Traditional architecture: FPGA (field programmable gate array) and ASIC (application specific integrated circuit) chips based on von Neumann architecture. FPGA and ASIC chips have formed a certain scale of production and application, whether in R & D or application. Although brain like chips are still in the early stage of research, they have great potential in the future and may become the mainstream of artificial intelligence industry in the future.
Today’s computers use von Neumann architecture. Its core architecture is that the processor and memory are arranged separately, so the CPU (central processing unit) and memory module are not integrated, but a cache with very small capacity is set in the CPU. Human brain like architecture is a structure that imitates the model of human brain nervous system. Neurons in human brain are both control system and storage system.
Therefore, CPU, memory module, bus, north-south bridge, etc. will eventually be integrated to form a huge chipset similar to human brain. As for the heating problem and the miniaturization of memory module, mankind will finally find a solution.
1. General purpose chip (GPU). GPU is a single instruction, multi data processing, using a large number of computing units and ultra long pipeline, mainly processing the operation acceleration in the field of image. GPU cannot be used alone. It is only an expert in processing big data calculation. It must be called by the CPU and issued instructions to work.
However, the CPU can act alone to process complex logical operations and different data types. However, when it is necessary to process big data computing, it can call GPU for parallel computing.
2. Semi customized chip (FPGA). FPGA is applicable to the analysis of multi instruction and single data stream. In contrast to GPU, it is often used in the prediction stage, such as the cloud. FPGA uses hardware to realize software algorithm, so it is difficult to realize complex algorithm, but the disadvantage is that the price is relatively high.
3. Fully customized chip (ASIC). ASIC is a special AI chip customized to meet the application requirements of specific scenarios. In addition to not being scalable, it has advantages in power consumption, reliability and volume, especially in mobile devices with high performance and low power consumption.
4. Brain like chip, brain like chip architecture is a new chip programming architecture that simulates the neural network model of human brain. This system can simulate the function of human brain for perception mode, behavior mode and thinking mode.
Some people say that ASIC is a main development direction of artificial intelligence chip, but the future development direction of real artificial intelligence chip is brain like chip. Brain like chip research is very difficult. The chip strategies of IBM, Qualcomm, Intel and other companies use hardware to imitate the neural synapses of the human brain. The future application direction of GPU is advanced complex algorithms and general artificial intelligence platform, which can be used when purchased.
FPGA is more suitable for various specific industries, and artificial intelligence will be applied to various specific fields. ASIC chip is a fully customized chip. Because the more complex the algorithm is, the more it needs a special chip architecture to correspond to it. Customization is to independently design a set of professional intelligent algorithm software for customers when they are in a special scene. ASIC is customized independently based on artificial intelligence algorithm, and its development prospect is promising. Brain like chip is the final development model of artificial intelligence, but it is still far from industrialization.