Another chip giant is on the verge of bankruptcy, and the competition in the GPU track is intensifying

The semiconductor industry is vertical

2024-05-15 18:48Posted in Guangdong science and technology creators

Another chip giant is on the verge of bankruptcy, and the competition in the GPU track is intensifying

In recent years, many chip companies have gone bankrupt. Last year, as many as 10,000 chip companies closed their doors, with an average of 31 per day. And this year is still miserable, China's IP design star Huaxia Core bankruptcy liquidation, Lisuan Technology also burned 300 million yuan, arrears of wages broke out in June last year, coupled with the withdrawal of the US dollar fund, almost bankrupt on the eve of tape-out.

The company was established in 2021, and its actual operation began in early 2022 with the establishment of Shanghai Calculation. Shanghai is a wholly-owned subsidiary of Nanjing Lisuan Technology Co., Ltd., and it is also the core of the company. The three founders are Xuan Yifang, Kong Dehai and Niu Yixin, all of whom have deep technical capabilities and engineering experience in S3, a Silicon Valley GPU company.

According to the information on the official website, Lisuan Technology adheres to the self-developed architecture and independent intellectual property rights, and is building domestic GPU chips that benchmark international mainstream products, serving the domestic 200 billion rendering GPU all-round market, realizing end, cloud, and edge graphics rendering, and meeting the needs of chips and computing power for learning, work and entertainment, metaverse, games and animation, digital twins, film and television production, ARVR, smart factories, professional design, smart cockpits, smart medical care, etc.

Fortunately, Dongxin Co., Ltd. recently issued an announcement, announcing that it intends to use its own funds or over-raised funds to increase the capital of Fatsu Technology (Shanghai) Co., Ltd., accounting for no more than 40% of the shares, with an upper limit of 200 million yuan. Behind the bankruptcy of Lisuan Technology is the intensification of competition in the GPU chip track.

The origin of the GPU

The full name of GPU is Graphics Processing Unit. Its initial functionality was mainly used to draw images and process specific chips for metadata, but many other functions were later added.

GPUs are better suited for intensive data processing than CPUs. From the perspective of chip microarchitecture, GPU is optimized and adjusted on the basis of the original CPU structure, through merging with the CPU, RAM sharing is realized, and SIMD architecture is used to build hundreds or thousands of logic operation units to achieve parallel processing at the same time.

GPUs are particularly good at handling many tasks at the same time, such as converting data from a computer into a beautiful image that we see on screen. In computing tasks, GPUs use their parallel computing advantages to accelerate the processing of large-scale data-intensive tasks such as deep learning training and scientific computing.

In 1962, the thesis of Ivan Suzel, Ph.D. of the Massachusetts Institute of Technology, and his sketchpad program, laid the foundations of computer graphics. From 1962 to 1984, there was no special graphics processing hardware, and graphics processing tasks were completed by CPUs.

With the development of computers, the demand for image processing has gradually increased. Tracing the history of GPUs starts with graphics display controllers. The world's first personal computer IBM5150 was released by IBM in 1981, and this PC was equipped with a monochrome display adapter (MDA) and a color graphics adapter (CGA), which was the first graphics display controller. In 1984, SGI-USA launched a high-end graphics workstation for the professional field. From 1984 to 1995, SGI continued to develop a series of better-performing graphics workstations.

In 1995, 3DFX released Voodoo, a consumer 3D graphics card. At this time, the graphics display hardware track has begun to become hot, and AMD, ATI (acquired by AMD in 2006), and NVIDIA have all begun to launch their own graphics card products. The CPU is able to get rid of some of the graphics processing tasks, but tasks such as vertex transformations still need to be done by the CPU.

Founded in 1993, NVIDIA is the originator of the GPU concept. In 1999, NVIDIA released the GeForce 256 graphics chip, which introduced the concept of GPU for the first time, and the GPU era came. The GeForce 256 features advanced technologies like "T&L" hardware, cubic environment material mapping, and vertex blending.

During this period, each hardware unit formed a graphics processing pipeline, and each pipeline level had a fixed function, and the hardware could not be programmed. As technology evolved, GPUs went from working in a fixed way to being smart chips that could be programmed and capable of creating more complex visual effects. In 2003, NVIDIA and ATI released new products with both programmable vertex processing and programmable pixel processors, which have good programmability.

In the 21st century, with the popularity of gaming and multimedia applications, the performance requirements of GPUs are growing rapidly. Major chip manufacturers have invested heavily in the research and development of a new generation of GPUs to continuously improve their performance. At this stage, the internal composition of the GPU has also undergone significant changes, including the graphics memory controller, compression unit, BIOS, graphics and computing column, bus interface, power management unit, video management unit, etc.

With the rise of artificial intelligence and deep learning technology, the application field of GPU has been further expanded. NVIDIA has developed CUDA technology, which allows GPUs to not only draw pictures, but also do more non-graphics work, such as processing big data. In addition to traditional graphics rendering, GPUs are also widely used in deep learning, machine vision, big data analysis, and other fields. This trend has driven the further development of GPU technology, which has gradually evolved from a dedicated graphics processor to a general-purpose computing platform.

A number of R&D GPU companies began to emerge in China: Jingjia Microelectronics, Xinyuan Microelectronics, Biqian Technology, etc.

It was not until April 2014 that Jingjiawei successfully developed the first domestic high-performance, low-power GPU chip - JM5400. In August 2018, the new generation of high-performance and high-reliability GPU chip - JM7200 independently developed by Jingjiawei was successfully taped out. In 2019, on the basis of JM7200, Jingjiawei launched a commercial version - JM7201 to meet the high-performance display needs of desktop systems and fully support domestic CPUs and domestic operating systems.

In 2020, the chip industry set off an entrepreneurial boom, this time mainly focusing on the GPU field. With the blessing of capital, domestic GPUs have developed rapidly, and a large number of companies have entered this field. In addition, the central and local governments continue to introduce policies to support the development of the integrated circuit industry, and under such a comprehensive influence, it can be said that GPU has ushered in the best era. These startups are mainly focused on developing GPUs that can meet the needs of both graphics processing and AI computing, which is now known as GPGPU (General-purpose GPU) in China. But as the financing funds ran out, some companies also fell into the brink of bankruptcy.

According to IDC data, the global AI chip market will reach $35.2 billion in 2022. GPUs account for the largest share of the total, with Goldman predicting that GPUs will account for 57% by 2025. With the United States restricting China's imports of the latest Nvidia GPU products, domestic GPUs have a bright future.

Global GPUs are in a three-legged oligopolistic competition pattern

From the perspective of the overall global GPU market, the global GPU market has formed a three-legged oligopoly competition pattern, and NVIDIA is still the absolute leader in the global independent GPU. Intel is now the dominant player in the integrated GPU market, thanks to its strength in the notebook and traditional PC industries, while the discrete GPU market is dominated by NVIDIA and AMD.

NVIDIA is now the largest manufacturer and distributor of independent graphics cards, including the Geforce series for civilian use, as well as the professional Quadro series. Its GPU has cuda universal computing, PureVideo HD video technology, PhysX physical acceleration, Optimus smart graphics switching, etc. In March this year, NVIDIA launched a new generation of AI graphics processing unit chip (GPU) architecture Blackwell, and released the GPU using this architecture - B200 and GB200 product series. It is understood that Blackwell can support models with up to 10 trillion parameters for AI training and real-time LLM inference. The B200 is a combination of two ultra-large die (bare die) packages, containing more than 208 billion transistors, more than double the 80 billion transistors of the previous generation, and 192GB of high-speed HBM3e video memory on the entire chip.

Intel is the world's largest supplier of integrated GPUs. Intel's GPUs are mainly integrated graphics cards, which are used in Intel's motherboards and Intel's notebooks. Intel's first generation of discrete graphics cards was released in 2022, about 2 years ago. Recently, some media reported that Intel has added bmg_g21 (Battlemage G21, BMG-G21) core related code to the LLVM documentation of the oneAPI DPC++ compiler. This is also the first time that code related to the Battlemage architecture has appeared in the document, suggesting that the Battlemage G21 GPU is expected to be the first model of Intel's next generation of GPU.

Founded in 1969, AMD is one of the few semiconductor companies in the industry that can provide high-performance CPUs, high-performance discrete graphics GPUs, and motherboard chipsets. In 2006, AMD acquired ATI for $5.4 billion, becoming the first semiconductor company to design both high-performance CPUs and GPUs. AMD has the Radeon series for civilian use, as well as the professional FireGL series. Its GPU has Stream general computing, ATI Video Converter video transcoding, UVD high-definition video technology, Havok, Bullet and Pixelux DMM three physics engines, etc. AMD's GPUs are divided into two distinct product segments, one for gaming and the other for supercomputers, big data analytics, and machine learning systems.

In 2023, AMD officially released a new generation of AI/HPC accelerators Instinct MI300X and Instinct MI300A, one is a traditional GPU accelerator, and the other is a unique CPU+GPU fusion accelerator. MI300A is oriented to HPC product positioning, so the product form is CPU+GPU/APU combined chiplet (24 Zen4 cores and their I/0, 128G HBM3, encapsulated 13 chiplets); And the underlying IF high-speed interconnection and UMA memory architecture have been transformed. The first purchaser was also the EI Capitan Exascale. In line with NVIDIA's CUDA, AMD has developed ROCm.

The domestic GPU market is full of flowers

Although domestic GPUs started late, thanks to the blessing of capital and the market catalysis of artificial intelligence in recent years, a large number of companies have entered this track. With the continuous emergence of the demand for graphics processing and accelerated computing in the domestic information industry, the domestic GPU market has shown a trend of "a hundred flowers blooming", including some outstanding enterprises and products.

Jingjiawei is the first enterprise in China to realize independent research and development of localized GPUs and industrialization. In order to break the long-term monopoly of ATI (now acquired by AMD) M9 chip in the field of military graphics display and control, the company successfully developed the first high-performance GPU chip JM5400 in China in 2014 after years of technical research, realizing the localization of military GPU. In August 2018, the company successfully developed the second-generation graphics processing chip JM7200, and in 2020, it successively completed the adaptation with localized equipment, and gradually expanded to the general field. In December 2021, the company officially released the third-generation image processing chip JH920. While the chip has greatly improved product performance and process design compared with the previous two generations, it has also brought new breakthroughs to localized GPUs in artificial intelligence, information systems and other fields.

Innosilicon is a company focusing on IP customization, design, and tape-out in the semiconductor field, from DDR to interface to GPU. At the end of 2021, Innosilicon launched "Fenghua No. 1" is a 12nm high-performance graphics card, supporting 4 channels of 4K60 frame output, and up to 16 1080P users online at the same time, filling the two gaps of domestic 4K-level desktop graphics cards and server-level graphics cards. The "Fenghua" series of GPUs comes with floating point and intelligent 3D graphics processing functions, and fully customized multi-level pipeline computing cores, which can achieve high-performance rendering and intelligent AI computing power, and are suitable for metaverse, cloud gaming, cloud desktop, AI computing and other fields.

Founded in 2020 by Zhang Jianzhong, former global vice president of NVIDIA, Moore Threads is committed to the research and development of the world's leading independent innovation GPU intellectual property, and its GPU product line covers general-purpose graphics computing and high-performance computing. The core members of the company are mainly from NVIDIA, Microsoft, Intel, AMD, ARM and other world-renowned chip companies, covering GPU R&D and design, manufacturing, marketing, service support and other complete architectures. One year after its establishment, it released the first generation of MUSA system architecture GPU "Su Di", which includes the top desktop graphics card MTT S60, with 2048 MUSA cores, the main frequency of 1.5GHz, providing 6T FLOPS single-precision floating point computing capabilities, onboard 8GB video memory, onboard DP1.4 and HDMI 2.1 interfaces and support up to 8K video output.

Loongson Zhongke established a GPU commando team in 2020 to accelerate the R&D and design of GPU products. At present, the GPU independently developed by Loongson Zhongke is integrated in 7A2000. Loongson 7A2000 is the second generation of Loongson No. 3 series processor supporting bridge for servers and personal computers, which is fully optimized and upgraded on the basis of 7A1000. In addition, for the first time, the self-developed GPU is integrated on the chip, using a unified rendering architecture, with a 32-bit DDR4 video memory interface, and a maximum of 16GB video memory capacity.

Founded in December 2019, the company's main business includes 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, the GB2062 will be officially released, which is the second generation of self-developed high-performance graphics processor chips launched by Xintong Semiconductor. The chip adopts 12nm CMOS process, with independent design, stronger functions, lower power consumption and other characteristics, which can be widely used in the field of three-dimensional graphics (CAD, GIS, animation models, computer games, etc.) and general computing (GPGPU).

Founded in 2019, Bichen Technology is committed to developing an original general computing system, establishing an efficient software and hardware platform, and providing integrated solutions in the field of intelligent computing. BR100 is the first general-purpose GPU chip independently developed by Bichen 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 of PFlops (1PFlops is equal to 1000 trillion floating-point instructions per second).

Founded in December 2015, Tiantian Zhixin officially launched the design and development of GPGPU chips in 2018, and is the first provider of GPGPU high-end chips and super computing power in China. In January 2021, it successfully independently developed a 7nm GPGPU high-end self-developed cloud training chip, which has the advantages of all-round ecological compatibility, high-performance effective computing power, instruction set programming architecture, and full-stack support of software and hardware.

In addition, there are many companies working on GPU products. However, in recent years, due to the restrictions of the US ban, some companies cannot use TSMC and other process tape-outs, which greatly restricts the development of enterprises.

The superposition of the US ban has brought huge market space

Due to historical reasons, the mainland has long been in the position of imitation and introduction in the field of information technology. International giants occupy a large market share and also monopolize the domestic information infrastructure. They have formulated domestic IT underlying technology standards and controlled the entire information industry ecology. With the continuous rise of China's national power, some countries have taken the initiative to provoke frictions in the fields of trade and science and technology in an attempt to suppress China's peaceful development. The field of information technology, which is supported by the bottom of the national economy, has naturally become their key target. In the face of increasing security risks, it is necessary to achieve autonomy and control, and "information innovation" has been officially proposed.

In 2022, the U.S. authorities banned Nvidia from exporting high-performance GPU chips (A100, H100) to China, and by October 2023, in addition to the original banned A100 and H100, the castrated version of the A800, H8009, L40, L40S, and even ordinary desktop graphics cards RTX4090 are all banned.

In recent years, the newly built data centers and intelligent computing centers in China have gradually adopted domestic chip design, creating a huge market space for domestic enterprises. As a result, many domestic players have poured into the relevant tracks in recent years, and competition has intensified. As the R&D and market development of some enterprises are less than expected, the financing funds are gradually burned out, and they begin to be gradually acquired and acquired, and even go bankrupt. However, the crisis and the opportunity are at the same time, and the huge market space created for the domestic market driven by the information innovation and the US ban will gradually determine the head enterprises in the future industry competition.

View original image 939K