AIACC can be applied to all AI training scenarios and AI inference scenarios.

Typical business scenarios of AI training

AI training is applied to the following typical business scenarios:

  • Image classification and image recognition
    • Common framework: MXNet.
    • Common storage: Cloud Paralleled File System (CPFS).
  • CTR prediction
    • Common framework: TensorFlow.
    • Model: Wide & Deep.
    • Common storage: Hadoop Distributed File System (HDFS).
  • Natural Language Processing (NLP)
    • Common framework: TensorFlow.
    • Model: Transformer and BERT
    • Common storage: Cloud Paralleled File System (CPFS).

Typical business scenarios of AI inference

AI inference is applied to the following typical business scenarios:

  • Video Ultra HD inference
    • Model: Ultra HD
    • Configuration: T4 GPU.
    • After the following performance optimization is performed, the performance is improved by 1.7 times.
      • Video decoding is ported to the GPU.
      • Preprocessing and postprocessing are ported to GPU.
      • The data set size that is automatically obtained from a single operation.
      • The deep optimization of convolution.
  • Online inference of image synthesis
    • Model: GAN.
    • Configuration: T4 GPU.
    • The following performance optimization improves the performance by three times.
      • Preprocessing and postprocessing are ported to GPU.
      • The data set size that is automatically obtained from a single operation.
      • The deep optimization of convolution.
  • Prediction and inference of CTR
    • Model: Wide & Deep.
    • Configuration: M40 GPU.
    • The following performance optimization improves the performance by 5.1 times.
      • Pipeline optimization.
      • Model splitting.
      • The child models are separately optimized.
  • Interface of natural language understanding
    • Model: BERT
    • Configuration: T4 GPU.
    • The following performance optimization improves the performance by 2.3 times.
      • Pipeline optimization of preprocessing and postprocessing.
      • The data set size that is automatically obtained from a single operation.
      • The deep optimization of Kernel.