In recent years,many chip companies have gone bankrupt.Last year,as many as tens of thousands of chip companies closed their doors,with an average of 31 companies going under each day.This year remains grim,with the bankruptcy and liquidation of Huaxia Core,a rising star in China's IP design,and Li Suan Technology,which burned through 300 million RMB,erupted in wage arrears in June last year,and nearly went bankrupt on the eve of tape-out due to the withdrawal of US dollar funds.
Li Suan Technology was established in 2021,and its actual operations began at the start of 2022 with the establishment of Li Suan Shanghai.Li Suan Shanghai is a wholly-owned subsidiary of Nanjing Li Suan Technology and is the core of the company.The three founders,Xuan Yifang,Kong Dehai,and Niu Yixin,all possess profound technical capabilities and engineering experience from their time at the Silicon Valley GPU company S3.
According to the official website,Li Suan Technology insists on self-developed architecture and independent intellectual property rights.
It is developing domestic GPU chips that are benchmarked against international mainstream products,serving the domestic 200 billion RMB rendering GPU market in all aspects.It aims to achieve graphic rendering on the end,cloud,and edge,meeting the needs for chips and computing power in learning,work,entertainment,metaverse,games,animation,digital twins,film and television production,ARVR,smart factories,professional design,smart cockpits,and smart medical applications.Fortunately,Dongxin Shares recently announced in a notice its intention to use its own funds or over-raised funds to increase its capital in Li Suan Technology (Shanghai) Co.,Ltd.,with a shareholding ratio not exceeding 40% and an amount limit of 200 million RMB.The bankruptcy turmoil of Li Suan Technology reflects the intensifying competition in the GPU chip race.
01
The Origin of GPU
The full name of GPU is Graphics Processing Unit,which is a specific chip originally used for drawing images and processing graphic element data.
Later,it added many other functions.Compared to the CPU,the GPU is more suitable for intensive data processing.From the perspective of chip microarchitecture,the GPU is optimized and adjusted based on the original CPU structure.By integrating with the CPU,it shares RAM while adopting the SIMD architecture,constructing hundreds or thousands of logical computational units to achieve parallel processing at the same time.GPUs are particularly adept at handling many tasks simultaneously,such as converting data within a computer into the exquisite images we see on our screens.In computational tasks,GPUs leverage their parallel computing advantages to accelerate the processing of large-scale data-intensive tasks like deep learning training and scientific computations.
In 1962,the thesis and drawing board program of Ivan Sutherland,a Ph.D.student at the Massachusetts Institute of Technology,laid the foundation for computer graphics.From 1962 to 1984,there was no dedicated graphics processing hardware,
and all graphics processing tasks were completed by the CPU.As computers evolved,the demand for image processing gradually increased.The history of GPUs can be traced back to the graphics display controller.The world's first personal computer,the IBM 5150,was released by IBM in 1981,equipped with a monochrome display adapter (MDA) and a color graphics adapter (CGA),which were the earliest graphics display controllers.In 1984,the American company SGI introduced a high-end graphics workstation for professional fields.From 1984 to 1995,SGI continued to develop a series of graphics workstations with better performance.
In 1995,3DFX released a consumer-level 3D graphics card called Voodoo.At this time,the graphics display hardware market had already begun to heat up,with AMD,ATI (acquired by AMD in 2006),and NVIDIA all starting to launch their own graphics card products.The CPU was relieved of some graphics processing tasks,but tasks such as vertex transformation still needed to be completed by the CPU.
NVIDIA,established in 1993,was the proponent of the GPU concept.In 1999,NVIDIA released the GeForce 256 graphics chip,which introduced the concept of the GPU for the first time,marking the advent of the GPU era.The GeForce 256 utilized "T&L" hardware,cubic environment mapping,and vertex blending,among other advanced technologies.
During this period,various hardware units formed a graphics processing pipeline,with each pipeline level having a fixed function and no programmability for the hardware.As technology advanced,GPUs evolved from being able to work in fixed ways to being programmable intelligent chips capable of creating more complex visual effects.In 2003,new products released by NVIDIA and ATI both featured programmable vertex processing and programmable pixel processors,offering good programmability.
Entering the 21st century,with the popularization of gaming and multimedia applications,the performance demands for GPUs grew rapidly.Major chip manufacturers invested heavily in the research and development of new generations of GPUs,
continuously improving their performance.During this stage,the internal composition of GPUs also underwent significant changes,including graphics memory controllers,compression units,BIOS,graphics and computing arrays,bus interfaces,power management units,and video management units.With the rise of artificial intelligence and deep learning technologies,the application fields of GPUs have been further expanded.NVIDIA developed the CUDA technology,enabling GPUs to do more than just graphics,such as processing big data.In addition to traditional graphics rendering,GPUs are also widely used in deep learning,machine vision,and big data analysis.This trend has driven the further development of GPU technology,gradually evolving from a dedicated graphics processor to a general-purpose computing platform.
Domestically,a number of companies have emerged that are developing GPUs: Jingjia Microelectronics,Xinyuan Microelectronics,and Biren Technology,among others.
It was not until April 2014 that Jingjia Microelectronics successfully developed the first domestically produced high-performance,
low-power GPU chip—JM5400.In August 2018,Jingjia Microelectronics' self-developed next-generation high-performance,high-reliability GPU chip—JM7200—was successfully taped out.In 2019,based on the JM7200,Jingjia Microelectronics launched a commercial version—JM7201—to meet the high-performance display needs of desktop systems and fully support domestic CPUs and operating systems.In 2020,the chip industry experienced a surge in entrepreneurship,this time mainly focused on the GPU sector.Domestic GPUs have seen rapid development with the support of capital,and a large number of companies have entered this field.In addition,central and local governments have continuously introduced policies to support the development of the integrated circuit industry.Under such comprehensive influences,it can be said that GPUs have ushered in their best era.These startups are primarily dedicated to developing GPUs that can meet both graphic processing and AI computing needs,also known as GPGPUs (General-purpose GPUs) in China.However,as financing funds run out,some companies are also on the brink of bankruptcy.According to IDC data,the global AI chip market will reach $35.2 billion in 2022.GPUs hold the largest share,and Goldman estimates that by 2025,the GPU share will reach 57%.Under the circumstances where the United States restricts China's import of the latest NVIDIA GPUs,the prospects for domestic GPUs are broad.
02
Global GPU Market Forms a Tripartite Oligopoly Competition Pattern
Looking at the overall global GPU market,a tripartite oligopoly competition pattern has been formed.NVIDIA remains the absolute leader in the global independent GPU market.Intel,benefiting from its advantages in the notebook and traditional PC industries,now dominates the integrated GPU market,while the independent GPU market is mainly occupied by NVIDIA and AMD.
NVIDIA is currently the largest independent graphics card producer and seller,with consumer-oriented GeForce series and professional Quadro series.Its GPUs feature CUDA general computing,
PureVideo high-definition video technology,PhysX physics acceleration,and Optimus smart graphics card switching.In March of this year,NVIDIA launched a new generation of AI graphics processor chip (GPU) architecture,Blackwell,and released the B200 and GB200 product series based on this architecture.It is understood that Blackwell can support models with up to 100 trillion parameters for AI training and real-time LLM inference.The B200 is composed of two ultra-large dies (wafers),containing over 208 billion transistors,more than double the previous generation's 80 billion transistors,and the entire chip is also packaged with 192GB of high-speed HBM3e memory.Intel is the world's largest integrated GPU supplier.Intel's GPUs are mainly integrated graphics cards used in Intel's motherboards and laptops.Intel's first-generation independent graphics card was released in 2022,roughly two years ago.Recently,media reports have indicated that Intel has added code related to the bmg_g21 (Battlemage G21,
BMG-G21) core to the LLVM documentation of the oneAPI DPC++ compiler.This is the first time that Battlemage architecture-related code has appeared in the documentation,suggesting that the Battlemage G21 GPU is expected to be the first model of Intel's next-generation discrete graphics card.AMD,founded in 1969,is currently a rare semiconductor company in the industry that can provide high-performance CPUs,high-performance independent graphics GPU,and motherboard chipsets.In 2006,AMD acquired ATI for $5.4 billion,becoming the first semiconductor company capable of designing high-performance CPUs and GPUs simultaneously.AMD has consumer-oriented Radeon series and professional FireGL series,among others.Its GPUs feature Stream general computing,ATI Video Converter video transcoding,UVD high-definition video technology,and three physics engines: Havok,Bullet,and Pixelux DMM.AMD's GPUs are divided into two distinct product areas,one for gaming and the other for supercomputers,big data analytics,
and machine learning systems.In 2023,AMD officially released the new generation of AI/HPC accelerators,Instinct MI300X and Instinct MI300A,one being a traditional GPU accelerator and the other a unique CPU+GPU fusion accelerator.The MI300A is positioned for HPC products,so the product form is a CPU+GPU/APU combined Chiplet (24 Zen4 cores and their I/O,128G HBM3,packaging 13 Chiplets),and has also transformed the underlying IF high-speed interconnect and UMA memory architecture.The first purchaser is also the United States National Supercomputing (EI Capitan Exascale).In response to NVIDIA's CUDA,AMD has developed ROCm.The domestic GPU market is flourishing with a variety of options.
Although the domestic GPU industry started late,it has benefited from recent capital support and the market catalysis of artificial intelligence,leading to a surge of companies entering this field.With the continuous emergence of graphic processing and accelerated computing demands in the domestic information industry,
the domestic GPU market has shown a "flourishing" trend,including some outstanding companies and products.Jingjia Microelectronics is the first company in China to achieve independent research and development of domestic GPU and its industrialization.To break the long-term monopoly of ATI (now acquired by AMD) M9 chip in the military graphics control field,the company has gone through years of technical research and successfully developed the first domestic high-performance GPU chip JM5400 in 2014,achieving the domestication of military GPU.In August 2018,the company successfully developed the second-generation graphics processing chip JM7200,and by 2020,it had gradually completed the adaptation work with domestic equipment and expanded to the general field.In December 2021,the company officially released the third-generation image processing chip JH920.This chip has greatly improved in product performance and process design compared to the previous two generations,and also brings new breakthroughs for domestic GPU in fields such as artificial intelligence and information systems.
Xindong Technology is a company specializing in semiconductor field IP customization,design,and tape-out,involving from DDR to interfaces to GPUs.At the end of 2021,Xindong launched the "Fenghua No.1",which is a 12nm high-performance graphics card,supporting 4-way 4K60 frame output,and up to 16 1080P users online at the same time,filling the gap of domestic 4K-level desktop graphics cards and server-level graphics cards.The "Fenghua" series GPU comes with floating-point and intelligent 3D graphics processing functions,fully customized multi-level pipeline computing core,which can achieve high-performance rendering and intelligent AI computing power,suitable for fields such as the metaverse,cloud gaming,cloud desktop,AI computing,etc.
Moore Threads,founded by former NVIDIA Global Vice President Zhang Jianzhong and established in 2020,is committed to developing world-leading independent innovative GPU intellectual property rights,and its GPU product line covers general graphics computing and high-performance computing.
The company's core members mainly come from globally renowned chip companies such as NVIDIA,Microsoft,Intel,AMD,ARM,etc.,covering the complete structure of GPU research and development design,production and manufacturing,market sales,and service support.One year after its establishment,it released the first-generation MUSA system architecture GPU "Sutie",among which the top-configured desktop-level graphics card MTT S60 has 2048 MUSA cores,a main frequency of 1.5GHz,and provides 6T FLOPS of single-precision floating-point computing power,with 8GB of on-board memory,on-board DP1.4 and HDMI 2.1 interfaces,and supports up to 8K video output.Loongson Technology established a GPU task force in 2020 to accelerate the research and development design of GPU products.Currently,the GPU independently developed by Loongson Technology is integrated in the 7A2000.Loongson 7A2000 is the second-generation Loongson 3 series processor supporting bridge chip for the server and personal computer field,achieving a comprehensive optimization and upgrade on the basis of 7A1000.
In addition,the chip integrates the self-developed GPU for the first time,using a unified rendering architecture,with a 32-bit DDR4 memory interface,and supports up to 16GB of memory capacity.Xintong Semiconductor was established in December 2019,with main businesses including GPU chip design,heterogeneous computing platform solutions,embedded display system solutions,and GPU application deployment solutions.The company's GPU architecture adopts the industry's mainstream unified rendering architecture,and has a highly scalable interconnection structure and computing array.In 2023,Xintong GB2062 was officially released,which is the second generation of independently developed high-performance graphics processor chips launched by Xintong Semiconductor.The chip uses 12nm CMOS process,with independent design,stronger functions,lower power consumption,and can be widely used in the field of three-dimensional graphics (CAD,GIS,animation models,computer games,etc.) and general computing (GPGPU) fields.
Biren Technology was founded in 2019,dedicated to the research and development of original general computing systems,the establishment of efficient software and hardware platforms,and providing integrated solutions in the field of intelligent computing.BR100 is the first self-developed general GPU chip released by Biren Technology,with a 16-bit floating-point computing power of more than 1000T,an 8-bit fixed-point computing power of more than 2000T,and a single-chip peak computing power reaching the PFlops level (1PFlops equals 1000 trillion floating-point instructions per second).Skywise Semiconductor,established in December 2015,officially launched the design and development of GPGPU chips in 2018,and is the first domestic provider of high-end GPGPU chips and supercomputing power.In January 2021,it successfully independently developed a 7-nanometer GPGPU high-end self-developed cloud training chip,featuring comprehensive ecosystem compatibility,high-performance effective computing power,
instruction set programming architecture,and full-stack support for both software and hardware.In addition,there are many companies developing GPU products.However,in recent years,some enterprises have been unable to use TSMC and other process technologies due to U.S.restrictions,which has greatly constrained their development.
04
The combination of domestic innovation and U.S.restrictions creates a huge market space
Due to historical reasons,China has long been in a position of imitation and introduction in the field of information technology.International giants have occupied a large market share and monopolized domestic information infrastructure.They have set domestic IT underlying technology standards and controlled the entire information industry ecosystem.As China's national strength continues to rise,some countries have actively provoked friction in trade and technology fields,attempting to suppress China's peaceful development.The field of information technology,as the underlying support for the national economy,
naturally becomes their main target for attack.Faced with increasing security risks,it is essential to achieve autonomy and controllability,and "domestic innovation" was officially proposed.In 2022,the U.S.authorities banned NVIDIA from exporting high-performance GPU chips (A100,H100) to China.By October 2023,in addition to the originally banned A100 and H100,the castrated versions A800,H8009,L40,L40S,and even the ordinary desktop graphics card RTX4090 were also on the banned list.
In recent years,newly built data centers and intelligent computing centers in China have gradually adopted domestic chip designs,creating a huge market space for domestic enterprises.As a result,many domestic participants have flooded into the relevant fields in recent years,intensifying competition.As some enterprises fail to meet expectations in research and development and market development,and financing funds gradually run out,they begin to be gradually acquired or merged,or even go bankrupt.However,opportunities and risks coexist,
and the huge market space created by domestic innovation and U.S.restrictions for the domestic market will gradually determine the leading enterprises in future industry competition.
Comment