Welcome to ecocompute-dynamic-eval Discussions!

[hongping-zh]

👋 Welcome to EcoCompute Discussions!

Thank you for your interest in energy-efficient LLM inference research!

---

🎯 What is this project?

We discovered that default bitsandbytes INT8 uses 17-147% MORE energy than FP16, contrary to common belief. Through systematic ablation experiments across three NVIDIA GPU architectures, we identified mixed-precision decomposition as the root cause.

🔥 Key Findings

Configuration	Energy vs FP16	Throughput vs FP16
INT8 Default	+17% to +147% ⚠️	-55% to -76% ⚠️
INT8 Pure (threshold=0.0)	-5.5% ✅	+79% to +98% ✅

Root cause: INT8↔FP16 type conversions in outlier-aware decomposition, not INT8 arithmetic itself.

---

📊 Research Quality

• ✅ 32 measurements across 3 GPU architectures (RTX 5090 Blackwell, RTX 4090D Ada Lovelace, A800 Ampere)
• ✅ High precision: Coefficient of Variation < 2% (n=10 per configuration)
• ✅ Causal analysis: Controlled ablation experiments on dual models × dual architectures
• ✅ Full reproducibility: Complete metadata, scripts, and protocols publicly available

📁 View Metadata →
📊 Interactive Dashboard →

---

🤝 How to Participate

💡 Share Ideas

Have suggestions for new experiments or visualizations? Start a discussion in Ideas!

🙋 Ask Questions

Confused about the methodology or results? Ask in Q&A and I'll respond within 24-48 hours.

📊 Share Your Results

Run benchmarks on your GPU and share findings in Results Sharing. We especially need:

• H100, A100 (Hopper, Ampere datacenter)
• AMD MI300, Intel GPUs
• LLaMA-3, Gemma, other models
• Different batch sizes

🎓 Academic Discussion

Discuss methodology, statistical approaches, or related research in Research.

🤝 Find Collaborators

Looking for co-authors for extended studies? Post in Collaboration!

---

🚀 Quick Links

• 📄 Paper (Draft): GitHub
• 📊 Dashboard: Live Demo
• 📁 Metadata: Complete Dataset
• 💻 Code: Benchmark Scripts
• 🎓 Cite: CITATION.cff

---

📧 Contact

Author: Hongping Zhang
Email: zhanghongping1982@gmail.com
Location: Changsha, Hunan, China

---

🌟 Current Focus

I'm actively working on:

1. Accuracy assessment: Perplexity and downstream task evaluation for pure INT8
2. Extended GPU coverage: H100, A100 validation
3. Alternative methods: GPTQ, AWQ, TensorRT-LLM comparison
4. Paper submission: Preparing for MLSys/ICML 2026

---

💬 Discussion Guidelines

• Be respectful: Constructive criticism is welcome, personal attacks are not
• Stay on topic: Keep discussions relevant to energy efficiency and LLM inference
• Share data: If posting results, include hardware specs and software versions
• Cite sources: Reference papers, code, or data when making claims
• Ask for help: No question is too basic!

---

Looking forward to your contributions and discussions! 🚀

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