Parallel Computing: The Graphics Processing Unit

In the era of parallel computing, the core available on the central processing unit (CPU) continues to grow. However, "Free lunch" is over and the CPU faces the end of the performance improvement which is easily obtained by Moore's Law. Other widely used processing devices, graphic processing units (GPUs) are also affected by Moore's Law. As a result, as graphics standards to promote market demand continue to grow, companies manufacturing GPUs can consistently provide products that meet these standards.

The GPU (Graphics Processing Unit) was originally designed for image processing, but has succeeded in AI. The GPU contains thousands of cores and can handle thousands of threads at the same time. This parallel computing design makes the GPU very powerful in large-scale data computation. AI chip, also known as AI accelerator, is the processor of AI related computing task. With machine learning technology, training algorithms that can not be realized with traditional computing hardware and the computing power of running applications are strongly required. As a result, demand for dedicated AI chips is rapidly growing. The AI ​​chip can be divided into three main application areas: training, cloud reasoning, and edge device inference.

The main catalyst for the advancement of artificial intelligence is to reuse the graphics processing unit (GPU) to train the large neural network model. Unlike a central processing unit (CPU) that calculates sequentially, the GPU provides a massively parallel architecture that can handle multiple tasks simultaneously. Considering that a neural network needs to process large quantities (usually high dimensional data), GPU training is much faster than using a CPU. As a result, since the 2012 AlexNet release, the GPU has grown really fast. This is the first neural network implemented in the GPU. NVIDIA continues to lead the way in 2017 ahead of Intel, Qualcomm, AMD, and recently Google.