Research map / Point clouds / Point cloud classification
Pushing the Limits of Narrow Precision Inferencing at Cloud Scale with Microsoft Floating Point
Bita Darvish Rouhani, Daniel Lo, Ritchie Zhao, Ming Liu, Jeremy Fowers, Kalin Ovtcharov, Anna Vinogradsky, Sarah Massengill, Lita Yang, Ray Bittner, Alessandro Forin, Haishan Zhu and 12 more
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 |
|---|---|---|---|
| increase mantissa bits from MSFP12 to MSFP13 (ablation) | MSFP · ImageNet | normalized top-1 accuracy: 0.967 -> 0.989 | Improved |
| increase mantissa bits from MSFP12 to MSFP13 (ablation) | MSFP on MobileNet-V1 · ImageNet | normalized top-1 accuracy: 0.863 -> 0.965 | Improved |
| use more bits for activations than weights (weights MSFP12, acts MSFP15 vs weights MSFP15, acts MSFP12) (ablation) | MSFP ResNet-50 · ImageNet | normalized top-1 accuracy: 0.959 -> 0.988 | Improved |
| lower precision on RNN/Transformer models (MSFP16 to MSFP12) (ablation) | MSFP on BERT-SQuAD1.1 · SQuAD1.1 | normalized F1: 1.000 -> 0.990 | No clear effect |
| light fine-tuning (1-10 epochs) of quantized model | MSFP low bit-width · ImageNet and NLP benchmarks | accuracy | Improved |
| share exponent per small bounding box rather than per whole tensor | MSFP · quantization error (KL / QNSR) | quantization error | Improved |
| double bounding-box size (ablation) | MSFP15 · QNSR | QNSR: approximately 0.52dB | Mixed |
Closest papers by meaning
- Scaling Laws for Floating–Point Quantization Training (ICML 2025 · similarity 0.83)
- Pushing the Limits of BFP on Narrow Precision LLM Inference (AAAI 2025 · similarity 0.82)
- Collage: Light-Weight Low-Precision Strategy for LLM Training (ICML 2024 · similarity 0.82)
- Intelligence per Watt: Measuring Intelligence Efficiency of Local AI (NeurIPS 2026 · similarity 0.81)
- HCInfer: An Efficient Inference System via Heterogeneous Error Compensation for Resource-Constrained Devices (NeurIPS 2026 · similarity 0.81)
- HybridFlow: Resource-Adaptive Subtask Routing for Efficient Edge-Cloud LLM Inference (ICML 2026 · similarity 0.81)
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.
