Sam Ultraman, CEO of the Open Artificial Intelligence Research Center (OpenAI) in the United States, and others believe that artificial intelligence (AI) will fundamentally change the world economy, and having a strong supply of computing chips is crucial. Chips are an important factor driving the development of the AI industry, and their performance and computing power directly affect the progress and application prospects of AI technology. The website of the British journal Nature recently reported that engineers are competing to develop cutting-edge chips, including graphics processing units (GPUs), to meet the computing needs of future AI. GPU accelerates machine learning and computing speed. GPU is Nvidia's iconic computer chip. Traditional central processing units (CPUs) process instructions in order, while GPUs can process more instructions in parallel, allowing for distributed training of programs, greatly accelerating the computational speed of machine learning. In 2022, NVIDIA's Hopper Superchip defeated competitors in all categories including image classification and speech recognition on MLPerf. MLPerf is one of the most authoritative and influential AI benchmark tests internationally, known as the "Olympic Games in the AI world". In March of this year, Nvidia officially showcased its next-generation AI chip Blackwell, which has better performance. It has 208 billion transistors and is Nvidia's first GPU to adopt a multi chip packaging design. With the development of technology, GPUs are becoming larger and larger. If they cannot be larger, more GPUs are combined together to become larger virtual GPUs. Blackwell integrates two GPUs on the same chip, and the new architecture will gradually build larger AI supercomputing clusters through chip to chip connectivity technology. If you want to train a GPT model with 1.8 trillion parameters, it requires 8000 Hopper chips, consumes 15 megawatts of energy, and takes 3 months. If using Blackwell chips, it only takes 2000 pieces and consumes 4 megawatts of energy to complete tasks in the same amount of time. The AI chip market continues to grow, and Nvidia currently supplies over 80% of its products. In 2023, the company sold 550000 Hopper chips. Recently, the company's market value exceeded $3 trillion for the first time, surpassing Apple and becoming the second highest company in the world after Microsoft. Although GPUs have always been the core of the AI revolution, they are not the only "protagonists" in the emergence of various chips. With the surge of AI applications, the variety of AI chips is also on the rise, and Field Programmable Gate Arrays (FPGAs) can be said to stand out. FPGA is a hardware device widely used in the fields of computing and digital circuits. It has become an ideal choice for various applications such as embedded systems and high-performance computing processing due to its unique programmability and flexibility. This is like building Lego bricks. Engineers can build FPGA circuits one by one into any design they can imagine, whether it is a washing machine sensor or an AI used to guide autonomous vehicle. However, compared to AI chips with non adjustable circuits (such as GPUs), FPGA runs relatively slower and has lower efficiency. But FPGA is very useful for processing certain tasks, such as data generated by particle colliders. David Salvatore, Product Marketing Director of NVIDIA Accelerated Computing Group, pointed out that the programmability of FPGA is also helpful for prototype design. The Tensor Processing Unit (TPU) is a chip specially designed by Google for neural network machine learning, aimed at performing matrix calculations and tensor operations. TPU, as an accelerator for Google's deep learning framework TensorFlow, was first launched in 2016. Its design goal is to provide low-power and efficient matrix operations to meet the needs of large-scale machine learning and neural network training. TPU achieves a good balance between performance and energy efficiency. Their power consumption is relatively low, which is crucial for large-scale data centers and applications on mobile devices. In addition, the metaverse platform is also independently developing its own chips. Google, Intel, and Qualcomm have established the UXL Foundation to develop software and tools that support multiple AI accelerator chips to compete against Nvidia's GPUs. Of course, the rise of AI chips such as GPUs does not necessarily mean the end of traditional CPUs. It has become a trend for the two to complement each other's strengths and weaknesses. For example, there is a version of Blackwell chip that allows GPU and CPU to work together; One of the world's most powerful supercomputers, Frontier, located at Oak Ridge National Laboratory in Tennessee, USA, also relies on a combination of CPU and GPU to perform high-performance computing. Given the revolutionary changes in the chip industry over the past decade, engineers may find it difficult to predict the future of chips. In the future, optical chips or quantum computing chips that use light instead of electrons may be developed, and further improving chip performance will accelerate the application of AI in the scientific field. (Lai Xin She)