VeriSilicon's Dai Weimin: VeriSilicon's business model has "counter-cyclical" attributes

"VeriSilicon's chip design service model has a 'counter-cyclical' attribute. During the epidemic or industry 'downturn' period, many chip customers have been keeping a low profile, accumulating strength, and actively deploying new products in order to surpass their competitors when the epidemic ends. However, during difficult times, it is inconvenient for companies to expand, so they need to seek first-class chip design service companies for cooperation; in addition, the low tide period is also a good time to acquire IP and IP companies."

Recently, Dai Weimin, founder, chairman and president of VeriSilicon, stated this at the "Artificial Intelligence Chip Innovation Theme Forum" sub-forum of the 2020 World Artificial Intelligence Conference Cloud Summit.

Dai Weimin, Founder, Chairman and President of VeriSilicon

AI layout from cloud computing to edge computing

The era of big data has ushered in an explosive growth in data volume. According to IBS data, the amount of data will grow 1,455 times from 2018 to 2030. With the surge in data volume, the way data is processed is also constantly changing. From data production and collection, to data storage and calculation, to data transmission and application, the proportion of human-machine participation is developing from the initial partial participation of machines to full participation of machines. Among them, artificial intelligence (AI) technology is used to convert data into high value.

Dai Weimin believes that in the field of cloud computing chips, ASIC chips are gradually becoming a trend. "Compared with other types of chips such as CPU, GPU, FPGA, etc., ASIC chips have great advantages in computing performance, size, cost, etc. In the future, with the development of general AI instruction set architecture, it is expected that the optimally configured AI computing chips will appear." He said.

Especially in the context of the hot short video market, the video resolution has been upgraded from 1080P to 2K, 4K, and even 8K. Therefore, major video cloud service providers including Inspur, Alibaba, and Kuaishou are gradually shifting from using CPU and GPU chips to customized ASIC chips.

According to Dai Weimin, "Compared with high-end CPU solutions, VeriSilicon's customized video transcoding ASIC has a smaller size, not only the power consumption is 1/13 of the former, but the transcoding capability is also 6 times that of the former. After the system is built, the transcoding capability of the entire system can be increased by 36 times, and the power consumption can be reduced by 50%."

Dai Weimin pointed out that although VeriSilicon is not a pure AI company, it currently has more than 30 AI customers and has realized the layout of artificial intelligence from cloud computing to edge computing. In the global AI chip company rankings released by market research company Compass Intelligence, VeriSilicon ranked 21st and ranked third in China.

It is worth mentioning that world-renowned semiconductor manufacturers NXP and Broadcom have also adopted VeriSilicon's AI processor IP. Among them, NXP's latest i.MX products are equipped with VeriSilicon's Vivante intelligent pixel processing IP solution with VeriSilicon's VIP neural network processor as the core, while Broadcom has selected VeriSilicon's Vivante VIP8000/VIPNano series AI processor IP for its next-generation set-top box system-on-chip (SoC).

Advanced technology drives the development of edge AI processors

The development of the entire semiconductor industry has brought certain changes to AI technology.

Taking Apple as an example, the number of transistors in Apple's mobile phone chips has increased significantly as the characteristic line width has decreased, which supports the continuous upgrading of mobile phone chip performance. IBS data shows that under the 16nm process, the number of transistors in Apple's mobile phone chips is 3.3 billion, under the 7nm process is 6.9 billion, and under the 5nm process is expected to reach 10 billion. The rapid increase in the number of transistors per unit area will cause the unit cost of transistors to drop rapidly. The production cost of each transistor of Apple chips is US$4.98/1 billion transistors under the 16nm process, and only US$2.65/1 billion transistors under the 7nm process.

Taking the design cost of the process in the mainstream application period as an example, when the process node is 28nm, the design cost of a single chip is about 41 million US dollars, while when the process node is 7nm, the design cost quickly rises to about 222 million US dollars. The cost of early use and mature use is more than doubled, but the cost of mature use is still very expensive.

To this end, Dai Weimin believes that with the development of advanced processes, the cost per unit of computing power will gradually decrease, making AI processors for edge computing more cost-effective. "For example, the cost per TOPS based on the 16nm process node is $0.67, and the cost per TOPS based on the 5nm process node is only $0.08, where 1TOPS means that the processor can perform one trillion operations per second."

Dai Weimin said that VeriSilicon's vision is to build an open, innovative, and privacy-protecting AI technology ecosystem.

Industrial transfer gives birth to light design model

The world's semiconductor industry is undergoing its third shift. From the 1960s to the 1970s, the transition from the military industry to the home appliance industry pushed the semiconductor industry to move from the United States to Japan, and system manufacturers began to differentiate into IDM manufacturers; from the 1970s to the 1980s, the home appliance era entered the early stage of PCs, and IDM manufacturers began to differentiate into wafer foundries, fabless and design service companies, and the semiconductor industry also moved from Japan to South Korea and Taiwan, China; around 2000, humans entered the mobile phone era from the PC era, the industrial division of labor was further refined, and light design gradually became the mainstream. With the advent of the era of smart Internet of Things, the semiconductor industry officially moved to mainland China.

VeriSilicon's business model is closely aligned with the development trend of the industry. According to Dai Weimin, VeriSilicon was founded in 2001 and has had a unique business model since its inception: Chip Design Platform as a Service (SiPaaS). The SiPaaS model specifically refers to a business model that provides customers with one-stop chip customization services and semiconductor IP licensing based on a technology platform built on VeriSilicon's proprietary semiconductor IP.

Through years of independent research and development and mergers and acquisitions, VeriSilicon can not only provide world-class one-stop solutions for system-on-chip (SoC) and system-in-package (SiP), but also build its own IP moat for consumer electronics, automotive electronics, computers and peripherals, industry, data processing, Internet of Things and other fields, and successfully develop into a first-class "IP Power House". According to IPnest statistics, VeriSilicon's semiconductor IP sales revenue in 2019 ranked first in mainland China and seventh in the world.

Dai Weimin said that VeriSilicon has advanced chip design capabilities and has a lot of successful tape-out experience in 14nm/10nm/7nm FinFET process technology, with 30-50 chips taped out each year, and has already started research and development of projects based on 5nm FinFET process.

Expanding the broad market with chiplets

As chip size and process technology continue to shrink, semiconductor manufacturing is facing increasingly severe challenges. The doubling rate of transistors has slowed to about 30 months. Therefore, many people can't help but wonder whether Moore's Law can continue, and some even think that Moore's Law is dead.

Regardless of whether Moore's Law will eventually fail, there is currently a technology that may help continue Moore's Law, namely Chiplets. What are Chiplets? Chiplets are "small chips" or "core particles". They are similar to IP, but they are provided in the form of chip dies rather than software. Many large companies have already developed this technology, such as TSMC, AMD, Intel, etc.

So, what is VeriSilicon's solution? According to Dai Weimin, VeriSilicon has proposed the IP as a Chiplet (IaaC) concept, which aims to achieve "plug and play" of special function IP with Chiplet. It strives to solve the balance between performance and cost in 7nm, 5nm and below processes, and reduce the design time and risk of larger-scale chips.

In May 2020, VeriSilicon passed the review of the Shanghai Stock Exchange Science and Technology Innovation Board Listing Committee and is about to be listed on the Science and Technology Innovation Board. "Although VeriSilicon does not make products, it is another realm to provide services and platforms. We hope that VeriSilicon can help China's AI wave rise and promote the development of the entire AI industry chain." Dai Weimin finally said.