As machine learning and artificial intelligence technologies advance, the rapid progress of intelligent applications has spurred semiconductor companies to innovate and introduce specialized processors and accelerators like TPUs and CPUs. While NVIDIA GPUs have traditionally been the go-to choice for deep learning practitioners, the landscape is now poised for transformation with the emergence of Google's TPU chip. In this article, we will delve into a comparison between TPUs and GPUs. However, before we explore the details, it's essential to grasp several key points.

The graphics processing unit refers to the key module of the 3D engine. The commands processed in this module can be divided into 6 types: configuration information commands, stack operation commands, buffer clearing commands, cumulative buffer commands, fragment data and commands not used in this unit. Therefore, this module is divided into 6 parts to achieve, as shown in "CAC module". The command parsing part is responsible for the preliminary parsing and dispatching of commands, the configuration information part is responsible for the parsing and saving of configuration information commands, the stack operation part is responsible for the parsing and clearing of stack operation commands and parameter access, the clear part is responsible for the parsing and clearing of buffer commands, and the accumulation part is responsible for the parsing and accumulating of cumulative buffer commands. The reading part is responsible for reading the target information required by this module and subsequent modules, and the sending part is responsible for sending processed data and commands that are not processed by this unit.

A graphics processing unit (GPU) is an electronic circuit that can perform mathematical calculations at high speeds. Computational tasks such as graphics rendering, machine learning (ML), and video editing require similar mathematical operations to be applied to large data sets. Gpus are designed to perform the same operation on multiple data values in parallel. This can make many compute-intensive tasks more efficient. This article will introduce what is GPU, and the origin of GPU, development history