Shulou(Shulou.com)11/24 Report--

With the new wave of AI reform set off by large language models such as ChatGPT, the shortage of AI chips is becoming more and more serious. For example, ChatGPT developer OpenAI relies heavily on supercomputers provided by Microsoft and uses a large number of Nvidia GPU. Recently, it has been reported that OpenAI is considering making or acquiring its own AI chips to solve the problem of high-performance and low-cost GPU needed by its business.

Today, let's talk. What on earth is the GPU that OpenAI wants to join?

GPU is "a thousand mobile phones, there are a thousand game experiences". When we play games on different phones, the sense of experience is different. In addition to response speed, the exquisite and three-dimensional of the game screen is also a major difference. One of the factors causing this difference is the different performance of the mobile phone GPU.

GPU (Graphics Processing Unit, graphics processor), as a superhero hidden in mobile phones and computers, is not only a professional painter who can quickly draw colorful pictures, but also a mathematician who can quickly complete a large number of computing tasks.

Professionally speaking, GPU is a special graphics processor, which can process a large number of graphics rendering calculations at a very fast speed, and can handle multiple graphics tasks at the same time, thus greatly improving the computing and processing speed of the computer.

GPU was originally designed for graphics processing, but because of its parallel processing and high-speed computing capabilities, GPU began to play more and more important roles. Now GPU is widely used in scientific computing, machine learning, big data analysis and other fields.

GPU in computers can be divided into two types, integrated GPU and independent GPU.

Integrated GPU, small in size, is generally built into the computer motherboard, and can even be integrated into CPU. Integrated GPU can make computers lighter and more common in laptops.

Independent GPU, larger in size, is an independent component with a special socket on the computer motherboard. It is more powerful than the integrated GPU and can be upgraded separately (replace the video card). However, because of its larger size, it will take up more space, consume more energy and generate more heat when running.

Some laptops have both types of GPU, generally using integrated GPU to save energy and reduce heat, and switch to a separate GPU to perform related tasks when more powerful graphics processing performance is needed.

What's the difference between GPU and CPU? do you think of another important role in the computer, CPU (Central Processing Unit, CPU)? so, what's the difference between GPU and CPU?

Although both can perform computing tasks, they each have their own strengths. If CPU is a knowledgeable mathematics professor who can solve all kinds of problems, then GPU is 10, 000 pupils with large numbers of people and great strength, and he is extremely fast in calculating simple math problems.

In fact, before the advent of GPU, almost all the tasks were done by CPU. With GPU, there is a division of labor between the two, and the following table lists the differences between the two.

Through the above comparison, we find that GPU and CPU have their own advantages. In mobile phones and computers, the two cooperate with each other, divide the work and work together to serve us.

GPU is more suitable for AI. Through the previous introduction, we learned that GPU is very suitable for large-scale parallel computing. The training of AI (Artificial Intelligence, artificial intelligence) just involves a lot of data processing, especially in the field of deep learning, the network model usually has millions or even billions of parameters, and needs to be trained through a large amount of data to obtain accurate prediction, so GPU is very suitable for AI algorithm.

Parallel processing ability GPU has a large number of core and high-speed memory, and is good at parallel computing. In the field of AI, the amount of computation is very large, and GPU is just suitable for this scenario. For example, when you need to calculate a large number of simple math problems, ten thousand pupils must be more suitable than a professor.

Some common GPU memory bandwidths are about 400 GB/s, while the best CPU memory bandwidths are about 50 GB/s, so GPU can get and access data in memory faster. In the AI world, data generally takes up large chunks of contiguous memory space, so it is obvious that GPU is more appropriate.

GPU has strong flexibility to support the use of programming frameworks and languages such as CUDA and OpenCL, so that developers can easily make use of the computing power of GPU, highly customize GPU computing tasks, and provide support for different types of AI algorithms.

CUDA

Compute Unified Device Architecture, Unified Computing device Architecture, a general parallel computing architecture introduced by NVIDIA, enables GPU to solve complex computing problems.

OpenCL

Open Computing Language, an open design language, is an open standard for cross-platform parallel programming of various accelerators in supercomputers, cloud servers, personal computers, mobile devices and embedded platforms.

Strong scalability with the increase of the complexity of the AI model and the growth of the amount of data, we can improve the processing power by adding more GPU, just like adding more primary school students to calculate, so that the system can better cope with the growing computing needs.

This article comes from the official account of Wechat: ZTE documents (ID:ztedoc)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.