EnvPipe: Performance-preserving DNN Training Framework for Saving Energy
summary
EnvPipe is a PyTorch library that profiles each pipeline stage, reschedules pipeline units to expose usable pipeline bubbles, and selectively lowers GPU SM frequency to stretch non‑critical stages. It keeps hyperparameters and accuracy unchanged, runs an online profiler plus a scheduler and iterative frequency planner, and reports up to 25.2% energy savings on 4×V100 single‑node and 28.4% on 16‑GPU multi‑node while keeping throughput loss under 1%.
AI-written summary · about
chapters (9)
- 0:00Introduction to sustainability
- 1:15Energy and performance trade-off
- 3:45Pipeline parallelism bubbles
- 7:19Goals of MVpipe
- 8:01System components
- 9:49Scheduler design
- 14:14Frequency planner
- 17:59Evaluation results
- 20:06Conclusion
related talks
20mHetPipe: Enabling Large DNN Training on (Whimpy) Heterogeneous GPU Clusters through
17mPPipe: Efficient Video Analytics Serving on Heterogeneous GPU Clusters via...
14mFlexPipe: Maximizing Training Efficiency for Transformer-based Models with...
14mFine-tuning giant neural networks on commodity hardware with automatic pipeline...
15mPipeThreader: Software-Defined Pipelining for Efficient DNN Execution
17mAttack of the Bubbles: Straggler-Resilient Pipeline Parallelism for Large Model Training
35mNNStreamer: Efficient and Agile On-Device AI Application Development with Neural-Net Pipelines
22mOptimizing Large MoE Inference on NVIDIA Blackwell: NVFP4, ADP, and DualPipe Strat...