With the rapid development of deep learning technology, especially the recent rise of large language models and AIGC technology, more and more people have joined the work of deep learning, and the graphics processing unit (GPU) has become the preferred hardware for deep learning research and application because of its parallel processing capabilities and efficient matrix computing performance. However, in the face of the numerous GPU options on the market, how to choose the one that best suits your needs? Should I choose the latest and most expensive model? Or is there a more cost-effective option? These questions can be confusing for beginners and researchers. Recently, I read a blog by Tim Dettmers (QLoRA), detailing the working principle of GPUs, comparing the performance and cost performance of different GPUs, and giving specific suggestions and guidance according to the deep learning projects and budget needs of different users to help you make an informed GPU choice. Also, when it comes to choosing cloud services and GPU desktop servers, Tim's advice is: If you expect to do deep learning for more than a year, it is cheaper to buy a GPU desktop. Otherwise, cloud services are a better choice unless you have extensive cloud computing skills and want to enjoy the benefits of scaling the number of GPUs up and down at will.
