June 29, 2022
AI chips
AI chips

AI chips (also known as AI hardware or AI accelerators) are accelerators designed specifically for artificial neural network (ANN) based applications. Most commercial ANN applications are deep learning applications. ANN is a subfield of artificial intelligence. ANN is a machine learning approach inspired by the human brain. It includes artificial neuron layers, mathematical functions inspired by how human neurons work.  Machine learning applications that use these networks are called deep learning. There are two main use cases for deep learning.

  • Training: Deep ANNs are provided with thousands of labelled data to identify patterns. Training is time consuming and intensive on computing resources.
  • Inference: As a result of the training process, the ANN can make predictions based on new inputs.

General purpose chips can also run ANN applications, but they are not the most effective solution for such software. There are different types of AI chips as different types of ANN applications require customization. For example, in some IoT applications where IoT devices must be battery-powered, AI chips must be physically small and built to operate efficiently with low power. This causes chipmakers to choose different architectures while designing chips for different applications.

General-purpose hardware uses arithmetic blocks for native in-memory calculations. Serial processing does not provide sufficient performance for deep learning techniques.

  • Neural networks require a lot of parallel/simple arithmetic operations.
  • A powerful general-purpose chip cannot support a large number of simple concurrent operations.
  • AI-optimized HW contains many less powerful chips that enable parallel processing.