Where did the domestic GPGPU go-to-market “landing” fall?

From a historical point of view, GPU was created to solve the problem of graphics rendering efficiency in games, but with the continuous iteration of graphics chip technology, GPU processing functions and computing capabilities continue to improve. In 2001, thanks to the appearance of shaders, GPU introduced programmability in the graphics pipeline. From then on, the GPU can no longer be limited to the processing of graphics data, but extends its antennae to other computationally intensive areas, opening up GPGPU era.

　　Skip the GPU

Why do local manufacturers choose self-developed GPGPU?

Some people say that GPGPU without graphics rendering is just a dedicated ASIC. Is this really the case? At the China Integrated Circuit Design Innovation Conference in 2021, the author also asked questions about this issue. Chen Weiliang, CEO of Muxi Integration, replied, “ASIC design deviates from the general GPU architecture, and its software adaptation flexibility will face great challenges, and the product’s application life cycle will be shorter than that of GPGPU.” This means GPGPU. In terms of the underlying architecture, the two technologies can be distinguished from ASIC, which serve the corresponding application scenarios, and cannot be confused.

As for the reason for “skip GPU and directly choose GPGPU as a new battlefield,” Chen Weiliang said, “The global GPU market has formed a semi-monopoly. In addition, graphics rendering is a stock market. It is not easy to confront head-to-head. Choosing incremental market computing This type of GPU, also known as GPGPU, may be a wise move.”

In fact, the statement of skipping the GPU itself is not very accurate. From the perspective of GPU architecture, we can divide GPGPUs into three major families.

　Traditional GPU

The dominant players include Nvidia and AMD/ATI. This type of architecture has a notable feature, that is, it contains an array of a large number of simple processing cores. It processes data in batches in a highly parallelized manner. These processing arrays with vector features are developed from multiple parallel rendering pipelines in traditional GPUs. In addition, there are still more or less components dedicated to image processing in this type of GPGPU, such as texture Cache, frame buffer, and so on. However, with the increasing demand for general-purpose computing, GPGPU is increasingly focusing on general-purpose computing capabilities, and gradually weakening its function as a graphics card.

　　Traditional multi-core CPU

The typical representative is Intel, which is famous for its CPU. Its architectural concept is to tailor the traditional CPU core to obtain a relatively lightweight processing core that constitutes its computing components. The advantage of this is that it can be compatible with the instruction set of some traditional CPUs. Its disadvantage is that compared with the fine-grained processing unit in the above-mentioned GPGPU, this type of processing core is still more complicated, so the core integration is far less than the first family of GPGPU.

　GPU and CPU marriage product

Integrating the architectural features of GPU and CPU, the typical representative is AMD’s APU product series. The approach of this type of architecture is to integrate the processing array in the GPGPU directly into the same chip as the acceleration component of the CPU. This brings two major benefits:

The integration of the CPU core enhances the scalar processing capability of GPGPU, which is more suitable for general computing requirements;

The fusion structure can alleviate the problem of limited communication bandwidth between GPGPU and CPU.

In addition, as a company founded by Chinese Americans, Nvidia did not actually invest too much research and development work in China. Therefore, from the perspective of talent composition, most of my country’s GPGPU entrepreneurs come from AMD, Imagination, and Intel, among which AMD accounts for leading. In fact, when these local companies are doing GPGPU, their R&D personnel have experienced GPU-related technology research and development before. Therefore, from a technical point of view, they did not skip GPU, but from a market perspective, facing the data explosion. With the advent of the times, self-developed GPGPU is the result of conforming to the natural selection of the market.

　　Preliminary formation of the head squad

Domestic GPGPU is brewing a “big landing”

According to incomplete statistics, there are about 7 relatively mainstream GPGPU companies in mainland China. They are Tianshu Zhixin, Denglin Technology, Biren Technology, Moore Thread, Zhuhai Core Power, Muxi Integration and Hongshan Microelectronics.

Why is the head team basically formed? Because the technical development of GPGPU is very difficult. “Usually, the front-end and back-end design of a high-end chip in the industry takes 1 to 3 years. After the design is completed, the tape-out process will take 3 to 6 months. There is a risk that the tapeout will fail and everything will come back. Even if it succeeds The tapeout still needs 3-12 months of product testing and tuning before mass production can finally start.” said Diao Shijing, chairman and CEO of Tianshu Zhixin.

For example, it is better than intel. It started its layout in 1997 through the acquisition of C&T and became a shareholder of Real3D. After the launch of the first discrete GPU i740 in 1998, there were few results. Until 2007, the GPGPU market of NVIDIA began to restart. GPU plans to launch the product Larrabee. Unfortunately, the performance and price are not competitive. In 2020, it will launch a new independent GPU architecture Xe. Unfortunately, as of now, Intel has not launched its own consumer-grade independent GPU products.

It can be seen that the barriers to entry in the GPGPU market are very high. At present, people who have accumulated technology in this area from major manufacturers and can attract investment are basically flowing. In addition, we know that the GPGPU process needs to use 12nm, 7nm or even 5nm technology. In addition to the salary expenses of high-end technicians, the cost of tape-out is also very high. High input means that high output is necessary to make a profit. Therefore, the product is launched. Mass production is imperative. Most of these local GPGPU manufacturers are geared toward the data center and cloud market, and a customer echelon that can cooperate with the “training” of the head has also been formed, and it is difficult to squeeze in later.

Combining the above arguments, let’s take a look at the progress of local GPGPU manufacturers.

From the perspective of product landing, only Tianshu Zhixin has announced that its first cloud 7nm GPGPU product card-“Tianga 100” has officially entered the mass production link (news on October 29 this year), and the single-core computing power is 147T@FP16 in seconds. Followed by Denglin Technology, Zhuhai Core Power and Biren Technology.

Denglin Technology announced on July 14 that its first GPU+ product has successfully passed the test and began to send samples to customers; Li Yuan, the co-founder of Zhuhai Core Power, revealed in July this year that its first chip used in edge servers R8 has been taped out, using Samsung’s 14nm process, with a computing power of 32 TOPS and a power consumption of less than 14W; On October 8 this year, Biren Technology announced that its first 7nm GPGPU—BR100 has been officially delivered to TSMC’s current chip, which is expected to be It will be released to the market next year.

Next is the Moore thread. In its November 25 release, Moore thread only indicated that its first domestically-made full-featured GPU was successfully developed. In other words, tape-out, testing, joint debugging, etc. are still something to be done.

As for the other two, Muxi Integration and Hongshan Microelectronics, the date of establishment is relatively late, and they should still be in the initial product development stage. It is worth mentioning that Muxi’s integrated R&D team is still very strong. Its founder, Weiliang Chen, was the senior director of AMD graphics R&D, and CTO Yang Jian is a former AMD Fellow, so there is a great chance of being among the top three in the future.

Of course, the previous is just sorted out according to the product landing situation of each company. As for the product structure, process, computing power, energy efficiency ratio, market segment and company strategy they use, they are not within the scope of consideration. Taking it into consideration, this progress may have to be overturned and restarted. However, judging from the product release progress of these manufacturers, everyone currently wants to be the “first”, and it can be said that a “go-to-market” grand landing is brewing.

　　Where is the downstream market for GPGPU?

The investment of manufacturers is inseparable from the promotion of the investment community, let alone the demand of the downstream market.

According to public data, 90% of China’s GPGPU market is currently being carved up by foreign companies represented by Nvidia. In 2019 alone, Nvidia and AMD two major U.S. companies earned about 8 billion yuan in the domestic GPGPU market. The semi-monopoly market brings high prices. The current market price of a high-end GPGPU board is as high as 100,000 yuan, which is equivalent to the price of an ordinary car. At the same time, some experts predict that by 2025, the market size of GPGPU chip cards in my country will reach 45.8 billion yuan, with a compound annual growth rate of up to 32%. It can be seen that the localization of GPGPU products is imperative.

So someone has to ask, in which industries are the domestic market demand for GPGPU distributed? According to the industry structure of China’s GPGPU shipments provided by industry insiders, nearly half of GPGPUs will be used in the Internet market in 2019, about one-third of GPGPUs will be used in the security and government markets, and about one-tenth of GPGPUs will be used for others. For AI applications in the industry, nearly one-tenth of GPGPUs are used in the HPC market. Of course, this is the data for 2019, and the current situation may be adjusted, for example, the proportion of HPC has increased.

　Write at the end

Nvidia has a market value of US$300 billion and AMD has a market value of US$100 billion. China’s GPGPU market and track are large enough. Maybe in three to five years, the local country will be able to make a “little NVIDIA” or “little AMD”. So, the next step for domestic GPGPU is how to make it bigger? In addition to continuous financing to provide economic power, all the competitiveness of these companies will focus on one point, that is, the cost-effectiveness of the product under high performance, the cost-effectiveness determines the shipment volume, the shipment volume represents the recognition of the market, and the market brings capital. Thus forming a virtuous circle.