Research map / Architectures and efficiency / Quantization
GPT3.int8(): 8-bit Matrix Multiplication for Transformers at Scale
Tim Dettmers, Mike Lewis, Younes Belkada, Luke Zettlemoyer
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 LLM.int8() vector-wise quantization and mixed-precision decomposition | Transformer feed-forward and attention projection layers · C4 validation, models from 125M to 13B | perplexity: 16.48 -> 12.45 | Improved |
| Add mixed-precision decomposition to vector-wise quantization (ablation) | Int8 transformer quantization · C4 validation, 13B | perplexity: 16.48 -> 12.45 | Improved |
| Use vector-wise rather than row-wise quantization (ablation) | Int8 absmax quantization · C4 validation, 13B | perplexity: 16.49 -> 16.48 | No clear effect |
| Use zeropoint rather than absmax quantization (ablation) | Vector-wise Int8 quantization · C4 validation, 13B | perplexity: 16.48 -> 13.47 | Improved |
| Add mixed-precision decomposition to row-wise absmax quantization (ablation) | Int8 transformer quantization · C4 validation, 13B | perplexity: 16.49 -> 12.46 | Improved |
| Convert the FFN to 8-bit (ablation) | Language-model training · WMT14+16 | BLEU: 28.9 -> 28.8 | No clear effect |
| Convert attention linear projections to 8-bit along with the FFN, without decomposition (ablation) | Language-model training · WMT14+16 | BLEU: 28.9 -> unstable | Hurt |
| Add 16-bit decomposition for 8-bit FFN and attention linear layers (ablation) | Language-model training · WMT14+16 | BLEU: 28.9 -> 28.0; 28.9 -> 27.6; 28.9 -> 27.5 | Hurt |
| Convert FFN layers to 8-bit (ablation) | RoBERTa-large fine-tuning · GLUE mean | score: 88.6 -> 88.8 | No clear effect |
| Convert attention linear projections to 8-bit without decomposition (ablation) | RoBERTa-large fine-tuning · GLUE mean | score: 88.8 -> 87.7 | Hurt |
| Add 2% decomposition to 8-bit attention linear projections (ablation) | RoBERTa-large fine-tuning · GLUE mean | score: 87.7 -> 88.7 | Improved |
| Use LLM.int8() matrix multiplication rather than FP16 | GPT-3 2.7B matrix multiplication · GPT-3 2.7B | speedup: 1.00x -> 0.64x | Hurt |
Closest papers by meaning
- FrameQuant: Flexible Low-Bit Quantization for Transformers (ICML 2024 · similarity 0.89)
- The Quantization Benefits of Residual-Free Transformers (NeurIPS 2026 · similarity 0.88)
- Outlier-aware Slicing for Post-Training Quantization in Vision Transformer (ICML 2024 · similarity 0.88)
- LUT-GEMM: Quantized Matrix Multiplication based on LUTs for Efficient Inference in Large-Scale Generative Language Models (ICLR 2024 · similarity 0.87)
- IM-Unpack: Training and Inference with Arbitrarily Low Precision Integers (ICML 2024 · similarity 0.87)
- MemoryFormer : Minimize Transformer Computation by Removing Fully-Connected Layers (NeurIPS 2024 · 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.
