China Securities Co., Ltd.: Domestic computing power sector welcomes intensive catalysis, computing power chips usher in a window of opportunity for domestic substitution.

China Securities Co., Ltd. released a research report stating that on August 22, the Shanghai Composite Index rose by 1.45% to surpass 3800 points, and the computing power and chip stocks collectively surged. Recently, domestic computing power sectors have experienced intensive catalysis. On August 13, during Tencent's performance meeting, the company mentioned having multiple choices for reasoning chip supply channels. On August 21, DeepSeek updated its model version to DeepSeek-V3.1, which is expected to support FP8 precision and domestic chips. Recently, server products based on Huawei's Ascend chips have gradually gained large orders in government, financial, and telecommunications industries. The domestic computing power chip market is experiencing a domestic substitution window. Considering that Nvidia's new products have seen significant performance upgrades and have been banned from sales in the Chinese market, the development of domestic computing power chips is urgent. A large number of domestic computing power chip manufacturers have emerged, with Ascend and Cambricon releasing their own AI chips, Hygon Information Technology's DCU gradually gaining recognition, and the domestication process of other supporting links accelerating. Development trend and outlook of domestic artificial intelligence chips: Overseas leading companies dominate the AI acceleration chip market. In the data center CPU market, Intel's market share has decreased slightly but still maintains a significant lead, while AMD continues to gain market share momentum. In the AI acceleration computing chip market, Nvidia dominates due to its hardware advantages and software ecosystem, leading in both training and inference. According to IDC data, by 2024, Nvidia is expected to have a shipment share of 70% in the domestic AI acceleration computing chip market, Huawei's Ascend with a 23% shipment share, and the remaining manufacturers a combined 7% share. While domestic companies started late, they are gradually making efforts, with some breakthrough companies emerging in the acceleration chip sector. Although there is a significant gap between high-end AI acceleration computing chips and overseas manufacturers in the domestic market, domestic companies have started to gain market share. According to IDC data, by 2024, Huawei's Ascend has shipped 640,000 units, Cambricon 26,000 units, and Suiyuan 13,000 units. As the U.S. continues to increase export restrictions on high-end chips to China, the domestication process of AI chips is expected to accelerate. In the GPU market, overseas leading companies dominate, while domestic manufacturers are catching up. Currently, Nvidia, AMD, and Intel dominate the global GPU chip market. Integrated GPU chips are commonly used in desktops and laptops, with lower performance and power consumption, with major manufacturers including Intel and AMD. Independent graphics cards are commonly used in servers, with higher performance and power consumption, with major manufacturers including Nvidia and AMD. In applications such as artificial intelligence, scientific computing, and video encoding, Nvidia and AMD dominate the server GPU market. According to JPR predictions, by Q1 of 2025, Nvidia's standalone graphics card market share (including AIB partner cards) will be 92%, while AMD and Intel will have shares of 8% and 0%, respectively. Graphics rendering GPUs: Nvidia has led the industry for decades and continues to lead with continuous technological iterations and ecosystem development. Since 2006, Nvidia's GPU architecture has been updated approximately every two years, with significant performance improvements in each generation of products. The GeForce series products have long maintained a leading market share, with the latest GeForce RTX 40 series representing the current peak of graphics card performance, using the new Ada Lovelace architecture, TSMC's 5nm process, featuring 7.6 billion transistors and 18,000 CUDA cores, with approximately a 70% increase in core count compared to Ampere. The power consumption is nearly doubled, capable of running DLSS 3.0 technology, far surpassing previous generation products. AMD's independent GPUs have a clear iterative path with the RDNA 3 architecture, which uses a 5nm process and Chiplet design, leading to a 54% performance improvement per watt compared to the RDNA 2 architecture. Currently, the gap between domestic manufacturers in graphics rendering GPUs and overseas leading companies is gradually narrowing. Chip Dynamics' "Fenghuang 2" GPU has a pixel fill rate of 48GPixel/s, FP32 single-precision floating-point performance of 1.5TFLOPS, INT8 AI calculation performance of 12.5TOPS, measured power consumption of 4-15W, supporting OpenGL4.3, DX11, Vulkan, and other APIs, achieving a breakthrough in domestic graphics rendering GPU capabilities. Although Changsha Jingjia Microelectronics lags behind Nvidia's counterparts in process technology, core frequency, and floating-point performance, the gap is narrowing. In 2023, the company successfully released the JM9 series graphics processing chip, supporting OpenGL 4.0, HDMI 2.0, and H.265/4K 60-fps video decoding, with a core frequency of at least 1.5GHz, equipped with 8GB of memory, and floating-point performance of approximately 1.5TFlops, similar to Nvidia's GeForce GTX1050 and potentially competing with the GeForce GTX1080. GPGPU: Nvidia and AMD are currently the leading companies in the global GPGPU market. Nvidia's general-purpose computing chips have excellent hardware design and have realized the generalization of GPU parallel computing through the CUDA architecture and other comprehensive stack software layouts. AMD introduced the Radeon Instinct GPU acceleration chips for data centers in 2018. The Instinct series is based on the CDNA architecture, with the MI250X using the CDNA2 architecture to significantly improve computing and interconnect capabilities in the general computing domain. AMD also launched the open-source AMD ROCm software development platform to compete with Nvidia's CUDA ecosystem. Domestic GPGPU manufacturers are gradually narrowing the gap with Nvidia and AMD. Nvidia leads the market due to the advanced performance of its hardware products and the completeness of its ecosystem. Although domestic manufacturers may have a gap in terms of hardware performance and industrial chain ecosystem compared to leading companies, efforts are being made to improve product layout and ecosystem construction, gradually closing the gap with industry leaders. In the ASIC market, due to its customized nature, the market landscape is more dispersed. ASICs also play a role in the field of artificial intelligence. Google holds a leading position in technology innovation and introduced the Tensor Processing Unit (TPU) designed specifically for machine learning since 2016. In 2025, Google introduced the seventh-generation Tensor Processing Unit (TPU) Ironwood, scalable to 9216 liquid-cooled chips, with breakthrough chip-to-chip interconnections, consuming close to 10 megawatts of power. The TPU v7p chip is Google's first TPU to support FP8 calculations in its tensor cores and matrix mathematical units, having previously supported INT8 format and inference processing as well as BF16 format and training processing. The Ironwood chip also features the third-generation SparseCore accelerator, which first appeared in the TPU v5p and was enhanced in last year's Trillium chip. Domestic manufacturers are rapidly growing, with companies like Cambricon emerging as a rising force. Through product comparisons, it can be seen that domestic manufacturers such as Cambricon, Hisilicon Ascend, and Tsuyuan Technology are continuously improving product performance, efficiency, and usability through technological innovation and design optimization, enhancing product competitiveness. In the future, domestic manufacturers are expected to continue making efforts in the ASIC field, breaking the monopoly of foreign manufacturers in the AI chip market. The ecosystem determines user experience and is the strongest moat for computing chip manufacturers. Although Nvidia's hardware platform offers excellent computing power, its powerful CUDA software ecosystem is the key force behind the popularization of its GPU computing ecosystem. From a technical standpoint, GPU hardware performance barriers are not high, and through product iterations, they can approach leading levels. However, downstream customers are more concerned with the usability of the ecosystem. Before CUDA was introduced, GPU programming required deep machine code integration into the graphics card kernel to complete tasks. With the introduction of CUDA, complex GPU programming was simplified into a user-friendly interface, benefiting developers. CUDA has become the most developed and widely used ecosystem for deep learning and AI training on GPU architectures. Since its introduction in 2007, Nvidia has continuously improved and updated CUDA, expanding into various toolkits and software environments, building a complete ecosystem. Through collaborations with numerous customers, Nvidia has created specialized acceleration libraries and AI training models, accumulating 300 acceleration libraries and 400 AI models. Particularly after deep learning became mainstream, Nvidia has optimized its solutions to achieve the best efficiency and performance. For example, supporting mixed-precision training and inference, adding Tensor Cores to GPUs to enhance convolutional computing capabilities, and introducing the latest Transformer Engine in the H100 GPU to improve performance for related models. These investments include collaborative design efforts in software and chip architectures, allowing Nvidia to maintain performance leadership with minimal cost. Even Nvidia's largest competitor, AMD, still has a gap in user ecosystem and performance optimization with its ROCm platform. As a complete GPU solution, CUDA provides direct access to the hardware, significantly reducing development barriers. This easy-to-use software ecosystem that fully leverages the chip architecture's potential has given Nvidia significant influence in the large model community. Since CUDA has a mature and high-performing underlying software architecture, almost all deep learning training and inference frameworks consider support and optimization for Nvidia GPUs essential, helping Nvidia maintain its leading position continuously. The U.S. continues to strengthen controls over the supply of AI chips to China, with H20 being included in the control range. In 2022, the U.S. BIS implemented export controls, restricting the export of advanced GPUs from Nvidia and AMD. To comply with regulations, Nvidia subsequently introduced the H800 and A800 for the Chinese market, with reduced interconnect bandwidth. In 2023, BIS announced new export control regulations for advanced computing chips, further expanding the restricted range based on "performance density" and "total processing power (TPP)," leading to restrictions on products such as A100, A800, H100, H800, L40, and L40S. Although Nvidia released the H20, which significantly reduced performance and complied with the new regulations, the H20 was included in the export control list by the U.S. in April of the same year.