AIACC 2.0-AIACC Communication Speeding (AIACC-ACSpeed) is a general-purpose training accelerator. AIACC-ACSpeed can be used to accelerate training tasks without requiring code changes to your training model. AIACC-ACSpeed can be applied to all AI training scenarios based on the PyTorch framework and can implement customized optimization for the PyTorch framework. AIACC-ACSpeed can also leverage the capabilities of the nccl-plugin component to accelerate AI training tasks based on the frameworks aside from PyTorch, such as TensorFlow and MXNet.
The following table describes the typical scenarios of AIACC-ACSpeed for AI training.
| Scenario | Applicable model | Common storage |
|---|---|---|
| Image classification and image recognition | ResNet and VGG-16 models | Cloud Paralleled File System (CPFS) |
| CTR prediction | Wide&Deep model | Hadoop Distributed File System (HDFS) |
| Natural Language Processing (NLP) | Transformer and BERT models | CPFS |