Tsinghua is a chip veteran who specializes in GPGPU! Seize the domestic gap
Aspect: GPGPU processor developed by Denglin Technology adopts software and hardware co-design and will be used by customers for testing at the end of this year.
Lead: The new concept of "AI chip" has gradually gone through the stage of popularization in the past year and is more and more familiar to the public. In the process that the industry has gone through barbaric growth and started to accelerate its landing and integration, more AI chip companies have also begun to step out of their own differentiated routes.
After the first season of previous AI chip series reports, Zhidongxi set out again, and further conducted differentiated in-depth follow-up reports on nearly 100 core enterprises in the whole AI chip industry chain. This is one of the second seasons of the series of reports on the AI chip industry in Zhidongxi.
When it comes to universal AI chips, no one will ignore NVIDIA.
With the super parallel processing capability of GPU, NVIDIA got on the high-speed train of artificial intelligence (AI), and in recent years, its market value soared by more than ten times, making it invincible in the field of AI chips in data centers and becoming an unshakable cloud AI overlord.
The outbreak of computing power provides the cornerstone for the accelerated development and landing of AI technology. The hot AI market has spawned a number of emerging AI chip entrepreneurs at home and abroad.
In the past, chip giants such as NVIDIA and Intel dominated the cloud AI market, and most AI chip startups chose to avoid its edge and specialize in the field of terminal AI-specific chips.
However, one company named "Denglin Technology" has a unique vision, aiming at GPGPU, which is now controlled by NVIDIA. GPGPU is the full name of general purpose computing on graphics processing unit, that is, a graphics processor (GPU) capable of general computing. It extends the application scope of GPU beyond graphics, and has been widely used in large-scale parallel computing in scientific research, education, financial computing, industry and other fields.
At present, only NVIDIA has realized the large-scale commercialization of GPGPU in the world, and the gap in this field in China is still obvious, but several players have been working hard to develop it.
▲ Li Jianwen, founder and CEO of Denglin Technology
On a warm morning, Zhi Dongxi chatted with Li Jianwen, the founder and CEO of Denglin Technology, to see how this veteran who has been deeply involved in GPU for 20 or 30 years led the team to develop innovative GPGPU architecture with versatility and high efficiency, and moved towards the vision of "becoming the NVIDIA of China".
GPU veterans have been in business for a year and a half, and more than a dozen patents are in hand.
As the mainstay of AI development, the opportunities for the underlying hardware of AI are undoubtedly extremely broad, but looking around, most of the solutions for cloud AI training in the world are provided by American company NVIDIA.
The design of high-performance processor is facing many challenges such as high complexity, technical difficulty and deep technical barriers. It is a key field in the field of integrated circuit design, and the relevant talent pool is very limited, even in Silicon Valley, where many IC elites gather.
Li Jianwen studied in Tsinghua University Institute of Microelectronics with a bachelor’s degree, and obtained a master’s degree from the Circuit Teaching and Research Group of Tsinghua Radio Department in 1990.
As early as 2017, Li Jianwen began to think, why don’t we build Chinese’s own server AI chip?
This idea became the bud of Li Jianwen’s founding of Denglin Technology.
Before the establishment of Denglin Technology, Li Jianwen had twenty or thirty years’ experience in GPU field. He served as the vice president of TUxin Technology (founded in 2004), and the GPU/GPGPU IP products he was in charge of were sold to famous semiconductor and technology companies such as Freescale, Intel, Meiman, Google, Samsung, Nokia and Dahua.
At first, Li Jianwen was also attracted by Google TPU, and wanted to have a try on special AI chips. However, during his time as a consultant of Northern Lights Venture Capital, Li Jianwen found that in the face of the rapid evolution of algorithms and the emergence of new applications, the special AI chip similar to Google TPU does not have the versatility needed by the market. Maybe after investing a lot of money and time to make the chip, the market no longer needs this thing.
After seeing the bottleneck of dedicated AI chip, based on years of accumulated industry experience, Li Jianwen is determined to make some achievements in the construction of heterogeneous general computing platform with GPGPU as the core, and make GPGPU solutions with better performance than NVIDIA.
In November 2017, Li Jianwen established Denglin Technology in Shanghai, which was hatched by the well-known high-tech venture capital institution "Northern Lights". The Angel Round and Pre-A Round of financing were completed at the end of 2017 and the first half of 2018 respectively, with a total financing amount of about 150 million RMB.
The core founding team of Denglin Technology consists of eight people, seven of whom have been deeply involved in the GPU field for a long time, and one has rich experience in network processors.
According to Li Jianwen, other members have worked for many years in world-renowned semiconductor, system and Internet companies such as TUX, NVIDIA, AMD, Cisco and Acacia. Each member not only has more than 20 years’ experience in high-tech industry, but also has the achievements of successful streaming and mass production in advanced technology from 28nm to 7nm, and comprehensively covers the parallel processor system architecture, software and hardware, core IP, processor verification platform construction, and the development of the overall SoC chip.
After a year and a half of development, there are about 60 people in the company now, most of whom have at least a master’s degree, no less than 78 years of work experience and very rich industry experience. At present, Denglin Technology has obtained more than a dozen patents, and dozens of core patents are being applied at home and abroad.
Software and hardware combined with innovative GPGPU will be delivered to customers at the end of this year.
With NVIDIA GPGPU Zhuyu in the front, if a startup wants its products to be recognized by the market, it usually chooses to cut prices and earn less money.
But Li Jianwen wants to do more than that. He hopes to do something different. He can not only inherit the advantages of NVIDIA, but also be as good as NVIDIA’s GPU in universality. At the same time, he has made some innovations in technology, which makes the calculation density higher, the efficiency further improved, and the demand for external bandwidth greatly reduced.
In people’s traditional cognition, NVIDIA’s GPGPU core is mainly oriented to graphics acceleration and high-performance computing, but in order to take into account the characteristics of all these applications, his hardware structure is fixed, the execution mode is based on instruction set, and the storage is centralized data storage based on the traditional von Neumann architecture, which limits the efficiency of hardware to some extent and is not the optimal solution for AI applications.
Therefore, Li Jianwen chose to adopt the concept of software and hardware co-design in the system architecture design, so as to solve the problem of both universality and high efficiency of AI computing.
Different from the customized AI processing chips provided by other manufacturers, the "boarding -Minsky" architecture (software-defined heterogeneous AI computing platform) independently innovated by Denglin Technology has higher flexibility in architecture design.
The software can predict the final application of customers. Through feature analysis at a higher application level, various hardware designs including pipeline, control model and storage model can be optimized according to the characteristics of tasks.
Through this architectural innovation, Denglin Technology can make the computing density and efficiency of AI processor hardware higher, the power consumption and area lower, and the requirements for bandwidth are much lower.
"It turns out that it’s like handing over the task to the right hand after the left hand has done something, and the way of software and hardware collaboration is that the two hands work together." Li Jianwen vividly metaphor.
According to him, compared with Tesla V100 and Tesla T4, the latest mainstream products in NVIDIA, the products of Denglin Technology have improved their performance by 5-10 times on the basis of lower-cost mature technology and reducing the chip area by more than 50%.
At the same time, it is unrealistic for a startup company to re-create a set of software ecology directly before it grows up. Therefore, Li Jianwen chose to inherit the ecology with NVIDIA as the main body and provide better solutions.
Li Jianwen revealed that the GPGPU processor at the core of Denglin Technology has passed the FPGA verification, and the design of the first generation product Goldwasser has also been completed, which is being fully verified before streaming. The product is planned to be available for customer testing before the end of this year.
Startups need to see three directions clearly.
Li Jianwen believes that it is very important for a new company to understand three aspects: one is the market, the other is the technology and product route, and the third is that the market and technology direction should be consistent with the team’s ability.
First of all, the AI hardware market is developing rapidly, which is recognized by everyone. According to the report of American research institute Tractica, by 2025, the market of cloud AI chips is expected to reach $14.6 billion.
At this time, the choice of technology and product route is particularly important. Because the iteration speed of the current AI algorithm is still very fast, Li Jianwen thinks that the risk of developing a special AI chip will be greater. Relatively speaking, the route of a general AI chip is more secure and feasible.
Even if you see the market opportunity and find the potential technology and product direction, you need to have the ability to be good at this field. Similarly, taking Denglin as an example, most of their members have mature experience in the GPU field, and they have practitioners with excellent professional and technical skills from software and hardware, IP to SoC, so they have the ability to solve problems quickly and develop products.
Previously, with the increasingly hot AI chips, a large number of fans poured into this market. However, as the capital market became calm, people began to question whether there was a bubble behind this carnival.
For this phenomenon, Li Jianwen thinks that the market is a little overheated. Some people and capital who jumped in underestimated the difficulty of cloud AI processors, but PPT can’t make chips. The ultimate winner must be someone who can have a deeper understanding of customers’ ideas and make products that truly solve customers’ core problems.
Conclusion: The landing tide of large-scale AI chips has not yet arrived.
Last year, it can be said that the domestic and international science and technology circles were involved in a frenzy of AI chips. Both traditional semiconductor companies and AI chip entrepreneurs stood their ground, and various cloud computing companies, AI algorithm companies and traditional industry giants poured in across borders, vying to announce their own research on AI chips.
They chose different technologies and product paths to meet the needs of different application scenarios such as data center, security and voice.
According to the AI chip plans announced by various companies before, there will be a number of products of AI chip companies at home and abroad one after another this year and next. From the underlying hardware to the adaptive solutions, the competition in the field of AI chips will become more intense. The bursting of any bubble will be accompanied by the disappearance of a large number of start-ups, but after the strong wind blows, the start-ups that really grasp the market needs and solve customer problems will stand out.
This account number is Netease News. NeteaseNo. has its own attitude.
Preview of Zhidongxi open class
In June, the open class of face recognition technology, Rainbow Soft, was launched! Scan the code to make an appointment and attend classes for free ~
This article first appeared on WeChat WeChat official account: Wise Things. The content of the article belongs to the author’s personal opinion and does not represent Hexun.com’s position. Investors should operate accordingly, at their own risk.