Brain-like chips and electronic devices

Abstract

The invention discloses a brain-like chip and an electronic device. Traditional artificial intelligence chips face the dilemma of Moore’s wall, memory wall and energy-efficiency wall. In comparison, although the computing energy efficiency of most brain-like chips has been significantly improved, they still face the problems of integration wall and energy-efficiency wall, which have not been substantially satisfied. Requirements for cost and power consumption in edge computing scenarios. In order to solve the single problem of cost, power consumption and task capability, the present invention integrates multiple types of dedicated pulse neural network cores in the chip, and the network can call dedicated cores with different resource scales and characteristics, which not only significantly improves power consumption and cost. , and it can also process multi-modal information in parallel with multiple networks and collaborative decision-making to improve the performance of the network. The ability of multitasking and parallel processing makes the brain-like chip disclosed in the present invention more similar to a biological brain, and the more intelligent, low-cost, and energy-efficient brain-like chip disclosed in the present invention provides the possibility for the intelligent interconnection of all things. The invention is suitable for the fields of brain-like chips and AIoT.

Description

technical field [0001] The invention relates to a brain-like chip and an electronic device, and in particular to a multimodal brain-like chip and an electronic device with a variety of dedicated neural network cores. Background technique [0002] The three most important indicators of artificial intelligence (AI) chips are cost, power consumption and performance. At present, the cost and power consumption of traditional AI chips based on artificial neural networks (ANNs) are relatively high, and it is difficult to adapt to the needs of edge computing, so people begin to turn their attention to neuromorphic computing inspired by the brain. [0003] Neuromorphic (Neuromorphic, also known as neuromorphic) computing is a new computing architecture developed in recent years. Real-time reasoning of input signals can be achieved by running spiking neural networks (SNNs) on neuromorphic hardware (commonly known as brain-like chips). [0004] IBM's True-North is a typical represent...

AI Technical Summary

Problems solved by technology

[0005] But the redundancy of this scheme is still too high, refer to image 3 , the neuron integration level of the chip is usually difficult to break through 3k / mm 2

In order to improve neuron integration / reduce silicon cost, one technical approach is to alleviate through advanced chip process nodes, such as Loihi-2 chips, but this does not essentially solve the problem

In addition, the known brain-like chips can only deploy a single spiking neural network, and can only meet the needs of a single type of task, and the chip function is relatively single

Images