Research map / Architectures and efficiency / RNNs and normalization
Piper: Multidimensional Planner for DNN Parallelization
Jakub Tarnawski, Deepak Narayanan, Amar Phanishayee
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 |
|---|---|---|---|
| allow heterogeneous stages instead of equi-partitioned stages (ablation) | Piper · BERT-32, 8GB, N=512 | Time-Per-Sample: 10-15% lower TPS | Improved |
| add tensor parallelism (ablation) | Piper · BERT-32, k > N devices or memory-limited | throughput / feasibility | Improved |
| add data parallelism (ablation) | Piper · BERT-32, 8GB, N=512 | Time-Per-Sample | Improved |
| add activation recomputation (ablation) | Piper · BERT-32, 2GB per device | throughput | Improved |
| add activation recomputation (ablation) | Piper · BERT-32, 80GB per device | throughput: 1.000× | No clear effect |
| add tensor parallelism (speedup in isolation) | Piper · BERT-32 | speedup | Mixed |
Closest papers by meaning
- Memory-Efficient Pipeline-Parallel DNN Training (ICML 2021 · similarity 0.88)
- Practical Performance Guarantees for Pipelined DNN Inference (ICML 2024 · similarity 0.86)
- Synergistic Tensor and Pipeline Parallelism (NeurIPS 2025 · similarity 0.86)
- UniAP: Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming (CVPR 2025 · similarity 0.84)
- FlashRNN: I/O-Aware Optimization of Traditional RNNs on modern hardware (ICLR 2025 · similarity 0.83)
- Pfeife: Automatic Pipeline Parallelism for PyTorch (ICML 2025 · similarity 0.83)
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.
