Research map / Architectures and efficiency / Quantization
Mixed Precision Training
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 mixed-precision training | CNNs including AlexNet, VGG-D, GoogLeNet, Inception v2/v3, and ResNet-50 · ILSVRC12 validation | top-1 accuracy: 56.77% -> 56.93%; 65.40% -> 65.43%; 68.33% -> 68.43%; 70.03% -> 70.02%; 73.85% -> 74.13%; 75.92% -> 76.04% | No clear effect |
| Add loss scaling (ablation) | Faster R-CNN mixed-precision training · Pascal VOC 2007 test | mAP: 68.6% -> 69.7% | Helped |
| Add loss scaling (ablation) | Multibox SSD mixed-precision training · Pascal VOC 2007 test | mAP: diverges -> 77.1% | Helped |
| Use mixed-precision training | DeepSpeech 2 English · WSJ '92 test | character error rate: 2.20 -> 1.99 | Helped |
| Use mixed-precision training | DeepSpeech 2 Mandarin · Mandarin internal test set | character error rate: 15.82 -> 15.01 | Helped |
| Maintain an FP32 master copy of weights for updates (ablation) | Mandarin speech recognition model · Mandarin speech training | relative accuracy: 80% | Helped |
| Omit the FP32 master copy of weights and update FP16 weights directly (ablation) | Mandarin speech recognition model · Mandarin speech training | relative accuracy: 80% | Hurt |
| Add loss scaling (ablation) | Multibox SSD · Pascal VOC 2007 test | mAP: diverges -> 77.1% | Helped |
| Add loss scaling with factor 128 (ablation) | bigLSTM language model · 1 billion word dataset | training perplexity | Helped |
| Use FP16 dot products with FP32 accumulation (ablation) | Mixed-precision neural network training · Various tasks | model accuracy | Helped |
| Use mixed precision rather than FP32 operations | DeepBench operations on Volta GPU · DeepBench | speedup: 2-6x | Helped |
Closest papers by meaning
- Training Deep Neural Networks with 8-bit Floating Point Numbers (NeurIPS 2018 · similarity 0.89)
- Accumulation Bit-Width Scaling For Ultra-Low Precision Training Of Deep Networks (ICLR 2019 · similarity 0.89)
- Collage: Light-Weight Low-Precision Strategy for LLM Training (ICML 2024 · similarity 0.87)
- Towards Learning High-Precision Least Squares Algorithms with Sequence Models (ICLR 2025 · similarity 0.85)
- Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators (ICLR 2024 · similarity 0.85)
- Guaranteed Approximation Bounds for Mixed-Precision Neural Operators (ICLR 2024 · similarity 0.85)
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.