Research map / Architectures and efficiency / Quantization
INT vs. FP: A Comprehensive Study of Fine-Grained Low-bit Quantization Formats
Mengzhao Chen, Meng Wu, Hui Jin, Zhihang Yuan, Jing Liu, Chaoyi Zhang, Yunshui Li, Jie Huang, Jin Ma, Zeyue Xue, Zhiheng Liu, Xingyan Bin and 1 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use symmetric integer clipping range [-127,127] with BF16 scale (ablation) | INT8 block-32 training · Training loss, 145M | training loss: 3.1354 -> 3.1251 | Improved |
| use symmetric integer clipping range with UE8M0 scale (ablation) | INT8 block-32 training · Training loss, 145M | training loss: 3.1299 -> 3.1269 | Improved |
| use symmetric clipping at per-channel granularity with UE8M0 scale (ablation) | INT8 per-channel training · Training loss, 145M | training loss: 3.3602 -> 3.4307 | Hurt |
| use symmetric clipping at per-channel granularity with BF16 scale (ablation) | INT8 per-channel training · Training loss, 145M | training loss: 3.2544 -> 3.2560 | No clear effect |
| replace MXFP8 with MXINT8 | direct-cast inference · KL divergence x1e6, Llama-3.1-8B | KL divergence: 359 -> 82 | Improved |
| replace MXFP4 with MXINT4 | direct-cast inference · KL divergence x1e6, Llama-3.1-8B | KL divergence: 7586 -> 12380 | Hurt |
| replace NVFP4 with NVINT4 (no rotation) | direct-cast inference · KL divergence x1e6, Llama-3.1-8B | KL divergence: 3718 -> 4224 | Hurt |
| add random Hadamard rotation to NVINT4 | direct-cast inference · KL divergence x1e6, Llama-3.1-8B | KL divergence: 4224 -> 3609 | Improved |
| add random Hadamard rotation to NVFP4 | direct-cast inference · KL divergence x1e6, Llama-3.1-8B | KL divergence: 3718 -> 4752 | Hurt |
| add random Hadamard rotation to MXINT8 | direct-cast inference · KL divergence x1e6, Llama-3.1-8B | KL divergence: 82 -> 65 | Improved |
| add random Hadamard rotation to MXFP8 | direct-cast inference · KL divergence x1e6, Llama-3.1-8B | KL divergence: 359 -> 409 | Hurt |
| replace MXFP8 with MXINT8 | low-bit training 1B · Training loss, 100B tokens | training loss: 2.6767 -> 2.6758 | Improved |
Closest papers by meaning
- Scaling Laws for Floating–Point Quantization Training (ICML 2025 · similarity 0.89)
- Optimizing Large Language Model Training Using FP4 Quantization (ICML 2025 · similarity 0.88)
- FALQON: Accelerating LoRA Fine-tuning with Low-Bit Floating-Point Arithmetic (NeurIPS 2025 · similarity 0.87)
- BlockDialect: Block-wise Fine-grained Mixed Format Quantization for Energy-Efficient LLM Inference (ICML 2025 · similarity 0.87)
- QuEST: Stable Training of LLMs with 1-Bit Weights and Activations (ICML 2025 · similarity 0.87)
- RaZeR: Pushing the Limits of NVFP4 Quantization with Redundant Zero Remapping (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
