GPU (Graphic Processing Unit) is the graphics processing unit, which is the cornerstone of the generation of graphics content in the digital world, whether it is desktop applications, games, movies, digital twins or the metaverse; at the same time, powerful parallel computing capabilities have become a variety of application acceleration The mainstream method is being widely used in scientific computing and artificial intelligence. However, my country’s GPU chips basically rely on imports. As the United States continues to impose restrictions on Nvidia and AMD's supply of high-end GPU chips to my country, my country's import of GPUs is also subject to great restrictions. These multiple factors have further prompted domestic GPU companies to catch up.
On June 15, the appraisal and press conference of the domestic independent GPU architecture "Sirius" was held in Beijing.
Industry leaders attending this event include: Dr. Yan Qun, Chinese president and fellow of the International Society for Information Display, foreign academician of the Russian Academy of Engineering, and chairman of the Beijing branch of the International Society for Information Display; Guo Yiwu, Secretary General of the Shanghai Integrated Circuit Industry Association; former National Energy Administration Electric Power Hu Hongsheng, deputy director of the Reliability and Quality Supervision Center; Liang Jingdong, executive director of the State Investment Commission of the State Development and Investment Group; Chen Hong, general manager of Hangzhou Shangcheng District State-owned Capital Operation Group Co., Ltd.; Zhang Jiahui, investment service center of Huzhou South Taihu New District; Lu Yuqian, OPPO Investment Department Shan; Cheng Quanfu, founder of Quantum Innovation (Beijing) Information Technology Co., Ltd., and Guo Song, vice president; Xu Yanmao, vice president, and Xiao Zhigang, vice president of Beijing Ruima Video Technology Co., Ltd.; Wang Miaowei, vice president of Beijing Deleng New Journey Technology Co., Ltd.
•GPU architecture “Sirius” is born
The GPU architecture "Sirius" was independently developed by the domestic company Advanced Technology Stellar (ATS).
The Zhongtian Xingxing R&D team is led by Dr. Deng Yangdong, Ph.D. in Electrical and Computer Engineering from Carnegie Mellon University, Associate Professor of the School of Software, Tsinghua University, Institute of Microelectronics, and NVIDIA Co-Professor. Dr. Deng is one of the earliest researchers of GPU general computing and is known as the "GPU general computing pioneer." Dr. Deng has been engaged in graphics processor architecture, parallel computing research and chip product development for a long time; he once designed the world's first FPGA-based GPU simulation platform; his research results have been published in top conferences and journals such as ISCA and MICRO. Dr. Deng has written several textbooks and monographs, among which "Structured Design and High-Level Synthesis of Digital Integrated Systems" was selected as a textbook for graduate students of Tsinghua University and several universities, and "Introduction to OpenCL Programming for Heterogeneous Processors" is the first in China for GPU heterogeneous computing. A textbook, "3-Dimensional VLSI" is the first monograph on three-dimensional integrated circuits.
The "Sirius" GPU architecture has several highlights: First, it has a 3D graphics engine + 2D graphics acceleration + video engine. The second is the independently controllable/flexibly optimized instruction set and the VLIW/SIMD machine instruction set (ICCD'13); among them, the independently controllable/flexibly optimized instruction set ensures the software compatibility of GPU chip iterations. The third is the SIMT computing framework, which fully utilizes the data parallelism of graphics applications (DATE'12.ICCD'13, TVLSI15). Fourth, it supports physically realistic rendering (ACM Computing Survey'14, SIGGRAPHASIA'14.15). The fifth is Shader Core (Graphics Processing Cluster), including SIMT independent instruction execution unit, unified graphics architecture (ICCD13) based on 32-bit floating-point ALU, and integrated register file, texture/data cache. The sixth is delayed aggregation global thread scheduling technology (ISCA20, TPDS21, TCAD'21). Seventh, on-chip interconnection architecture with good performance scalability (MICRO'20, TPDS'21).
The Sirius released this time has three main technical advantages: first, facing the order market of 100 million orders, with the mainstream products of discrete graphics card GPU chips as a breakthrough, aiming at the rigid demand market of 100 million orders with broad demand and strong growth, highlighting the cost-effective advantage; It has completely independent intellectual property rights and a self-developed core structure based on more than 10 years of research at Tsinghua University to ensure independent and controllable product iterations. Based on basic theoretical research, starting from the derivation of mathematical formulas, all aspects of architecture design, algorithm model, principle verification, hardware implementation, and driver development are all forward-designed. The core architecture has complete intellectual property rights, and hundreds of patents and copyrights have been applied for, dozens of which have been authorized; relevant research results have been published in top conferences and journals such as ISCA, Micro, IEEE TPDS, and IEEE TCAD; third, complete delivery capabilities, The upstream and downstream industry chains are fully prepared to ensure independent and controllable product mass production; the chip design has been fully verified to ensure successful tape-out; the software and hardware interfaces comply with international standards to ensure that the chip can be used after installation, including operating system certification: Windows WHQL; API certification : Open GL Conformance Test; Peripheral interface certification: HDMI DP CTS; Professional test certification: PHY layer and LIINK layer; Quality system certification: Graphics card 3C certification (China), Japan VCCI, EU CE, US FCC.
•Highly recognized by industry experts
Currently, Imagination and Vivante are the main IP sources. These IPs are mainly GPU cores for mobile applications and are not suitable for desktop applications. In addition, it is difficult to form a mature graphics card-level GPU by purchasing third-party IP, and the core circuit patents cannot be controlled and iterated independently.
Dr. Deng Yangdong, co-founder and chief architect of Zhongtian Stellar, said: "Many domestically produced GPU IP licenses are mainly purchased from third parties. Zhongtian Stellar's route is different. The core graphics engine is completely designed independently, so the 3D graphics engine This does not involve other people’s intellectual property rights, and there are no IP issues. It is all owned by us, and the realization of the complete architecture from scratch is all our own technology.”
There are extremely high technical challenges in the self-developed GPU architecture. Deng Yangdong analyzed and pointed out: "There are several aspects of GPU design that are very complicated: One is the architecture-level design, which is actually an art of overall planning. The resources of the GPU itself, in the so-called uni version of the shader, all computing resources are the same. Both use 32-bit or 64-bit floating-point units, which is a matter of coordination. There are many cores sharing L2 cash, and each core has its own computing unit, as well as various channels leading to off-chip , in fact, there are some on-chip for graphics, such as texture caches; although it is easy to know what kind of modules these are, the difficulty is how to cooperate to achieve the best overall performance. In other words, how to adapt to the absolute Most graphics applications. In most cases, a balanced design maintains performance at least 30 frames per second, which is a very challenging place. Second, the architecture simulation takes a very long time, and requires experience and intuition to find out Where does the architectural problem lie? This aspect is also a big challenge, and it takes many years of accumulation to do this."
The GPU architecture "Sirius" has been highly recognized and praised by industry experts.
Regarding the Sirius architecture, Dr. Yan Qun, Chinese president and fellow of the International Society for Information Display, foreign academician of the Russian Academy of Engineering, and chairman of the Beijing branch of the International Society for Information Display, pointed out: “Now all displays can be called passive display technology, and the viewer’s information on it Transmission is only reception, no interaction. Once the display screen is equipped with many GPU functions, and there is an immersive three-dimensional image presentation and interactive experience, it will no longer be a traditional TV. No one watches TV now, especially young people. Very few people watch it. TV, this experience has been completely replaced by portable devices such as mobile phones and PADs. There is no need to watch such a big TV. But once we need an interactive experience, we will return to the big screen and these renderings. It’s a real, real experience that you can be a part of, and it’s incredible.”
Dr. Yan Qun said that ChatGPT is very popular now. We think this artificial intelligence is still in the kindergarten and elementary school stage, but this is already terrible. If there is interaction, the soul will be in the data in the future. The big data captured at this time, It's not chat generative, but interactive GPT. At this time, after the big data is learned by artificial intelligence, it will far exceed the intelligence and ability of human beings. This is the goal of the real metaverse. "
"I think now is a really good opportunity to seize it. The trend is also moving in this direction, and there are many opportunities. If we can gradually move up, we may not necessarily lag behind some countries in the West, because The points you grasp are higher-level points." Yan Qun shared.
Guo Yiwu, secretary-general of the Shanghai Integrated Circuit Industry Association, believes: "Now is a very good juncture, that is, the rise of the industrial revolution and the new technology revolution. You have also seen cloud computing, digitization, smart cars, cloud-to-edge, etc. After industrialization, large computing power is required, which is what we call GPU. Therefore, the time for us to catch up is very good. In the future, with the development of new technological industrial revolution, there will be a very large space in this aspect."
"Zhongtian Xingxing's self-developed architecture has the following characteristics: First, 3D graphics rendering. Second, the independent controllability of the instruction set, which is not easy. Third, its framework structure, including the entire memory DDR4 realizes high-speed storage. This design is also a relatively advanced architecture. Zhongtianxingxing has used it, which determines that future products will be very versatile. Zhongtianxing’s entry point is in the display field. I think the display field is very broad. I believe Under the leadership of Mr. Huang, Zhongtian Xingxing will definitely move from architecture to products, and finally enable our products to be launched globally." Guo Yiwu pointed out.
•Overcome many challenges and finally "show your sword"
Relying on the research results of the R&D team for many years, Zhongtian Xingxing starts from the derivation of mathematical formulas, and all links such as architecture design, algorithm model, principle verification, hardware implementation and driver development are all forward-designed. The core IP is completely independent and controllable, and it has complete intellectual property rights of graphics GPU. , and has applied for hundreds of patents and copyrights, 25 of which have been authorized, and related research results have been published in top conferences and journals such as ISCA, Micro, IEEE TPDS, and IEEE TCAD.
In 2019, the design verification of the first-generation "Sirius" architecture chip was completed. In 2021, the first-generation "Sirius" architecture GPU was born; in 2022, the second-generation GPU architecture "Arcturus" was defined; in 2023, the first-generation "Sirius" architecture GPU was mass-produced.
The name of the GPU architecture "Sirius" has a unique meaning. Dr. Huang Yong, founder of Zhongtian Star, pointed out: "All our architectures are named after stars. Stellar means stars, stars, and constellations. By extension, it means super-class, The meaning of excellence. The second generation architecture is named Arcturus; Arcturus is the second brightest star. Just because it is further away from the earth, it does not appear to be as bright as Sirius; in fact, Arcturus is brighter than Sirius, and its brightness is that of the sun 110 times."
It is reported that in 2024, Zhongtian Stellar will continue to optimize GPUs based on the "Sirius" architecture. In 2025, the second-generation GPU architecture "Arcturus" will be mass-produced.
The launch of the domestically developed "Sirius" GPU architecture will undoubtedly push domestic GPU chips to a new height.
Currently, the U.S.’s crackdown on China’s technology continues, and its pressure on China to restrict the supply of high-end GPU chips will not weaken. Domestic domestic substitution is in a period of in-depth advancement. The domestic downstream application market has greater autonomy and urgency in purchasing domestically independent and controllable chips. This further stimulates the enthusiasm of Chinese enterprises to self-research GPU chips and provides huge opportunities for the development of domestic GPUs. of assistance.