Neural network processor based on TTA architecture

Abstract

The invention provides a neural network processor based on TTA architecture. The neural network processor based on TTA architecture aims at giving consideration to low power consumption and high speedwhile increasing the neural network scale supported by the processor, and can improve flexibility of design of the processor. The neural network processor based on TTA architecture includes a data exchange network and a plurality of function units which are connected with the data exchange network through the TTA architecture mode, wherein the data exchange network and the function units except for an SD card module and a DDR3 SDRAM module are implemented through FPGA; the function units are used for realizing storage, control and decoding of commands and storage and operating of data; and the data exchange network is used for realizing transmission of the data among each function list. The neural network processor based on TTA architecture can be applied to an image processing system forperforming high speed and real-time processing and identification on images.

Description

technical field [0001] The invention belongs to the technical field of image processing, and relates to a neural network processor, in particular to a TTA-based neural network processor, which can be applied to an image processing system for high-speed parallel real-time image processing and recognition. Background technique [0002] The TTA architecture is an extension of the classic VLIW architecture. The classic VLIW architecture is relatively complex. In order to reduce the complexity of the VLIW architecture, Corporaal and others proposed the TTA architecture. The TTA architecture includes two parts: functional units and data exchange networks. It is a programmable architecture with a simple structure and high instruction execution efficiency. . Each functional unit works relatively independently with high parallelism. [0003] The artificial neural network is based on the basic principles of neural networks in biology. After understanding and abstracting the structur...

AI Technical Summary

Problems solved by technology

When using ASIC to realize the artificial neural network, the general method is that there is one neuron in the neural network, and the hardware circuit is a neuron. This approach is feasible for small-scale neural networks. However, with the development of neural networks , the scale of the neural network is increasing day by day, so that the number of neurons is increasing, and it is unrealistic to adopt this approach; ASIC adopts integrated circuit technology, the design process is complicated, and once the circuit design is completed, it cannot be modified and lacks flexibility. sex

[0005] To sum up, the current artificial neural network processor, despite its high speed and low power consumption, lacks flexibility in design and cannot support arbitrarily large-scale neural networks.

Images