New Players Added to Domestic Intelligent Computing Chip Track

It is reported that houmo intelligence was established at the end of 2020, focusing on the creation of original new intelligent computing chips and integrated software and hardware platforms. Houmo intelligence focuses on memory computing integrated technology and storage technology, and is committed to breaking through the bottleneck of Intelligent Computing chip performance and power consumption, so as to accelerate the inclusive implementation of artificial intelligence. The high-energy efficiency chips and solutions with high computing power and low power consumption can be applied to large edge terminals such as pan robots / unmanned cars, as well as cloud reasoning and training.

The completion of financing by houmo intelligent means that a new player will be added to the hot domestic chip track. Looking back on the whole 2020, the chip industry also has continuous financing.

According to the investment and financing report of China's chip semiconductor brands in recent ten years released by enterprise investigation big data research institute, in 2020, 458 investment and financing events occurred in domestic chip semiconductors, involving 392 enterprises, with a total financing amount of 109.769 billion yuan.

Nowadays, chip semiconductor has become a crowded track.

For domestic chip enterprises, it may be one of the optional paths to stand out from many industry players, find a new computing paradigm, and catch up with and surpass giants in new segments.

01. Cloud, edge and end are in urgent need, and energy-efficient chips have become a breakthrough for domestic chips

In 2020, the United States cut off the supply of Huawei's chips and imposed sanctions on SMIC, making chip localization a science and technology topic of national concern.

Data show that China accounts for one third of the global chip market and is the world's largest chip importer, with an annual chip import value of more than 200 billion yuan. Among them, a considerable part comes from the United States. Although the market share has decreased year by year, the United States still controls 50% of the global chip market.

Today, American enterprises are still the industry leader in the traditional chip market for PC, mobile phone, server and other hardware. However, in the emerging segment of Intelligent Computing chip, the power of Chinese chip players is rising.

At present, Huawei Hisilicon, Cambrian, horizon and other enterprises have completed the flow of a variety of chips and put them into the market. The industry's market expectation for domestic chips has also turned to occupation and Transcendence in subdivided fields.

At the same time, domestic intelligent computing chip manufacturers are also constantly looking for new breakthroughs. High power efficiency chip with large computing power and low power consumption is one of them.

With the gradual commercialization of artificial intelligence and the continuous upgrading of 5g and Internet of things technologies, there will be more and more intelligent devices equipped with complex AI capabilities. These devices have a high demand for computing power, but limited by power supply, endurance and other problems, the existing intelligent computing chips and computing platforms are difficult to meet their needs.

For example, the common unmanned cargo robot in modern logistics center needs to do real-time Perception Analysis and calculation through cameras and various sensors. Traditional chips can not meet the computing power needs of such devices, and using large computing power chips can not give consideration to power supply and endurance. The contradiction between computing power and power consumption directly limits the intelligence of the equipment.

In fact, not only at the edge, AI computing in the cloud also urgently needs AI solutions with large computing power and low power consumption.

At present, the demand for Intelligent Computing in most cloud computing centers is increasing, and the high power consumption and price brought by the high computing power of mainstream intelligent computing chips on the market (NVIDIA Tesla series GPU) are also very "moving", which has also become a dilemma perplexing cloud computing centers. At the same time, the high power consumption of the existing intelligent computing chip also brings great challenges to the computing density, heat dissipation and cooling system of the server.

02. How to have both high computing power and low power consumption?

In the chip industry, the traditional view is that large computing power and low power consumption can not have both.

Take the automatic driving scenario as an example. At present, the industry generally believes that the computing power of the vehicle computing platform should reach at least 1000 TOPS to achieve the highest level of L5 automatic driving. At the GTC 2020 conference, NVIDIA once demonstrated the robotaxi computing platform with up to 2000 tops, and its power consumption reached an amazing 800W.

The high-power computing platform not only reduces the endurance of the vehicle, but also brings great challenges to automobile design links such as heat dissipation. It has always been the common pursuit of the chip industry to make the chip have both high computing power and low power consumption and achieve high performance.

The "compute in memory" (also often translated as "in memory computing") technology is regarded as an effective technical means to solve the chip performance bottleneck and improve the energy efficiency ratio.

Wu Qiang, founder of houmo intelligence, told 36 krypton that at present, compared with more cutting-edge technologies such as quantum computing, the integration of storage and computing is in a critical period of migration from academia to industry. The chip equipped with memory computing technology is expected to increase the computing power of the chip per watt by 5-10 times.

What is the integration of storage and computing called "black technology"? The story starts in 1946.

In that year, von Neumann, the father of computer, designed the world's first modern computer - EDVAC. On this computer, von Neumann laid two theoretical foundations for the computer industry - the design mode of separating binary operation from memory and calculator, and the "von architecture" was born.

Since then, Feng's architecture has dominated the computer industry. However, in the history of computer development, the chip as a calculator has developed rapidly, while the reading and writing speed of memory has increased slowly. Especially on the AI computing platform, the speed of memory data access can not keep up with the speed of data processing, which leads to the problem of "storage wall" becoming more and more serious, which has also become the constraint of the whole computing platform.

If the chip is imagined as a production workshop in a factory, the memory is a warehouse for stacking production raw materials. People come and go between the workshop and the warehouse, which is bound to waste a lot of resources. The memory computing integration technology integrates the memory and calculator on the same module, breaking the "storage wall", which can directly and greatly reduce the chip power consumption.

"The integrated storage and computing architecture is a subversive system, which is expected to break the bottleneck of traditional computing power and make the chip more suitable for all kinds of Intelligent Computing scenarios at the edge and in the cloud." Chen Erdong, partner of IMO venture capital, a smart investor of houmo, told 36 krypton.

To some extent, the integration of storage and computing technology breaks the traditional von Neumann architecture, which also means that the AI industry has ushered in a new upgrade.

In the AI industry, algorithms, data and computing power have always been regarded as the three basic elements of the industry. In the past, these three elements were separated from each other - the algorithm stayed at the software level; Data is stored in memory; The computing power mostly depends on GPU chip.

In recent years, there are more and more special chips for Intelligent Computing in the AI industry. Most of them adopt the form of FPGA or ASIC, which are customized or optimized for specific AI algorithms. This means that in the AI industry, the two elements of algorithm and computing power begin to integrate.

Now, the emergence of memory computing integrated chip also means that data has begun to be integrated into the chip. The relationship among algorithm, data and computing power is also closer.

In the field of integrated memory and computing technology, houmo intelligence has selected the implementation path of core technologies based on SRAM / RRAM to achieve the balance between read-write performance, power consumption, erasability, cost and process maturity. In order to ensure the chip landing as soon as possible, houmo intelligence has also established a star level founding team.

In the founding team, founder Wu Qiang is the former CTO of horizon. In his early years, he worked for Intel, AMD and Facebook. Wu Qiang's doctoral thesis in Princeton University, namely, high energy efficiency ratio computing chip and compiler, won the best paper of the year at the computer architecture summit conference micro. He worked in AMD and Facebook successively. After returning to China, he joined horizon as CTO. He has accumulated profound experience in architecture and software, and has a deep understanding and engineering practice of edge and cloud application scenarios.

In addition, there are many core members from Penn State University, National University of Singapore, University of California, Tsinghua University, Peking University, University of Electronic Science and technology and other well-known universities at home and abroad, as well as front-line chip enterprises such as AMD, NVIDIA, Huawei hisilic and horizon. The team members have nearly 15 years of research accumulation in the direction of advanced memory devices and memory computing integration technology, and have won the edaa Excellent Doctoral Thesis Award and the "international best young scholar" at the International Microelectronic solid state device conference ", the results were published at the international chip top academic conference ISSCC, ISCA, hpca, etc. the team members have nearly 20 years of experience in high-performance CPU / FPGA / ASIC chip core design and mass production, and have led the design and delivery of many world-class chips (0.18um-6nm), including GPU, CPU, high-performance vehicle specification AI chips, etc.

"This founding team gathered in academia and industry makes houmo intelligent not only have core technology advantages, but also quickly realize chip streaming, mass production and commercial realization. This has also become the core technology advantages of houmo intelligent," said he Zhiqiang, senior vice president of Lenovo Group and President of Lenovo venture capital group.

03. Who will stand out among domestic intelligent computing chips?

From the "sandpaper core making" in the early years to the flourishing of domestic chips, China's chip industry is ushering in an unprecedented spring.

Wu Qiang told 36 krypton: "the current domestic chip industry can be compared with the Internet in 1999." the soil for chip entrepreneurship in China is also solid enough. China not only has the political appeal and public voice of chip autonomy, but also has the talent and technology reserves accumulated over more than ten years. "Therefore, in the past two years, more and more Chinese chip enterprises have gone beyond simple application chips and began to design high complexity computing chips."

The outbreak of China's chip industry, especially the AI chip industry, is inevitable. The report released by AI media consulting in September 2020 points out that the global AI chip market will be US $11 billion in 2019; by 2025, this figure is expected to reach US $72.6 billion, with an average annual compound growth rate of 93.3%.

China is the largest AI chip market in the world. According to the above report data, the scale of China's AI chip market will be about 11.55 billion yuan in 2019 and 133.88 billion yuan in 2023. Cloud computing, intelligent security and robots will become the main landing markets of AI chips.

Such a large market will naturally accommodate a large number of new players. In the future, what characteristics of chip manufacturers can stand out? Wang Huadong, a partner of Jingwei China, an investor of houmo intelligence, points out that there are two main winning opportunities in the future: one is to make subversive products in a specific field with new technology; the other is to form the most powerful team to make the same achievements of foreign enterprises And other products with higher computing power, and then Mozi energy is a combination of the two.