Research map / Architectures and efficiency / Quantization
FP8 Quantization: The Power of the Exponent
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 FP8 (best flexible-bias format) instead of INT8 for PTQ | BERT-base · GLUE PTQ | score: 71.03 -> 82.80 | Helped |
| use FP8 instead of INT8 for PTQ | ViT · ImageNet PTQ | score: 76.39 -> 77.71 | Helped |
| use FP8 flexible-bias instead of INT8 for PTQ | ResNet18 · ImageNet PTQ | score: 69.55 -> 69.66 | No clear effect |
| use FP8 instead of INT8 for PTQ across networks | seven networks · PTQ suite | score | Helped |
| use fully flexible per-tensor m/e bias instead of fixed m/e with flexible bias (ablation) | FP8 PTQ · PTQ suite | score: SalsaNext 55.80 -> 55.02 | No clear effect |
| use a fixed format (single bias) instead of flexible per-channel bias (ablation) | FP8 PTQ · DeepLabV3 PTQ | score: 9.42 -> 72.58 | Helped |
| add per-channel/tensor flexible bias to fixed format (ablation) | FP8 PTQ · MobileNetV2 PTQ | score: 68.49 -> 71.06 | Helped |
| learn clipping value c and mantissa bits m during QAT (ablation) | FP8 flex-bias init · ResNet18 QAT | accuracy: 69.97 -> 70.25 | Helped |
| learn c and m during QAT from fixed init (ablation) | BERT · BERT QAT | accuracy: 83.89 -> 83.85 | No clear effect |
| switch from PTQ to QAT | FP8 vs INT8 · QAT | accuracy | Mixed |
Closest papers by meaning
- Scaling Laws for Floating–Point Quantization Training (ICML 2025 · similarity 0.91)
- F8Net: Fixed-Point 8-bit Only Multiplication for Network Quantization (ICLR 2022 · similarity 0.91)
- Shifted and Squeezed 8-bit Floating Point format for Low-Precision Training of Deep Neural Networks (ICLR 2020 · similarity 0.90)
- BiE: Bi-Exponent Block Floating-Point for Large Language Models Quantization (ICML 2024 · similarity 0.87)
- Scaling FP8 training to trillion-token LLMs (ICLR 2025 · similarity 0.86)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.86)
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.