Research map / Architectures and efficiency / Quantization
Computation Error Analysis of Block Floating Point Arithmetic Oriented Convolution Neural Network Accelerator Design
Zhourui Song, Zhenyu Liu, Dongsheng Wang
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 BFP with 8-bit weight and input mantissas | VGG-16 · VGG-16 | accuracy drop: 0.0002 | No clear effect |
| Use BFP with 8-bit weight and input mantissas | ResNet-18 · ResNet-18 | accuracy drop: 0.0026 | Hurt |
| Use BFP with 8-bit weight and input mantissas | ResNet-50 · ResNet-50 | accuracy drop: 0.0008 | No clear effect |
| Increase input mantissa length from 6 to 8 bits at fixed 8-bit weight mantissa (ablation) | VGG-16 · VGG-16 | accuracy drop: 0.1772 -> 0.0002 | Improved |
| Increase weight mantissa length from 6 to 8 bits at fixed 8-bit input mantissa (ablation) | VGG-16 · VGG-16 | accuracy drop: 0.1246 -> 0.0002 | Improved |
| Increase input mantissa length from 6 to 8 bits at fixed 8-bit weight mantissa (ablation) | ResNet-50 · ResNet-50 | accuracy drop: 0.0664 -> 0.0008 | Improved |
| Increase input mantissa length from 7 to 8 bits at fixed 7-bit weight mantissa (ablation) | CIFAR-10 · CIFAR-10 | accuracy drop: 0.0042 -> 0.0028 | Improved |
| Increase input mantissa length from 6 to 8 bits at fixed 7-bit weight mantissa (ablation) | CIFAR-10 · CIFAR-10 | accuracy drop: 0.0042 -> 0.0014 | Improved |
| Choose the second tested BFP partition rather than the first (ablation) | VGG-16 · VGG-16, ILSVRC12 | top-1 accuracy: 0.6672 -> 0.6832 | Improved |
| Use BFP partition choice 2 rather than floating-point representation | VGG-16 · VGG-16, ILSVRC12 | top-1 accuracy: 0.6808 -> 0.6832 | No clear effect |
| Round rather than truncate shifted-out mantissa bits (ablation) | BFP CNN arithmetic · CNN convolution | classification accuracy | Improved |
| Use an 8-bit rather than shorter mantissa for deep CNNs (ablation) | VGG-16, GoogLeNet, ResNet-18, ResNet-50 · ILSVRC12 | accuracy drop: less than 0.3% | Improved |
Closest papers by meaning
- Training DNNs with Hybrid Block Floating Point (NeurIPS 2018 · similarity 0.88)
- A Block Minifloat Representation for Training Deep Neural Networks (ICLR 2021 · similarity 0.85)
- Training Deep Neural Networks with 8-bit Floating Point Numbers (NeurIPS 2018 · similarity 0.84)
- Pushing the Limits of BFP on Narrow Precision LLM Inference (AAAI 2025 · similarity 0.84)
- Scaling Laws for Floating–Point Quantization Training (ICML 2025 · similarity 0.82)
- JAQ: Joint Efficient Architecture Design and Low-Bit Quantization with Hardware-Software Co-Exploration (AAAI 2025 · similarity 0.82)
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.
