Neural network weight storage method, reading method and related equipment

Abstract

The present application provides a neural network weight storage method, reading method and related equipment, wherein the storage method includes: curing the weight of the first layer of the neural network in the static memory of the static random access memory when the system is initialized ; Obtain the cumulative length of the weights of each layer of the neural network in the circular cache of the static random access memory; solidify the weights of the target layer in each layer in the static memory according to the cumulative length of the weights; execute the next acquisition of the weights The operation of accumulating the length repeats the operation of obtaining the accumulative length of the weight for several times until there is no target layer in each layer; the size of the circular buffer is set according to the accumulative length of the corresponding weight when there is no target layer in each layer. The embodiments of the present application help to solve the problems of overflow and underflow of the circular buffer, so that the chip can achieve better throughput, and further reduce the power consumption and cost of the chip.

Description

technical field [0001] The present application relates to the field of computer technology, in particular to a neural network weight storage method, reading method and related equipment. Background technique [0002] As the computing power of AI (Artificial Intelligence, artificial intelligence) chips is getting stronger and the neural network model is getting deeper and deeper, the AI ​​processor needs to read a large amount of weight data for calculation when performing training or reasoning. In order to reduce the power consumption and cost of the chip, relevant personnel have done in-depth research on the storage of neural network weights. The current mainstream storage solutions include high-speed DDR (Double DataRate, double-rate synchronous dynamic random access memory) combined with Cache cache, or using super large capacity of SRAM (Static Random-Access Memory, static random access memory), and further there is a scheme using ping-pong cache. However, the above-men...

AI Technical Summary

Problems solved by technology

However, the above solutions still have various disadvantages, such as large transmission bandwidth, large storage space or difficulty in making the throughput of the chip reach an optimal state, resulting in high power consumption and cost of the chip

Images