(Report producer/author: West China Securities, Liu Zejing)
01 Why NVIDIA
Global technology giant, GPU king - NVIDIA
NVIDIA is a global GPU giant. NVIDIA was founded in 1993 and is headquartered in California, USA. The company focuses on GPU development and manufacturing, and in 2009 released the Fermi architecture, establishing its dominance in the gaming field. The company's businesses include data centers, gaming, scientific computing, and autonomous driving. In the field of artificial intelligence, TensorCore acts as a processing unit for deep learning, providing efficient computing and learning capabilities for AI. Excellent software development supports the company's continuous development, and the CUDA platform and deep learning libraries are widely used in scientific research and big data fields.
Since 2016, the stock price has increased by nearly 60 times, and AI has opened a new growth cycle. NVIDIA's total market capitalization as of May 30, 2023 was $963.18 billion, and its closing price was $389.46 per share. According to JonPenddie Reasearch data, in the discrete graphics card market, NVIDIA has an 84% share of the global GPU market. In the past 20 years, NVIDIA's performance revenue has increased by 15 times, and its performance has continued to grow explosively, which we believe is the fundamental reason for the continuous increase in NVIDIA's stock price. In addition, the company began to lay out the corresponding artificial intelligence field in 2015, and in 2019 emerged and gradually became a global AI giant, and now, with the outbreak of large models, NVIDIA as an AI hardware leader, opened the second wave of growth curve.
30 years of the road to king, the rise of AI hardware giants
Fierce competition and courage to break the ice (1993-1998): NVIDIA entered the market in 1993, when the display chip industry was highly competitive. In 1995, NVIDIA launched NV1, but the results were not obvious, and the financial situation was tight. However, in 1997, the company launched the world's first 128-bit 3D processor, the RIVA 128 (accelerated graphics processing chip), which sold more than one million units in the first four months alone. After that, NVIDIA continued its efforts in 1998, releasing two high-performance 3D processors, RIVA 128ZX and RIVA TNT.
Successful IPO and rapid growth (1999-2005): In January 1999, NVIDIA conducted an initial public offering on the NASDAQ Stock Exchange at a price of $12 per share. In August of the same year, the world's first GPU (GeForce 256) was released, defining the GPU as a monolithic processor with integrated transform, lighting, triangulation, cropping, and rendering engines, capable of processing at least 10 million polygons per second. Nvidia is one of the fastest-growing semiconductor companies with $1 billion in revenue and is included in the S&P500 index.
CUDA comes out with an emphasis on ecology (2006-2009): In 2006, CUDA was introduced, a revolutionary architecture for general-purpose GPU computing, enabling scientists and researchers to perform more complex calculations. In 2009, the first full GPU computing architecture (Fermi) was released, in which the Quadro 7000 represented a leap forward, achieving a double improvement in gaming performance and computing performance.
Product: Graphics display and control + computing center acceleration card two-wheel drive
The company's GPU product features are divided into 1) GPGPU for compute & networking; 2) GPU for graphics processing. GPGPU (General Purpose GPU): General-purpose computing graphics processor. In the architecture design, the acceleration hardware unit designed by the GPU for graphics processing is removed, and the SIMT architecture and general-purpose computing unit are retained. The parallel computing power of GPUs can be applied to scientific computing, data analysis, machine learning and other fields to improve computing speed and efficiency. GPU (Graphics Processing Unit): A microprocessor that does image computing. As a separate module, either the discrete graphics core or the motherboard integrated graphics core. It is mainly used to provide high-performance graphics rendering capabilities.
According to the FY2023 annual report, the company's computing and networking GPGPU revenue was $15.068 billion, and graphics processing GPU revenue was $11.906 billion. Among them, 1) Computing and networking revenue +36% year-over-year, mainly used in data center accelerated computing platforms, artificial intelligence cockpits, autonomous driving solutions, electric vehicle computing platforms, NVIDIA AI enterprises and other software, cryptocurrency mining, etc. 2) Graphics processing revenue -25% year-over-year, mainly used in GeForce graphics processors for gaming and personal computers, solutions for gaming platforms, cloud-based vision and virtual computing software, and full-service enterprise software for building and operating 3D Internet applications.
Business: Build a diversified product matrix, with data center and games as the core
The company's software and hardware combined with platform layout to create 4 product lines covering downstream mainstream applications:
Data centers: Revenue of $15.01 billion in fiscal 2023, accounting for 56%. The NVIDIA accelerated computing platform, built on three next-generation architectures of GPU, DPU, and CPU, introduces the NVIDIA DGX AI supercomputer, enabling modern data centers to process workloads involving deep learning, machine learning, and high-performance computing (HPC) faster.
Gaming: Revenue in fiscal 2023 was $9.07 billion, accounting for 34%. Products include GeForce RTX and GeForce GTX graphics processors, cloud gaming GeForce NOW, shielding for streaming, and system-on-a-chip (SOCs) and game console development services. Launch of the GeForce RTX 40 series gaming graphics processors based on the Ada Lovelace architecture in FY2023.
Professional visualization: Fiscal 2023 revenue of $1.54 billion, accounting for 6%. Applied to many leading 3D design and content creation, such as the full universe, virtual reality and augmented reality. Leverage GPUs to power design, manufacturing, and digital content creation. The NVIDIA RTX platform leverages ray tracing to render film quality, photorealistic objects and environments in real time.
Autonomous driving: Revenue of USD 900 million in fiscal 2023, accounting for 3% of the total. Includes AV, AI cockpit, EV computing platform, and infotainment platform solutions. According to the FY2023 Annual Report, the company is working with hundreds of automotive ecosystem partners, including automotive industry chain manufacturers, automotive research institutes, mapping companies, and startups, to develop and deploy AI systems for autonomous vehicles. The launch of Drive is an artificial intelligence vehicle platform covering a variety of autonomous driving fields.
CUDA opens up the software and hardware ecology and forms a moat
GPUs are suitable for processing large data sets, and CUDA cores are the essential reason. Initially, the GPU (Graphics Processing Unit) was a specialized computer processor that could meet the needs of implementing computationally intensive tasks for high-resolution 3D graphics. By 2012, GPUs had evolved into highly parallel multicore systems, giving them the ability to process large amounts of data. In short, the CPU does focus on linear computing, the GPU does parallel computing (there is no direct relationship between the data), and the essential reason is the difference in CUDA cores, the more CUDA cores, the stronger the computing performance, and the number of CUDA cores of the GPU is hundreds of times that of the CPU, such as AMD EPYC 7003 series 7763 cores is 64, while NVIDIA A100 40GB cores are 6912.
The essence of CUDA is "software-defined hardware", which implements "software call hardware". CUDA is a parallel computing platform and application programming interface (API) that allows software to use a specific type of graphics processing unit (GPU) for general-purpose processing, called general-purpose graphics processing unit computing (GPGPU). CUDA provides a software layer with direct access to the GPU virtual instruction set and parallel computing elements for executing compute cores. CUDA-backed GPUs can also use programming frameworks to use HIP by compiling code to CUDA. CUDA integrates a variety of different codes into one go, which greatly speeds up the training of development models. It can be simply understood that CUDA is a kind of "class compiler" of NVIDIA to achieve software and hardware adaptation, converting software code into hardware assembly code, CUDA is NVIDIA's moat to achieve software and hardware ecology.
NVIDIA released a number of AI products this year to help the global AI ecosystem
At the GTC Conference on March 23, 2023, NVIDIA's new AI-related products helped the global AI ecosystem. 1) Basic software: launched a new acceleration library; 2) Chips: Launched the data center Grace CPU, which has the advantages of high energy efficiency and high running speed; 3) Server: Launch of DGX supercomputer; 4) New AI service platform (DGX cloud and generative AI service), the "iPhone" moment of AI has arrived, AI foundations cloud service can build, improve and operate customized large-scale language models and generative AI models, helping startups have the ability to have generative AI, and already have a variety of generative AI models and corresponding cases.
The platform is actually a "bridge" between the model and computing power, which is a necessary element for AIGC or large model generation, whether it is a database or a compiler, it is necessary to achieve the rational allocation of resources through the platform to achieve the optimal combination of software and hardware, thereby greatly improving the efficiency of the model. The platform adapts the structure between software and hardware by calling data packets to achieve the optimal combination of models, thereby improving the efficiency of the model and even the entire virtual machine.
02 It is imperative that AI hardware be autonomous and controllable
NVIDIA successfully transformed into a global AI hardware leader
NVIDIA has successfully transformed from a graphics processor company to a comprehensive hardware company that explodes artificial intelligence. GPU technology plays an important role in AI applications, and its parallel computing capabilities are leading the way. AI platform—NVIDIA DGX systems are a high-performance computing solution that integrates NVIDIA's GPUs and software tools to make it easier for developers to build, train, and deploy AI. In terms of software, NVIDIA CUDA and NVIDIA cuDNN were launched, providing fast neural network training and inference capabilities. NVIDIA's technology is widely used in various fields such as intelligent driving, deep learning, financial services, healthcare, game development, and construction engineering in the field of AI.
Healthcare: Accelerate drug discovery, with NVIDIA introducing NVIDIA Clara Discovery's AI-accelerated computing platform to support cheminformatics research, protein structure prediction, drug screening, and molecular dynamics. Launch of the large language model BioNeMo can be used to train and deploy large biomolecular language models. Updating medical equipment and using NVIDIA's powerful GPU analysis and imaging technology to create a variety of medical devices for imaging diagnosis, digital surgery, and patient monitoring, greatly improving treatment accuracy and efficiency.
Game development: NVIDIA game technology for game development, such as NVIDIAACE, an intelligent game character casting platform supported by generative AI, NVIDIA DLSS, an AI neural graphics technology, and a scalable multi-GPU real-time inference development platform (for 3D policy and design) - NVIDIA Omniverse platform. These platforms help game development build realistic and accurate games at this record speed.
Architecture: Accelerate the design process and increase productivity, with NVIDIA RTX-powered workstations augmenting building and infrastructure design workflows with real-time ray tracing, virtual reality, engineering simulation, and AI-enabled applications. The Omniverse platform overcomes the problem of operational conflicts between design software. NVIDIA Quadro GPUs can quickly analyze complex fluid scenarios, accelerated using NVLink technology, which can greatly reduce the time of design, modeling, simulation, and inspection.
It has been repeatedly emphasized that computing power is bound to usher in an outbreak under the background of large models
ChatGPT opens the computing power arms competition: We have mentioned in "ChatGPT: Baidu Wenxin One Word" that data, platform, and computing power are the necessary foundation for building a large model ecology, and computing power is the underlying power source for training large models, and an excellent computing power base has efficiency advantages in the training and reasoning of large models (AI algorithms); At the same time, we mentioned in "ChatGPT Starts the AI Computing Power "Armament War"" that computing power is the "ticket" to AI technology competition, of which AI servers and AI chips are core products; In addition, we also mentioned in "ChatGPT, NVIDIA DGX Detonates AI "Nuclear Fusion" that technology companies represented by NVIDIA are rapidly replenishing the global demand for AI computing power and adding necessary "fuel" to large models.
Large model parameters show exponential scale, detonating massive computing power demand: According to CaiLian News and OpenAI data, there is a huge gap in computing power under the wave of ChatGPT, and according to OpenAI data, the growth rate of model computing volume far exceeds the growth rate of artificial intelligence hardware computing power, and there is a 10,000-fold gap. The growth of computing scale has driven the demand for the single-point computing power of AI training chips, and put forward higher requirements for data transmission speed. According to Zhidong data, in the past five years, the development of large models has shown an exponential level, and some large models have reached the trillion level, so the demand for computing power has also increased.
The United States restricts the flow of high-end chips into China, seriously interfering with the development ecology of domestic large models
The U.S. government banned NVIDIA and AMD from exporting top-tier computing chips for artificial intelligence to China. According to titanium media, in September 2022, the US Department of Commerce announced restrictions on the export of advanced computer graphics processors (GPUs) to China by American companies such as NVIDIA and AMD, which mainly restricted the export of NVIDIA's A100, H100 high-end chips and AMD's MI250 to China, with the aim of targeting domestic advanced computing for containment and affecting the development of domestic artificial intelligence.
Inspur Group was added to the "Entity List": According to titanium media, in March 2023, the U.S. Department of Commerce released Inspur to be added to the U.S. "Entity List", restricting the support of U.S. technology companies for LaoChao's technology and products. Inspur's server business is highly dependent on foreign manufacturers in key chip technology vendors such as CPUs and GPUs, in addition, as of the end of 2022, Inspur servers and components accounted for 99.17% of the total revenue, if the sanctions are strictly implemented, its server business will be seriously stagnant.
The industry side responded positively, and the construction of intelligent computing power continued to accelerate
Beijing Ascend Artificial Intelligence Computing Center officially lit up: Beijing Ascend Artificial Intelligence Computing Center was officially lit, which will promote the high-quality development of Beijing's artificial intelligence industry. The intelligent computing center adopts Ascend AI basic software and hardware to fully release hardware computing power and accelerate the innovative application and model incubation of artificial intelligence enterprises.
Guizhou Provincial Big Data Bureau issued the "Nationwide Computing Power Support Base Construction Plan": The overall goal is that by 2025, the construction task of the nationwide computing power support base will be fully completed, the position of Guizhou's ultra-large-scale data center cluster will be more consolidated, the storage-computing ratio will be more reasonable, the infrastructure layout, structure, function and system integration will be optimized, the data center will achieve intensive, large-scale and green development, network interconnection, energy security and reliability will be improved to a new level, and an internationally competitive digital industrial cluster will be created.
"Sharing AI capabilities and computing power", the demand for AI cloud is increasing
It is difficult to deploy generative AI applications, and AI cloud provides platform customization capabilities. With the rise of the artificial intelligence industry brought by large models, AI applications such as text generation, automatic customer service, automatic driving and other fields are expanding rapidly. It is very difficult for most enterprises to deploy such capabilities on their own, we believe that one is due to the current computing power gap and the high marginal cost of training, and the other is that the full-chain deployment application requires deep software and hardware combined with ecological technology. AI Cloud can integrate the AI capabilities of professional suppliers such as NVIDIA into the cloud, allowing enterprises to directly access applications or pre-train from the base layer to form their own models and applications.
The demand for AI cloud is growing rapidly, and the cloud computing power revolution has begun. Under the continuous upward trend of enterprises' demand for AI capabilities such as large model training and artificial intelligence application deployment, AI cloud products are favored by the market. Platform companies represented by Alibaba and Tencent pay more attention to the construction of general clouds in the cloud market layout; Huawei and Dawning are more from the perspective of hardware to join the cloud market layout. We believe that AI cloud players (NVIDIA, Capital Online) that combine software and hardware meet the needs of the market. At the same time, the high demand for AI cloud also means that enterprises that master intelligent computing cards in the future will continue to occupy the market highland, and we insist that enterprises with computing power have a first-mover advantage in the next stage of application and platform competition.
Excerpts from the report:
(This article is for informational purposes only and does not represent any investment advice from us.) For information, please refer to the original report. ）
Selected report source: [Future Think Tank]. 「Link」