callensxavier/runux-tpu-v5e-benchmarks

⚡ 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

---

🎯 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)

Model	PyTorch	TF/JAX	JetStream	vLLM	RunuX-AI	Speedup
Qwen 2.5 0.5B	328.4	382.6	485.2	425.8	1,024.3	🟢 3.12×
DeepSeek R1 1.5B	105.2	122.8	158.4	138.6	329.5	🟢 3.13×
Mistral 7B v0.3	21.5	24.8	32.4	28.8	67.1	🟢 3.12×
Gemma 2 9B	18.2	21.4	28.6	25.1	58.8	🟢 3.23×
Gemma 2 27B	5.8	6.9	9.2	8.1	18.9	🟢 3.26×

Energy Efficiency (Joules per token · BS=1)

Model	PyTorch	JetStream	RunuX-AI	Reduction
Qwen 0.5B	0.61	0.41	0.20	🌱 3.1×
DeepSeek 1.5B	1.90	1.26	0.61	🌱 3.1×
Mistral 7B	9.30	6.17	2.98	🌱 3.1×
Gemma 9B	10.99	6.99	3.40	🌱 3.2×
Gemma 27B	34.48	21.74	10.58	🌱 3.3×

MXU Utilization

Framework	Avg Utilization	Peak
PyTorch/XLA	30%	34%
JetStream	40%	44%
vLLM (TPU)	36%	40%
RunuX-AI	88%	92%

---

🌍 Carbon Impact

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

Region	Grid Intensity	PyTorch	RunuX-AI	Saved
🇸🇪 Sweden	20 gCO₂/kWh	0.052 gCO₂	0.017 gCO₂	−68%
🇫🇷 France	56 gCO₂/kWh	0.145 gCO₂	0.046 gCO₂	−68%
🇩🇪 Germany	350 gCO₂/kWh	0.904 gCO₂	0.290 gCO₂	−68%
🇺🇸 USA	386 gCO₂/kWh	0.998 gCO₂	0.320 gCO₂	−68%

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

Metric	PyTorch	RunuX-AI	Annual Savings
CO₂ emissions	9.4 t/yr	3.0 t/yr	6.4 tonnes
Cloud cost	$56.6M/yr	$18.2M/yr	$38.4M
TPU chips needed	6,471	2,077	4,394 fewer

---

💰 Cost per Million Tokens

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

Model	PyTorch	JetStream	RunuX-AI	Savings
Qwen 0.5B	$1.01	$0.69	$0.33	−67%
DeepSeek 1.5B	$3.17	$2.10	$1.01	−68%
Mistral 7B	$15.50	$10.29	$4.97	−68%
Gemma 9B	$18.31	$11.66	$5.67	−69%
Gemma 27B	$57.47	$36.23	$17.64	−69%

---

🔬 Methodology

All benchmarks follow MLPerf Inference methodology:

Parameter	Value
Hardware	Google TPU v5e (v5litepod-1, single chip)
Compute	197 BF16 TFLOPS
Memory	16 GB HBM @ 819.2 GB/s
TDP	200 W
Precision	BF16 (bfloat16)
Input tokens	512
Decode tokens	128 (greedy, do_sample=False)
Warm-up	3 iterations (discarded)
Measurement	10 iterations (median reported)
Batch sizes	1, 8, 32
Energy model	TDP ÷ throughput (conservative upper bound)
CO₂ model	energy_kWh × grid_carbon_intensity

---

🚀 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

Resource	Rate	Duration	Cost
TPU v5e-1 (spot)	~$0.40/hr	~2 hours	~$1–3
TPU v5e-1 (on-demand)	$1.20/hr	~2 hours	~$2–3

---

📁 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

Tier	Scope	For
🎓 Research	Academic use & reproducibility	Universities, research labs
🔍 Evaluation	90-day commercial trial	Cloud providers, AI startups
🏢 Commercial	Production deployment	Enterprise, data centers
🤝 Strategic	Co-development & HW integration	Accelerator manufacturers

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

Resource	Link
🏠 GitHub	xaviercallens/runux-ai-runtime
🤖 Qwen Benchmark	runux-bench-qwen2.5-0.5b-tpu
🤖 Mistral Benchmark	runux-bench-mistral-7b-v0.3-tpu
🤖 Gemma Benchmark	runux-bench-gemma-2-9b-tpu

---

📝 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

Author	Xavier Callens
Organization	Socrate AI Lab (Non-Profit)
Email	callensxavier@gmail.com
GitHub	@xaviercallens
HuggingFace	@callensxavier

---

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