Method and system for reducing eDRAM (Embedded Dynamic Random Access Memory) refreshing energy consumption in neural network chip

Abstract

The invention provides a method and system for reducing eDRAM (Embedded Dynamic Random Access Memory) refreshing energy consumption in a neural network chip. The method comprises the steps of trainingan original neural network model, and determining a target neural network model with the maximum fault-tolerant capability, and data hold time of an eDRAM corresponding to the target neural network model; scheduling each layer of the target neural network model, and determining a computing mode of each layer and data survival time of each layer under the lowest computing energy consumption; and setting the target neural network model to be performed on the neural network chip according to the computing model of each layer, for storage subareas of each layer, if effective data is not stored inthe storage subareas or the data survival time of each layer is shorter than the data holder time, not refreshing the storage subareas. According to the method, the unnecessary refreshing operation can be removed to the greatest extent, and the eDRAM refreshing energy consumption in the neural network chip is greatly reduced.

Description

technical field [0001] The invention belongs to the field of neural network chip acceleration, and in particular relates to a method and system for reducing energy consumption of eDRAM refreshing in a neural network chip. Background technique [0002] With the advent of the era of artificial intelligence, intelligent tasks such as image recognition, speech recognition, and natural language processing are ubiquitous in life. Neural networks have achieved world-leading results in such intelligent tasks and are widely used in the industry. For example, Baidu image search, Microsoft speech recognition and Google online translation, etc., are all implemented based on neural networks. Due to its regular calculation and large parallelism, the neural network needs and is especially suitable for using neural network chips to accelerate its calculation. However, due to the large amount of data and the limited internal memory capacity of the chip, a large number of external memory ac...

AI Technical Summary

Problems solved by technology

[0004] The invention is used to solve the defect that the eDRAM type neural network chip in the prior art has relatively large refresh energy consumption and weakens the energy consumption benefit brought by the reduction of external storage access

Images