Deep learning is a field with high requirements for computing. The selection of GPU will fundamentally determine our deep learning experience. When individuals start deep learning, having a fast GPU is essential because it can quickly gain practical experience. In other words, it is the key to building professional knowledge, and we can use this knowledge to apply deep learning to new problems.
Why GPU not CPU ?
The CPU is designed for more traditional computing workloads. Although GPUs are less flexible, they are designed to compute the same instructions in parallel. The structure of a deep neural network (DNN) is very uniform. In other words, in each layer of the network, thousands of identical artificial neurons perform the same calculations. Therefore, the structure of DNN is very suitable for various calculations that the GPU can perform efficiently.
In comparison, GPU has more advantages than CPU, including more computing units and higher bandwidth to retrieve from memory. Besides, in applications that require image processing (ie, convolutional neural networks), GPU graphics' specific functions can be used to speed up the calculation further.
Compared with CPU, GPU's main disadvantage is that the memory capacity on GPU is lower than the CPU. For instance, the highest known CPU contains 1TB RAM, yet the highest known GPU can only reach 24G RAM.
The next disadvantage is that the GPU will need to transfer the data to the CPU. It will need to process through the PCI-E connector, and the speed is much slower than the CPU or GPU memory.
The last weakness is that the GPU clock speed is 1/3 of the high-end CPU, so in sequential tasks, the GPU has a lower performance.
Why GPU can work well with DNN calculation?
1) GPU has more resources and faster memory bandwidth
2) The DNN calculation is very consistent with the GPU architecture. Computational speed is essential because in-depth neural network training may take several days to several weeks.
Hence, if it was not for GPUs' availability, many successes of deep learning might not have been discovered.