AI Chips Compete With Each Other: Financing Reaches New Heights And Faces Challenges Of Commercial Landing

As we all know, in the past two years, the commercial landing of AI chips is getting colder, but in 2021, the investment circle's enthusiasm for AI chips will rise sharply. At the beginning of the year, the financing information of start-ups appeared frequently, and a new wave of companies, such as Suiyuan, Hanbo, Muxi, Biren, Moore thread and tiantianzhixin, emerged one by one.

With the continuous emergence of start-up companies, domestic technology giants have also increased the investment in AI chips. Recently, Tencent released the information of chip R & D positions on the official recruitment website, including chip architects and chip design engineers. In response, Tencent said that it was based on some business needs. Tencent had some chip R & D attempts in specific fields, such as AI acceleration and video codec, but they were not general-purpose chips.

So far, bat has entered the market. In the first half of this year, baidu announced that its Kunlun chip business has established an independent new company, aiming at the cloud AI general chip. The second generation Kunlun chip has been successfully taped and will be mass produced in the second half of 2021; Byte skipping also means that we are setting up a relevant team to do some exploration in the field of AI chips; Huawei's rising ecology continues to unfold, and Alibaba's optical chip has also been put on the market.

From the global point of view, it is another ecology. NVIDIA still dominates the AI chip market, with GPU in the terminal and cloud. Its challengers are also fierce. Intel, AMD, Google, Amazon, and even Tesla, which has self-developed chips, have not only invested heavily in some key technological breakthroughs, but also are fully armed and integrated from end to end. We can see that after years of cooperation, the battlefield of AI chip has come to the stage of chasing the Central Plains.

Domestic AI investment boom reappears

At present, domestic AI chip enterprises are in the stage of technological innovation and commercial exploration. In the previous wave, Cambrian has been listed, horizon is said to be on the market this year, and Shenjian technology was acquired by Xilinx in 2018; Among the emerging AI start-ups, many enterprises directly enter the GPU track to compete positively with NVIDIA, and some compete in the subdivision field through innovation dislocation such as architecture level. Among these new forces, many of the founding members come from NVIDIA, AMD and other big factories. In 2021, when AI and chips are intertwined, the team has been favored by the capital circle since its inception. The amount of financing and valuation are quite high. Some enterprises have reached the scale of "unicorn".

Specifically, the large amount of investment related to GPU was mainly concentrated in the first quarter of this year, and these enterprises targeted NVIDIA's powerful cloud training field. For example, Moore thread completed two rounds of super billion financing within 100 days of its establishment; Founded more than a year ago, Bilin technology has completed the B round of financing, with an accumulated financing amount of more than 4.7 billion yuan, and the founding team has gathered experts; It is reported that the first GPU chip of Bi Ren technology locates high-end general intelligent computing and supports cloud based training and reasoning. At present, it has reached the end stage and is expected to be released this year; Muxi has also completed several hundred million yuan pre a + round of financing, and plans to develop high-performance GPU chips fully compatible with CUDA and ROCM ecology to meet the computing needs of HPC, data center and AI; Days smart chip completed the C round of 1.2 billion yuan financing, and its first TSMC OEM 7-nm cloud training chip Bi has entered the stage of mass production, and will enter the large-scale commercial sector.

At the same time, many enterprises choose AI cloud reasoning field to break through. From the perspective of market demand, the market of AI reasoning is expanding. According to the report of CCID consultant, a research institute, from 2019 to 2021, China's AI chip market size will still maintain a growth rate of more than 50%, and by 2021, the market size will reach 30.57 billion yuan. Among them, with the completion of the construction of large-scale local data centers, the growth rate of cloud training chips slows down; With the release of market demand in various fields, the growth rate of cloud inference chip and terminal inference chip market will continue to show an upward trend.

For example, Qian Jun, founder and CEO of Hanbo semiconductor, told the 21st century economic reporter that the reason side breakthrough was selected because of market demand. Some data show that the reasoning market in 2021 has been larger than the training market; On the other hand, because GPU is not the best architecture on the reasoning side, Hanbo semiconductor has a better DSA architecture. At the same time, Hanbo semiconductor also strengthens the video processing ability; Founded in 2018, Suiyuan technology has both training chips and reasoning chips. Suiyuan technology has obtained the latest round C financing, with a total financing of 1.8 billion yuan. The latest launch of the second generation chip will also accelerate the pace of commercialization. Zhao Lidong, CEO of Suiyuan technology, said that Suiyuan's products are mainly aimed at Pan Internet, traditional industries (finance, transportation, power, medical, medical, etc.) Industry, etc.) and new infrastructure.

To compete with international giants

Although the AI chip market is huge, it will eventually concentrate on the leading enterprises from the perspective of development law. Some investors pointed out to the reporter of the 21st century economic report that there are many AI enterprises in the domestic market in recent two years, reaching dozens, and will face integration and elimination in the future.

For domestic enterprises, the similar background also brings about the homogenization of start-ups. At present, the commercialization situation is even more tested. Some chip companies have formed cooperative relations with Internet enterprises and scientific research institutions. How to continue to expand the scale and continuously export high-performance products is a challenge to NVIDIA.

Some chip practitioners told reporters that no matter what the team background is, the most important thing is how to realize the commercial landing. In the internal environment, hardware architecture and software are needed to run in. When contacting customers after the combination of hardware and software, greater challenges have begun. In the face of the actual requirements of customers for products, chip companies need to constantly optimize. At the same time, in the process of cooperation with customers, it also means direct PK with giant NVIDIA in performance and other aspects, and needs to prove its high performance. The competition is still very fierce.

After all, the artificial intelligence ecology built by NVIDIA is still powerful. As early as 2007, NVIDIA put forward the concept of GPGPU (GPU for general purpose computing), which widely applied GPU from image processor in traditional impression to computing training, and seized the development opportunities in the fields of graphics rendering, deep learning and blockchain. At the same time, NVIDIA has also built CUDA platform, which is a complete set of software and hardware ecological standards for parallel computing established by NVIDIA. Most AI chips are compatible with it, especially the training end chip. If the acquisition of arm is successful, NVIDIA will become more powerful.

NVIDIA's ecology is very strong, but other technology companies are also rubbing their hands. Google began to manufacture its own chips and TPU chips in 2015; Amd acquired Xilinx for AI data center products; After acquiring Annapurna labs in 2016, Amazon began transferring Alexa's brain to its inferentia chip last year.

Intel made up for its AI chip layout through acquisition. In 2015, it acquired Altera, a FPGA giant, with us $16.7 billion. FPGA has great potential in cloud computing, Internet of things and edge computing. In 2016, it acquired nervana with $408 million, and Habana labs, an AI chip enterprise, with $2 billion in 2019. In addition, Intel has acquired movidius, a visual processing chip start-up, and Mobileye, a self driving company. At the same time, Intel released the idm2.0 plan, but also to strengthen the chip manufacturing end. It is not difficult to see that various enterprises are carrying out the layout of the whole business chain.

It is worth noting that in addition to these enterprises, Tesla's progress has also attracted much attention. After its decision to use self-developed chips, the idea of software and hardware integration is also challenging the traditional chip design manufacturers. AI practitioners told reporters that AI enterprises have been looking for new application scenarios, and NVIDIA is no exception. Besides mobile phones, computers, servers and other markets, the automatic driving industry is setting off a new wave of upsurge. In the scene of autopilot chip, Tesla has the advantages of practice. It remains to be seen who has stronger deep learning ability and AI comprehensive strength.

?