In recent years, the semiconductor industry has seen significant upheaval, with thousands of chip companies declaring bankruptcyJust last year, the overwhelming toll amounted to over ten thousand companies shutting their doors, averaging around 31 closures dailyThis year continues to paint a bleak picture with the recent bankruptcy liquidation of Huaxia Chip, a rising star in Chinese IP design, alongside Lixuan Technology, which burned through 300 million RMB before facing a near bankruptcy situation last June due to wage arrears and the withdrawal of dollar-based funding before crucial production stages

Founded in 2021, Lixuan Technology initiated operations with the establishment of its Shanghai branch in early 2022. This branch, a wholly-owned subsidiary in Nanjing, represented the core of the company

The founding trio: Xuan Yifang, Kong Dehai, and Niu Yixin, all brought extensive technical expertise and engineering experience from their tenure at the Silicon Valley GPU company, S3.

According to the company's official information, Lixuan Technology is committed to self-developed architectures and proprietary intellectual property rightsIt is actively working to create domestic GPU chips that meet international standards, aimed at servicing China's extensive 200 billion RMB rendering GPU marketThe focus is on diverse applications, ranging from graphic rendering in learning and entertainment to training in metaverse, gaming, digital twin, film production, augmented reality/virtual reality, smart factories, professional design, intelligent cockpits, and smart healthcare.

Fortunately, a recent announcement from Dongxin Co

• What is a 'Slow Bull' Market?
• Is the Dollar's Reign as Global Currency at Risk?
• Financial Platform Defaults: Key Warning Signs for Investors
• German Economic Slump Fuels Europe Fears
• Semiconductor M&A Surge Drives Industry Consolidation

indicated plans to inject capital into Lixuan Technology (Shanghai), with their stake not exceeding 40% and an upper limit of 200 million RMBThis news underscores the intense competition brewing in the GPU chip arena.

History of GPU Development

The term GPU stands for Graphics Processing UnitInitially, these specific chips were designed to draw images and handle graphics data, but they have since expanded greatly to encompass a wide range of functionalities.

Compared to CPUs, GPUs are more adept at intensive data processing

From a micro-architecture perspective, GPUS are optimized and adjusted versions of traditional CPU structures, enabling shared RAM and employing a SIMD (Single Instruction, Multiple Data) architecture for hundreds or thousands of logical operation units, thereby achieving parallel processing.

GPUs excel at handling numerous tasks simultaneously, such as converting computer data into stunning visuals on our screensThey leverage their parallel computing strength to accelerate processing large-scale, data-intensive assignments such as deep learning training and scientific calculations.

The groundwork for computer graphics was laid in 1962 by MIT’s doctoral student Ivan Sutherland and his pioneering sketchpad program

In the period from 1962 to 1984, no dedicated graphic processing hardware existed, and all graphic tasks were executed by CPUs.

With the evolution of computers, the need for image processing has risen steadilyThe first graphics display controllers emerged with the introduction of the IBM5150, the world’s first personal computer, in 1981, equipped with monochrome and color graphics adaptersThe American company SGI launched high-end graphics workstations targeting professional fields in 1984, continuing to innovate through the mid-90s with a series of superior performance workstations.

The evolution continued in 1995 with 3DFX's release of the consumer 3D graphics card, Voodoo

This marked the heating up of the graphics hardware market, with companies like AMD, ATI, and NVIDIA following suit with their own offeringsWhile CPUs were relieved of some graphic processing duties, tasks like vertex transformation remained CPU dependent.

Founded in 1993, NVIDIA is credited with introducing the GPU conceptIn 1999, Nvidia launched the GeForce 256, which was the first to embody the GPU concept, ushering in a new eraThis chip introduced advanced technologies such as T&L hardware, cubic environment texture mapping, and vertex blending.

During this period, graphics processing units operated along a fixed-function pipeline, without programmability for hardware

With advancements, GPUs transformed from fixed-function units to programmable chips capable of generating sophisticated visual effectsThe 2003 releases from NVIDIA and ATI both included programmable vertex and pixel processors, demonstrating significant programmability.

Entering the 21st century, the rapid growth of gaming and multimedia applications dramatically increased the performance demand for GPUsMany chip manufacturers invested heavily in developing the next generation of GPUs, leading to notable internal changesGPU components now encompass memory controllers, compression units, BIOS, both graphical and computational parallel processing, bus interfaces, power management units, and video management units.

With the rise of artificial intelligence and deep learning technologies, the applications of GPUs have grown extensively

NVIDIA's development of CUDA technology expanded GPU capabilities beyond mere graphics processing to encompass large data handling needsAside from traditional graphics rendering tasks, GPUs are now widely utilized in deep learning, machine vision, and big data analysis, evolving from dedicated graphics processors to general-purpose computing platforms.

China has witnessed the emergence of several companies focused on GPU development, such as Jingjia Microelectronics, Chipron Microelectronics, and Birun Technology.

Until April 2014, Jingjia successfully developed the first domestically high-performance, low-power GPU chip, JM5400. By August 2018, the company had crafted the next-generation high-performance, highly reliable GPU chip, JM7200. In 2019, the commercial version of JM7201, built on the JM7200 foundation, emerged to meet the performance display needs of desktop systems while fully supporting domestic CPUs and operating systems.

The onset of 2020 heralded a new wave of startups within the semiconductor industry, primarily concentrated in the GPU sector

The inflow of capital accelerated the development of domestic GPUs, propelling numerous companies into this spaceAdditionally, continual central and local policies have bolstered integrated circuit industry growth, making it a promising epoch for GPUsThese nascent firms are primarily striving to develop GPUs that effectively balance graphics processing and AI computations, a concept termed GPGPU (General-purpose GPU) within ChinaHowever, as funding dwindled, several companies found themselves on the brink of collapse.

According to IDC data, the global AI chip market is projected to reach $35.2 billion in 2022, with GPUs holding the largest shareGoldman Sachs estimates that by 2025, GPU market share will reach 57%. Given the U.Srestrictions on the import of the latest NVIDIA GPU products into China, the future for domestic GPUs appears promising.

Global GPU Market Landscape

A look at the global GPU market reveals a triopoly featuring NVIDIA as the dominant independent GPU leader

Benefitting from its stronghold in the laptop and traditional PC sectors, Intel has carved out an area for integrated GPUs, while independent GPUs are predominantly shared between NVIDIA and AMD.

NVIDIA stands as the largest manufacturer and seller of independent graphics cards, with its consumer-friendly Geforce series and professional Quadro seriesIts GPU lineup incorporates features like CUDA for general computation, PureVideo for high-definition video playback, PhysX for physical simulations, and Optimus for intelligent GPU switchingIn March of this year, NVIDIA unveiled its latest AI graphics processing unit architecture, Blackwell, releasing the Blackwell-based products—B200 and GB200. Reports suggest that Blackwell can support model training for AI with up to 100 trillion parameters

The B200 combines two ultra-large dies with over 208 billion transistors, more than double the previous generation's 80 billion, and houses 192GB of high-speed HBM3e memory.

Intel is the world’s largest integrated GPU supplier, primarily producing integrated graphics cards for Intel motherboards and laptopsIntel's first iteration of a dedicated graphics card launched in 2022, approximately two years agoRecently, media reports indicated that Intel added core-related code for its Battlemage G21 GPU to the LLVM documentation, hinting at its potential first launch.

AMD, established in 1969, has become a rarity in the semiconductor industry by offering high-performance CPUs, independent GPUs, and motherboard chipsets

In 2006, AMD's acquisition of ATI for $5.4 billion positioned it as the first company to design high-performance CPUs alongside GPUsAMD produces both consumer-grade Radeon and professional FireGL series GPUs, offering features like Stream computing, ATI Video Converter, and multiple physics engine supportAMD's GPUs are categorized into gaming and supercomputing pipelines, catering to both sectors.

In 2023, AMD launched its new generation of AI/HPC accelerators, Instinct MI300X and MI300A, the former serving as a traditional GPU accelerator and the latter a unique fused CPU+GPU acceleratorMI300A is explicitly designed for HPC, integrating two Chiplets comprising 24 Zen4 cores, 128GB HBM3 memory, and optimizing interconnections and memory architectureThe first buyers include national supercomputing centers.

Diverse Developments within China's GPU Market

Although the GPU development in China started late, recent financial investments and the catalyst of AI market demand have driven numerous companies into this space

The emerging demand for information technology and accelerated computing within China has created a flourishing GPU market landscape, showcasing many proficient enterprises and products.

Jingjia Microelectronics became the first company to achieve the independent development and industrialization of domestic GPUsTo counteract ATI's M9 chip monopoly in military graphics control, the company invested years into technical research and in 2014 successfully produced China's first high-performance GPU chip, the JM5400. The subsequent development of the second-generation chip, JM7200, followed in 2018, and by 2021, the third iteration, JH920, promised substantial enhancements in product performance and design.

Chip Motion Technology focuses on customized IP design and has released the “Fenghua 1” card at the end of 2021, a high-performance graphics card that supports four lanes of 4K output and simultaneous online usage by up to 16 users at 1080p resolution

The “Fenghua” GPUs carry floating-point and intelligent 3D graphics processing capabilities that allow for high-performance rendering suited for applications like the metaverse and cloud gaming.

Founded in 2020 by former NVIDIA Vice President Zhang Jianzong, Moore Threads strives to develop industry-leading, self-innovated GPU intellectual property rightsTheir GPU line covers a wide spectrum, targeting general graphics computation and high-performance computingThe core team comprises members from renowned companies like NVIDIA and Intel, encompassing all aspects of GPU development, manufacturing, marketing, and support

One year post-establishment, they unveiled their first generation GPU with the MUSA system architecture, the SudiThe desktop-grade MTTS60 boasts 2048 MUSA cores, 1.5GHz frequency, and 6T FLOPS capability while being equipped for advanced video output.

Loongson Technology formed a GPU task force in 2020 to expedite the development of GPU productsTheir in-house GPU is integrated into the 7A2000 series, designed for servers and personal computingBuilt on the 7A1000 foundation, it delivers extensive optimizations and the first integration of self-developed GPU technology.

Xintong Semiconductor, launched in December 2019, focuses on GPU chip design and offers solutions for heterogeneous computing platforms and embedded display systems

Their GPU architecture features a widely-accepted unified rendering structure with a scalable interconnection and computing arrayIn 2023, they launched the GB2062, their second-generation high-performance graphics processor chip, fabricated using 12nm CMOS technology.

Birun Technology, established in 2019, aims to develop original general computing architectures and establish efficient hardware-software platformsThe BR100, their independently developed general GPU chip, has a floating-point computing power exceeding 1000T and integer computing power exceeding 2000T, achieving peak computing performance on a single die.

Tianshu Zhixin commenced GPGPU chip R&D in December 2015, becoming China’s first high-end GPGPU chip and super computing power supplier

In January 2021, they successfully developed a cloud-based training chip on a 7nm GPGPU basis with comprehensive ecosystem compatibility and exceptional calculation performance.

Additionally, numerous firms are delving into GPU product developmentHowever, recent U.Srestrictions have hindered several companies from utilizing TSMC processes for production, significantly constraining their growth.

Confluence of Domestic Innovation and U.SRestrictions Creates Market Opportunities

Due to historical factors, China's information technology sector has long operated in a mode of imitation and importation, with international giants holding substantial market shares and monopolizing the nation’s infrastructure

They have set the underlying tech standards and controlled the information industry’s ecosystemWith the escalation of China’s national power, certain nations have instigated trade and tech tensions, attempting to suppress China's peaceful developmentThis crucial sector has naturally become a focal point for their attacks, necessitating the realization of independent capabilities, thus giving birth to the notion of “Xinchang” or self-innovation.

In 2022, the U.Sgovernment prohibited NVIDIA from exporting high-performance GPU chips (A100, H100) to ChinaBy October 2023, not only were the originally banned A100 and H100, but also stripped versions like A800, H8009, L40, L40S, and even ordinary desktop cards RTX4090 faced restrictions.

Recently, China’s newly established data centers and intelligent computing hubs have gradually embraced domestic semiconductor designs, creating substantial market opportunities for Chinese enterprises