Board card and neural network operation method

Abstract

The invention discloses a board card. A mapping unit of the board card receives an input neuron and a weight value and then processes the input neuron and the weight value to obtain a processed inputneuron and a processed weight value; And an operation unit of the board card performs artificial neural network operation on the processed input neurons and the processed weights. By adopting the embodiment of the invention, the extra expenditure of the device can be reduced, the access amount is reduced, and the neural network operation efficiency is improved.

Description

technical field [0001] The invention relates to the field of neural networks, in particular to a board and a neural network operation method. Background technique [0002] Artificial Neural Networks (ANNs) are referred to as neural networks (Neural Networks, NNs) for short. It is an algorithmic mathematical model that imitates the behavior characteristics of animal neural networks and performs distributed parallel information processing. This kind of network depends on the complexity of the system, and achieves the purpose of processing information by adjusting the interconnection relationship between a large number of internal nodes. [0003] Neural network is an algorithm with high computation and high memory access. The more weights, the more computation and memory access. will increase. In order to reduce the amount of calculation and the number of weights, thereby reducing the amount of memory access, a sparse neural network is proposed. The definition of sparse neu...

AI Technical Summary

Problems solved by technology

[0004] In the prior art, GPU involves the calculation of sparse neural network when performing graphics and image operations and scientific calculations, but since GPU is a device specially used to perform graphics and image operations and scientific calculations, there is no provision for sparse convolutional neural network operations. Dedicated support, still requires a lot of front-end decoding work to perform sparse artificial neural network operations, bringing a lot of additional overhead

In addition, the GPU has only a small on-chip cache, and the model data (weights) of the multi-layer artificial neural network need to be repeatedly moved from off-chip. The off-chip bandwidth has become the main performance bottleneck, and it has brought huge power consumption overhead.

Images