CodePudding user response:
On the one hand is the GPU for general computing support is more and more big; On the other hand is the present stage model are mainly composed of CNN, the CNN would eventually convolution operation, and the operation of the CNN is right support parallel computing algorithm, and of course many layers and support parallel computing; In addition, the GPU can share the load of CPU,People feel a little regret, however, at this stage in the GPU are mainly composed of floating-point reasoning, the reasoning CPU reasoning based on fixed-point reasoning is given priority to, hope to see the GPU as soon as possible with fixed-point reasoning framework,