Research map / Point clouds / Point cloud classification
Training DNNs with Hybrid Block Floating Point
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 HBFP with 8-bit dot products and 16-bit weight storage instead of FP32 | Image classification models · CIFAR-100, SVHN, ImageNet | validation error: 26.07% -> 25.12%; 20.35% -> 20.78%; 26.03% -> 26.27%; 1.89% -> 1.98%; 2.00% -> 1.98%; 1.80% -> 1.79%; 23.64% -> 23.88% | Mixed |
| Use HBFP with 12-bit dot products and 16-bit weight storage instead of FP32 | Image classification models · CIFAR-100, SVHN, ImageNet | validation error: 26.07% -> 25.10%; 20.35% -> 20.78%; 26.03% -> 25.82%; 1.89% -> 1.96%; 2.00% -> 1.94%; 1.80% -> 1.85%; 23.64% -> 23.58% | Mixed |
| Use HBFP with 8-bit dot products and 16-bit weight storage instead of FP32 | LSTM · PTB | perplexity: 61.31 -> 61.86 | Hurt |
| Use HBFP with 12-bit dot products and 16-bit weight storage instead of FP32 | LSTM · PTB | perplexity: 61.31 -> 61.35 | Hurt |
| Use 4-bit rather than wider mantissas (ablation) | HBFP WideResNet-28-10 · CIFAR-100 | validation error: 4.1% larger error | Hurt |
| Use mantissas wider than 8 bits (ablation) | HBFP WideResNet-28-10 · CIFAR-100 | validation error difference from FP32: within 1% | No clear effect |
| Add 16-bit weight storage to the 8-bit and 12-bit mantissa configurations (ablation) | HBFP · CIFAR-100, WideResNet-28-10 | validation accuracy: 0.21%; 0.43% | Helped |
| Use 24×24 or 64×64 tiles (ablation) | HBFP WideResNet-28-10 · CIFAR-100 | validation error difference from FP32: within 0.5% | No clear effect |
| Remove tiling (ablation) | HBFP WideResNet-28-10 · CIFAR-100 | validation error difference from FP32: 0.8% | Hurt |
| Use 8-bit dot products instead of FP32 arithmetic | HBFP training · Forward and backward passes | memory bandwidth requirement: up to 4× | Helped |
| Use 8-bit BFP multiply-and-add units instead of FP16 multiply-and-add units | FPGA accelerator · Stratix V 5SGSD5 | throughput: 8.5× | Helped |
Closest papers by meaning
- A Block Minifloat Representation for Training Deep Neural Networks (ICLR 2021 · similarity 0.90)
- Hybrid 8-bit Floating Point (HFP8) Training and Inference for Deep Neural Networks (NeurIPS 2019 · similarity 0.89)
- Scaling Laws for Floating–Point Quantization Training (ICML 2025 · similarity 0.87)
- Pushing the Limits of BFP on Narrow Precision LLM Inference (AAAI 2025 · similarity 0.86)
- Floating-Point Neural Networks Can Represent Almost All Floating-Point Functions (ICML 2025 · similarity 0.84)
- Collage: Light-Weight Low-Precision Strategy for LLM Training (ICML 2024 · similarity 0.84)
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.