The following statements about overflow and underflow in numerical calculations are correct?

Deep learning neural network training requires a lot of matrix calculations. Generally, we need to use hardware to enable the computer to have parallel computing capabilities. The following hardware devices can provide parallel computing capabilities.:

On-Device Execution, that is, the entire image is offloaded and executed, and the computing power of the Yiteng chip can be fully utilized, which can greatly reduce the interaction overhead, thereby increasing the accelerator occupancy rate. On-Device The following description is wrong?

Convolutional neural networks are more suitable for image recognition problems than cyclic neural networks.

TensorFlow is Google's first generation of proprietary machine learning system.