Accelerating Model Training on Ascend Chips: An Industrial System for Profiling...
summary
Hermes is an industrial system that combines cost-to-find profiling (lightweight monitor that triggers fine-grained traces) with a hierarchical inter- and intra-operator bottleneck analysis and an MST advisor that maps observed costs to fixes from 135 real-world cases. It reports low runtime overhead, fixes CPU scheduling and AI‑CPU underutilization, and catches issues like port flapping and Python GC. Real experiments show speedups such as PanGu-α 3.05×, MobileNetV1 1.91×, MoE 1.19× and per-case gains up to 5.34×.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 0:41Model training complexity
- 2:06Profiling tool limitations
- 2:51Ascend chips insights
- 4:16Bottleneck analysis framework
- 5:39Hermas profiling mechanism
- 6:22Interoperator parallelism
- 7:05IO bottleneck detection
- 7:28CPU bottleneck causes
- 7:50Roofline model categorization
- 8:12Communication bottleneck analysis
- 8:33MST advisor tool
- 8:53Five lessons from cases
- 9:57Case studies speedups
- 11:49Conclusion
related talks
15mScaling the Next Paradigm of Heterogeneous Intelligence
22mCentimani: Enabling Fast AI Accelerator Selection for DNN Training with a Novel...
40mOptimize Your AI Cloud Infrastructure: A Hardware Perspective
11mPyTorch in Production: Boosting LLM Training and Inferencing on Ascend NPU
24mAccelerating AI on Edge
17mCharacterization of Large Language Model Development in the Datacenter
12mDemo: NVIDIA's use of Catapult HLS for Building ML Inference Accelerators
25mRace to Superintelligence: Building Low-Friction AI Infrastructure for a Heterogen...
