Smart chips can be classified according to the way they process signals, design concepts, and application fields. For example, analog chips and digital chips are divided according to how they process signals; According to the design concept, the chip can be divided into general purpose chips and special chips. In addition, if viewed from the application field, it can also be divided into aerospace grade chips, automotive grade chips, and so on.
< br >
When discussing artificial intelligence (AI) related smart chips, we focus primarily on chips that continue traditional computing architectures, such as graphics processors (Gpus), semi-customized chips (FPgas), and application-specific integrated circuits (ASics). These chips are designed to speed up hardware computing power. Among them, GPU is widely used in the field of image processing computing acceleration, due to its large number of computing units and ultra-long pipeline design, making it very suitable for processing big data computing. However, the GPU cannot be used alone and must be called and instructed by the CPU to work.
< br >
Semi-custom chips (FPgas) are suitable for multi-instruction, single-data stream analysis and are often used in predictive phases such as the cloud. Unlike Gpus, FPgas implement software algorithms with hardware, so there is a certain difficulty in implementing complex algorithms, and the price is high.
< br >
The other is an application-specific integrated circuit (ASIC), which is a custom type of chip designed for a specific user or specific electronic system.
< br >
It is worth noting that in addition to these traditional AI chips, there is also a brain-like neural structure chip that subverts the classic von Neumann computing architecture, such as the IBM TrueNorth chip, which enables more efficient computing by simulating the structure and function of the human brain neural network. Although the current stage, GPU and CPU-based AI chips are still the mainstream, but with the continuous optimization of vision, voice, and deep learning algorithms on FPGA and ASIC chips, these two chips will gradually occupy more market share, so as to achieve a long-term coexistence with GPU.