div align="center">

⚡ RunuX-AI — TPU v5e Inference Benchmarks

Achieving 3× Throughput & 3× Energy Reduction on Google TPU v5e

   

Xavier Callens · Socrate AI Lab (Non-Profit)

Reproducible benchmark data & scripts — No proprietary code included

</div>

🎯 What Is This?

This repository contains benchmark results and Apache-2.0 reproduction scripts for comparing LLM inference performance across 5 frameworks on Google TPU v5e. The goal is to enable independent verification of our claims and foster collaboration in energy-efficient AI inference.

📄 Read the full scientific article: `scientific_article.md`

📊 Key Results at a Glance

Throughput (tokens/second · BS=1 · BF16 · 128 decode tokens)

Energy Efficiency (Joules per token · BS=1)

MXU Utilization

🌍 Carbon Impact

CO₂ per 1,000 Tokens (Mistral 7B · BS=1)

🏢 Datacenter Projection — Mistral AI Sweden (200 MW · 10B tok/day)

💰 Cost per Million Tokens

TPU v5e on-demand at $1.20/chip-hour

🔬 Methodology

All benchmarks follow MLPerf Inference methodology:

🚀 Reproduce the Baselines

Prerequisites

• —Google Cloud account with TPU v5e quota
• —Python 3.10+

Quick Start (TPU v5e)

# 1. Provision a TPU v5e
gcloud compute tpus tpu-vm create bench-vm \
  --zone=us-west4-a \
  --accelerator-type=v5litepod-1 \
  --version=tpu-ubuntu2204-base

# 2. SSH and install
gcloud compute tpus tpu-vm ssh bench-vm --zone=us-west4-a
pip install torch==2.4.0 torch_xla[tpu]==2.4.0 \
  -f https://storage.googleapis.com/libtpu-releases/index.html
pip install transformers==4.44.2 accelerate sentencepiece

# 3. Run baselines
python benchmark_baselines.py

# 4. Teardown (important!)
gcloud compute tpus tpu-vm delete bench-vm --zone=us-west4-a --quiet

Quick Start (CPU — methodology verification only)

pip install torch transformers accelerate
python benchmark_baselines.py --device cpu --models qwen2.5-0.5b

Cost Estimate

📁 Repository Structure

📦 runux-tpu-v5e-benchmarks
├── 📄 README.md                    ← You are here
├── 📄 scientific_article.md        ← Full scientific paper
├── 📊 benchmark_results.json       ← Complete results (5 models × 5 frameworks)
├── 🐍 benchmark_baselines.py       ← Reproduction script (Apache-2.0)
├── 📋 methodology.md               ← Detailed measurement protocol
└── 🌍 carbon_factors.json          ← Regional grid CO₂ intensities

🤝 Collaboration & Licensing

RunuX-AI is developed by Socrate AI Lab, a non-profit research organization.

Available License Tiers

We're Looking For Partners In

• —☁️ Cloud Providers — TPU/GPU runtime integration, joint benchmarking on next-gen hardware
• —🤖 AI Companies — Green datacenter deployment, CO₂ reduction certification
• —🔧 Hardware Makers — RISC-V edge inference, custom silicon co-design
• —🎓 Academia — Collaborative publications, internship programs

🔗 Related Resources

📝 Citation

@article{callens2026runux,
  title     = {RunuX-AI: Achieving 3× Inference Throughput and Energy
               Reduction on Google TPU v5e Through Runtime-Level Optimization},
  author    = {Callens, Xavier},
  year      = {2026},
  note      = {Socrate AI Lab, Non-Profit Research Organization},
  url       = {https://huggingface.co/datasets/callensxavier/runux-tpu-v5e-benchmarks}
}

📬 Contact

<div align="center">

© 2026 Xavier Callens / Socrate AI Lab Benchmark data & scripts: Apache-2.0 · RunuX-AI runtime: Proprietary (Patent Pending)

</div>