Research map / Architectures and efficiency / Quantization
RDO-Q: Extremely Fine-Grained Channel-Wise Quantization via Rate-Distortion Optimization
Zhe Wang, Jie Lin, Xue Geng, Mohamed M. Sabry Aly, Vijay Chandrasekhar
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 RDO-Q channel-wise allocation vs best prior 2-bit method | ResNet-18 · ImageNet 2-bit | top-1 accuracy (%): 67.8% -> 68.8% | Improved |
| use RDO-Q vs best prior 2-bit method | ResNet-34 · ImageNet 2-bit | top-1 accuracy (%): 72.1% -> 72.6% | Improved |
| use RDO-Q vs best prior 2-bit method | ResNet-50 · ImageNet 2-bit | top-1 accuracy (%): 73.8% -> 75.0% | Improved |
| use RDO-Q vs SAT at 4 bits | MobileNet-v2 · ImageNet 4-bit | top-1 accuracy (%): 71.1% -> 71.3% | Improved |
| add hardware-aware constraints, ResNet-50 vs uncompressed (ablation) | ResNet-50 · Google TPU | inference rate (images/s): 361 -> 1254 | Improved |
| add hardware-aware constraints, ResNet-50 vs uncompressed (ablation) | ResNet-50 · MIT Eyeriss | inference rate (images/s): 5.6 -> 17.0 | Improved |
| hardware-aware RDO-Q vs equal-bit DoReFa+PACT at 76.5% accuracy | ResNet-50 · Google TPU | inference rate (images/s): 646 -> 769 | Improved |
| hardware-aware RDO-Q vs HAQ | MobileNet-V2 · Google TPU | inference rate (images/s): 2067 -> 2197 | Improved |
| rate-distortion search instead of reinforcement learning (HAQ) | ResNet-50 · bit allocation search time | time: 117 hours -> 7 minites | Improved |
| rate-distortion search instead of reinforcement learning (HAQ) | MobileNet-v2 · bit allocation search time | time: 79 hours -> 6 minites | Improved |
| use channel-wise instead of layer-wise bit allocation (ablation) | MobileNet-V2 · ImageNet | top-1 accuracy (%) | Improved |
| add hardware-aware constraints to channel-wise allocation (ablation) | MobileNet-V2 · Eyeriss and TPU | inference rate | Improved |
Closest papers by meaning
- AdaBits: Neural Network Quantization With Adaptive Bit-Widths (CVPR 2020 · similarity 0.89)
- One-Shot Model for Mixed-Precision Quantization (CVPR 2023 · similarity 0.88)
- HAQ: Hardware-Aware Automated Quantization With Mixed Precision (CVPR 2019 · similarity 0.88)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.87)
- Quantization without Tears (CVPR 2025 · similarity 0.87)
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 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.
