Benchmark Tool
The training benchmark tool can start the training service based on the specified training script and provide multiple records, including the training result, running time, and performance and precision indicators of the training model.
Benefits
You can quickly and easily run a training model to obtain statistics.
Instructions
./benckmark.sh -e ResNet50 -hw 8p -y ./yaml/ResNet50.yaml -docker
For details, see the CANN V100R020C10 Training Benchmark Tool User Guide.
Scenarios
- Single-server training on the host
Single-server single-device, that is, 1P (P indicates an Ascend AI Processor.)
Single-server multi-device, that is, 2P, 4P, and 8P
- Single-server training on a Docker
Single-server single-device, that is, 1P
Single-server multi-device, that is, 2P, 4P, and 8P
- Cluster training on the host
Multi-server multi-device: Multiple servers use the same dataset for training.
- Cluster training on a Docker
Multi-server multi-device: Multiple containers use the same dataset for training.
Applicable Models
Atlas 800 AI training server (model 9000)
Atlas 800 AI training card (model 9010)
Atlas 300T AI training card (model 9000)
Implementation
Use the training script to start the training service.
Add the timing function to the key position of the training to complete the performance test.
