HPC


High Peformance Computing

  • About distributing storage and processing over many processors.
  • Not just hardware, but also software to manage the execution of the tasks on ha rdware
    • This is why we need the orchestration softwares.
    • e.g. The AlphaGo distributed processing on many different GPUs.

Distributed Machine Learning

  • Different models can be trained in parallel using different subsets of the data, and then the results can be merged.
  • The operation of a single but large model can be distributed over many processors.
    • e.g. A deep NN which have many layers, parts of it can be run on different processors working as a pipeline

Time and Space

  • Machine learning is not just a component in a product.
  • Time and space sometimes are more important than accuracy
    • Time: a self-driving car in a real world only has limited amount of time to decide whether there is pedestrian on the road or not.
      • The accuracy of the recognizer becomes irrelevent when it can not decide fast enough.
    • Space: IoT devices have limited computing power.
      • Such systems may be making decisions on site or may just work as smart processors of the signal.

Last modified March 10, 2021