A Neural Network Accelerator Device with Low Power Consumption

Abstract

The invention relates to a low-power-consumption neural network accelerator framework. In the prior art, the power consumption is large, the start is slow and the data throughput is large. The framework comprises a CPU, a neural network acceleration module, a read-only memory module, an internal readable and writable random access memory, an external readable and writable random access memory, a readable and writable non-volatile memory, a power management module, and two power domains. The neural network acceleration module is used for achieving the hardware acceleration of a command of a neural network, and supporting a weight splitting and structure splitting neural network. The read-only memory module is used for storing the solidified neural network weight parameters and structure parameters. The internal readable and writable random access memory is an SRAM, and the external readable and writable random access memory is a DRAM. The two power domains are a high-frequency start power domain A and a low-frequency start power domain B. According to the invention, not all masks need to be replaced in an upgrading process, and only one mask needs to be replaced, thereby greatly reduces the upgrading cost.

Description

technical field [0001] The invention belongs to the technical field of neural network algorithms, and in particular relates to a low-power consumption neural network accelerator device. Background technique [0002] In recent years, with the deepening of research on neural network algorithms, its accuracy has surpassed all traditional machine learning algorithms in many applications. Neural network algorithms gradually began to replace traditional algorithms and began to be deployed on terminal devices. However, although the accuracy of the neural network algorithm is very good, the amount of calculation is very large, which leads to a large consumption of memory bandwidth and overall power consumption. Terminal devices are often embedded devices, and some are even powered by dry batteries, which require high power consumption and narrow bandwidth. [0003] In order to solve this contradiction, most of the current solutions adopt the solution of cloud plus terminal for dep...

AI Technical Summary

Problems solved by technology

However, the ultra-low power consumption required in the Internet of Things still cannot be solved

The reason is that although the compression technology is used, the weight is still too large, and the weight needs to be imported every time it runs, resulting in a slow system startup speed, and it is impossible to start in real time when the power is off, so that it cannot achieve long-term standby

In addition, although its bandwidth requirements have decreased, the bandwidth bottleneck is still very obvious for low-cost solutions

Images