Independent AI Researcher | Energy Efficiency & Sustainable Computing

Paper (Draft) | Dashboard | Batch Size Analysis | Metadata

Breakthrough Finding: Discovered that bitsandbytes INT8 increases energy by 17-147% due to mixed-precision decomposition. Causal diagnosis via ablation recovered +79-98% throughput and −35-41% energy across consumer (RTX 4090D) and datacenter (A800) GPUs.

NEW — Batch Size Optimization: A800 sweep (BS 1→64) shows 95.7% energy reduction and 55.5× throughput scaling. BS=1 wastes 55% GPU capacity. Interactive results: View Dashboard →

Research Scope: This work focuses on energy efficiency diagnosis. Accuracy assessment (perplexity, downstream tasks) is not yet complete. Pure INT8 (threshold=0.0) shows major performance gains, but accuracy impact requires validation. Next steps: PPL and MMLU evaluation—contributions welcome!

Key Contributions:

• 🎯 Root cause identified: Mixed-precision decomposition, not INT8 itself
• 📊 93+ measurements (CV < 1-2%) across 3 GPU architectures: RTX 5090 (Blackwell), RTX 4090D (Ada Lovelace), A800 (Ampere)
• 📈 Batch size scaling law: 95.7% energy reduction (BS=1→64), validated on A800 with 70 measurements
• ✅ Cross-platform validation: Consistent results across consumer & datacenter GPUs, multiple models
• 🔓 Full reproducibility: Complete metadata with software versions, configs, and protocols
• 🌐 Open data: All raw data, scripts, interactive dashboard, and provenance publicly available

Impact: Prevents industry from drawing wrong conclusions about INT8 quantization. Provides actionable guidance for practitioners deploying quantized LLMs in production.

Status: Preparing for arXiv submission. bitsandbytes Issues filed: #1851 (NF4) | #1867 (INT8 Energy)

All my benchmarks follow rigorous reproducibility standards:

• ✅ Complete metadata with hardware specs, software versions, and model commits
• ✅ Statistical rigor (n=10, CV < 2%, significance tests)
• ✅ Open data with full provenance and reproducible scripts
• ✅ Causal analysis via controlled experiments and ablations
• ✅ Cross-architecture validation (Blackwell + Ada Lovelace + Ampere)

📁 View Metadata Standards →

Languages: Python, TypeScript, Bash
ML/AI: PyTorch, Transformers, bitsandbytes, CUDA
Data: Pandas, NumPy, SciPy, Matplotlib
Tools: Git, Docker, Jupyter, VS Code
Cloud: AutoDL, AWS (occasional)

Interactive dashboard for comparing AI models by accuracy, cost, and carbon footprint. Features a Batch Size Analysis page with interactive charts and cost calculator.

Tech: TypeScript, React, Recharts, TailwindCSS, GitHub Pages
Data: 93+ measurements, 8 models, 3 GPU architectures (RTX 5090, RTX 4090D, A800)

Systematic study of quantization energy efficiency on modern GPUs. Discovered two paradoxes (NF4 and bitsandbytes INT8) and provided causal diagnosis via ablation. Latest: batch size sweep reveals 95.7% energy reduction potential.

Tech: Python, PyTorch, NVML, bitsandbytes
Impact: Prevents ~30-96% energy waste in production LLM deployments

• Email: zhanghongping1982@gmail.com
• GitHub: @hongping-zh
• Dashboard: ecocompute-dynamic-eval

I believe in open science and reproducible research. Every benchmark I publish includes:

• Complete metadata (hardware, software, models)
• Raw data and analysis scripts
• Reproduction commands
• Known issues and resolutions

Goal: Make AI research more transparent, reproducible, and energy-efficient.

"Measure, don't assume. Reproduce, don't trust. Share, don't hoard."