Research map / Architectures and efficiency / RNNs and normalization
DropCompute: simple and more robust distributed synchronous training via compute variance reduction
Niv Giladi, Shahar Gottlieb, Moran Shkolnik, Asaf Karnieli, Ron Banner, Elad Hoffer, Kfir Yehuda Levy, Daniel Soudry
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 |
|---|---|---|---|
| Apply DropCompute at 2.5–3% drop rate without compensation | BERT-Large pretraining · SQuAD-v1.1 dev | F1 score: 91.32 ± 0.15 -> 91.34 ± 0.04 | No clear effect |
| Apply DropCompute at 5.5–6% drop rate without compensation | BERT-Large pretraining · SQuAD-v1.1 dev | F1 score: 91.32 ± 0.15 -> 91.44 ± 0.02 | No clear effect |
| Apply DropCompute at 10–11% drop rate without compensation | BERT-Large pretraining · SQuAD-v1.1 dev | F1 score: 91.32 ± 0.15 -> 91.19 ± 0.02 | No clear effect |
| Add 11% extra training steps to compensate for dropped samples (ablation) | DropCompute on BERT-Large · SQuAD-v1.1 dev | F1 score: 91.19 ± 0.02 -> 91.40 ± 0.08 | Improved |
| Increase batch size by 11% to compensate for dropped samples (ablation) | DropCompute on BERT-Large · SQuAD-v1.1 dev | F1 score: 91.19 ± 0.02 -> 91.38 ± 0.08 | Improved |
| Use re-computation to compensate for dropped samples (ablation) | DropCompute on BERT-Large · SQuAD-v1.1 dev | F1 score: 91.19 ± 0.02 -> 91.19 ± 0.11 | No clear effect |
| Clip compute at a threshold to reduce the effect of worker latency outliers | BERT1.5B synchronous training · BERT1.5B training, 162 workers, 64 gradient accumulations | performance boost: ∼18% | Improved |
| Clip compute at a threshold after hardware and software optimizations | BERT1.5B synchronous training · BERT1.5B training, 196 workers | performance boost: 5% | Improved |
| Use DropCompute and run the additional steps needed to reach baseline loss | BERT1.5B synchronous training · BERT1.5B training, 64 workers | training time: approximately 3% more steps; 13% less time | Improved |
| Allow stochastic batch sizes with up to 10% dropped samples | ResNet-50 · ImageNet validation | Top-1 accuracy | No clear effect |
| Apply learning-rate correction for low drop rates (ablation) | ResNet-50 · ImageNet validation | Top-1 accuracy | No clear effect |
| Add DropCompute to Local-SGD when stragglers occur | Local-SGD with ResNet-50 · ImageNet with simulated stragglers | relative speedup | Improved |
Closest papers by meaning
- Moshpit SGD: Communication-Efficient Decentralized Training on Heterogeneous Unreliable Devices (NeurIPS 2021 · similarity 0.88)
- Pipe-SGD: A Decentralized Pipelined SGD Framework for Distributed Deep Net Training (NeurIPS 2018 · similarity 0.87)
- From Promise to Practice: Realizing High-performance Decentralized Training (ICLR 2025 · similarity 0.86)
- A Quadratic Synchronization Rule for Distributed Deep Learning (ICLR 2024 · similarity 0.85)
- Towards a Better Theoretical Understanding of Independent Subnetwork Training (ICML 2024 · similarity 0.85)
- AsyncMesh: Fully Asynchronous Optimization for Data and Pipeline Parallelism (NeurIPS 2026 · 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.
