Research map / Architectures and efficiency / RNNs and normalization
Tempo: Accelerating Transformer-Based Model Training through Memory Footprint Reduction
Muralidhar Andoorveedu, Zhanda Zhu, Bojian Zheng, Gennady Pekhimenko
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 Tempo (In-place GELU, In-place LayerNorm, dropout recomputation) vs Baseline | BERT-Large · 2080 Ti seq len 512 | throughput speedup: 16% | Improved |
| apply Tempo vs Baseline on larger max batch size | BERT-Large · 2080 Ti seq len 512 | max batch size: 1 -> 2 | Improved |
| apply Tempo vs Baseline on V100 | BERT-Large · V100 seq len 512 | max batch size: 4 -> 7 | Improved |
| use default layer checkpointing instead of Baseline | BERT-Large · 2080 Ti seq len 128 | throughput: 142.3160342 -> 120.0995009 | Hurt |
| Tempo versus Checkpoint on memory footprint at batch size 15 | BERT-Large · 2080 Ti seq len 128 | total memory: 8.3 GB -> 9.2 GB | Hurt |
| Tempo versus Baseline on memory at batch size 15 | BERT-Large · 2080 Ti seq len 128 | total memory: 11.3 GB -> 9.2 GB | Improved |
| Tempo versus Baseline on pre-training loss | BERT-Base · pre-training phase 1 | loss: 0.5% | No clear effect |
| Tempo versus Baseline on fine-tuning accuracy | BERT-Large · MRPC | accuracy | No clear effect |
| run on A100 with larger hidden sizes (ablation) | Tempo · A100 BERT H=2048/3072 | throughput speedup: 39% | Improved |
| run on longer sequence lengths up to 3072 on A100 (ablation) | Tempo · BERT-Large 12 layers | throughput speedup: 27% | Improved |
| apply Tempo to GPT2 and RoBERTa (ablation) | Tempo · 2080 Ti | throughput speedup: 19% and 26% | Improved |
| apply Tempo to GPT2 and RoBERTa on V100 (ablation) | Tempo · V100 | throughput speedup: 5% and 4% | Improved |
Closest papers by meaning
- A Multi-Level Framework for Accelerating Training Transformer Models (ICLR 2024 · similarity 0.88)
- TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting (ICLR 2024 · similarity 0.87)
- DeepSpeed Data Efficiency: Improving Deep Learning Model Quality and Training Efficiency via Efficient Data Sampling and Routing (AAAI 2024 · similarity 0.87)
- GPAS: Accelerating Convergence of LLM Pretraining via Gradient-Preserving Activation Scaling (NeurIPS 2025 · similarity 0.87)
- IceFormer: Accelerated Inference with Long-Sequence Transformers on CPUs (ICLR 2024 · similarity 0.86)
- MemoryFormer : Minimize Transformer Computation by Removing Fully-Connected Layers (NeurIPS 2024 · similarity 0.86)
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.
