AI Paper Atlas

Research map / Architectures and efficiency / Quantization

GPT3.int8(): 8-bit Matrix Multiplication for Transformers at Scale

Tim Dettmers, Mike Lewis, Younes Belkada, Luke Zettlemoyer

NeurIPS 2022 · Architectures and efficiency › Quantization · approach: Ultra-low-bit LLM quantization

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 changedWhereMeasuredEffect
Use LLM.int8() vector-wise quantization and mixed-precision decompositionTransformer feed-forward and attention projection layers · C4 validation, models from 125M to 13Bperplexity: 16.48 -> 12.45Improved
Add mixed-precision decomposition to vector-wise quantization (ablation)Int8 transformer quantization · C4 validation, 13Bperplexity: 16.48 -> 12.45Improved
Use vector-wise rather than row-wise quantization (ablation)Int8 absmax quantization · C4 validation, 13Bperplexity: 16.49 -> 16.48No clear effect
Use zeropoint rather than absmax quantization (ablation)Vector-wise Int8 quantization · C4 validation, 13Bperplexity: 16.48 -> 13.47Improved
Add mixed-precision decomposition to row-wise absmax quantization (ablation)Int8 transformer quantization · C4 validation, 13Bperplexity: 16.49 -> 12.46Improved
Convert the FFN to 8-bit (ablation)Language-model training · WMT14+16BLEU: 28.9 -> 28.8No clear effect
Convert attention linear projections to 8-bit along with the FFN, without decomposition (ablation)Language-model training · WMT14+16BLEU: 28.9 -> unstableHurt
Add 16-bit decomposition for 8-bit FFN and attention linear layers (ablation)Language-model training · WMT14+16BLEU: 28.9 -> 28.0; 28.9 -> 27.6; 28.9 -> 27.5Hurt
Convert FFN layers to 8-bit (ablation)RoBERTa-large fine-tuning · GLUE meanscore: 88.6 -> 88.8No clear effect
Convert attention linear projections to 8-bit without decomposition (ablation)RoBERTa-large fine-tuning · GLUE meanscore: 88.8 -> 87.7Hurt
Add 2% decomposition to 8-bit attention linear projections (ablation)RoBERTa-large fine-tuning · GLUE meanscore: 87.7 -> 88.7Improved
Use LLM.int8() matrix multiplication rather than FP16GPT-3 2.7B matrix multiplication · GPT-3 2.7Bspeedup: 1.00x -> 0.64xHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy