If the backward state of high-end integrated circuits is due to the late start of China's related industries and the low accumulation, the Chinese have already participated deeply in the wave of artificial intelligence that is once again emerging, and have occupied the top echelon of the world in several technologies and applications. position.
However, if the AI industry chain is stratified, the basic layers represented by AI chips, big data, and algorithms directly determine the development of the intermediate technology layer and the application of ecological superstructure. Then the AI voice right in each major power of the world On this battlefield where no smoke can be seen, the AI chip has become a strategic point of command that cannot be lost.
In an interview article of Tianyi Zhixin CEO Li Yunpeng in the recent AI science and technology base camp (Interview with Tian Yunzhi Li Yunpeng: Self-controllable, world-class AI chip in two years), Li Yunpeng stated that it takes two years Create self-controllable, world-class AI chips. It is natural to hear that Zhuang language is natural at this stage. However, the outside world's composition of the core team of Tianzhi Zhixin, the arrival path of the AI chip, etc. , This AI science and technology base camp once again entered the number of days core intellectual core, exclusive interview with the core members of its hardware team, in order to readers unveil the mystery of this core dark horse.
Those who accepted the interview were Mr. Dong Xu, head of the technical department of Zhixin Systems, and Alex Lee, an AI system architect.
构成 The composition of the hardware team: 100-person team, Silicon Valley, Shanghai
At present, Tianzhizhixin's hardware team is mainly concentrated in Silicon Valley and Shanghai. The Shanghai team is based on senior GPU chip experts. The team size is about 90. The core research directions of the Silicon Valley team are artificial intelligence and high-end computing chips. The scale is about 20 people.
As the person in charge of Tianxie Zhixin's system technology department, Dong Hao graduated from Zhejiang University in 1985 with a profound chip background. He has worked in Silicon Valley and has lived for more than 20 years. Prior to joining Tianxin Zhixin, he worked in chip giant AMD. 20 years. 'It can be said that my major career is doing the chip', Dong Ye said that many of the core members of the hardware team are from the chip industry, worked together for more than 10 years, and they know each other well. In terms of perspectives, Tianji Zhixin is not a start-up company.
Alex Lee graduated from Tsinghua University in Taiwan, China. He received his master's degree from UCLA, and he entered Intel Labs as a full-time senior researcher for one year. He specializes in the implementation of machine learning in full-duplex communication baseband chips. Michigan During his Ph.D. degree at the University, he was an intern at the Nvidia Architecture Research Group to research and develop deep neural network processor architectures. The current research direction is domain-specific hardware and software acceleration systems.
▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌ ▌
With regard to the current industry-based craze, Alex said that in order to succeed, it needs to have three conditions:
First, the market traction. Any entrepreneurship is not the so-called 'technical promotion'. Only the 'market traction' requires people to have real technological innovation. 'Academia has done a lot of chip research, throwing away the papers. We jokingly call them 'silicon trash'. Alex stressed that NVIDIA was able to become a world-class chip company because they not only focused on R&D, but all the research was for products, not for research. Research. If there is no market for support, it will cost hundreds of millions of dollars to make a film and it may fail. The startup company has no conditions to make such an attempt.
Second, the team. To determine whether a start-up company can successfully build cores, it is necessary to first study their member structures first. If there are not a group of professionals working in the chip industry for 15-20 years, they have excellent understanding and efficient execution. Force, the possibility of success is quite embarrassing. For the current number of days core intellectual core, the core team composition most of them are very senior experts. On this basis, to inject new blood, Alex said: 'We now have a research and development team that is closest to the ideal.'
Third, capital. Chips are a very burning industry. There is no long-term investment in large sums of money that cannot be sustained. At the same time, chips are an area of infrastructure that needs long-term cultivation. Its development requires capital, talent and technology. All-round support. For investors' attitudes, Alex urges: 'Whether it is a venture capital institution or a national strategic investment, it is important to give chip startups enough time to accumulate and accumulate.'
Regarding funding issues, Dong Hao, who is familiar with the chip's process flow, said that a tape-out cost 3 million to 5 million US dollars. If it is the latest 7-nanometer technology, it needs tens of millions of dollars. The premise of this price is to succeed once. To achieve performance standards, cost-controlled products will be competitive, but the middle of the production process is also very test the team's management capabilities, a little carelessness will lose money.
到达 AI chip's arrival path: FPGA is replenished along the way, ASIC is the final solution
For the architecture of AI chips, currently there are several solutions in the industry such as CPU, GPU, FPGA, ASIC, etc. Dong Xuan disclosed: 'The number of days core wisdom core is to do GP-GPU, because only GP-GPUASIC can meet the high-end computing The market demand also provides the highest price/performance ratio, but in some special occasions, FPGA still has a great advantage and a huge market.
In response, Alex added: 'Sky is currently working on the cloud computing machine, edge computing machine, and AI chip. At the same time, the cloud and edge computing machines all use FPGA solutions. Although FPGA is positioned as A number of relay stations on the strategic road of Zhixin core, however, the acceleration they can demonstrate, the improvement of computing power, the reduction of power consumption, and the integrity of the system all make it worth far beyond a transitional product. At the time of the project, our R&D team developed applications for rapid iteration, and further improved its efficiency and efficiency on High-Level-Synthesis technology. In the edge computing machine product, a customized FPGA board, In addition, high-level synthetically generated artificial intelligence and database acceleration IP is a complete industrial-grade solution. Inserting the card in the cloud server becomes the core of the cloud computing machine, and it can easily provide customers with hardware. Accelerate service (H aaS) services. So cloud, edge-computing machines are the same, they provide a quick and effective solution for existing customers. The plan is the ground cash flow business developed by the company. '
The AI chip is the most valuable vision for the company's development. Its ultimate goal is to make common, standard, high-performance and scalable solutions. However, it takes time to precipitate and baptize. Alex believes that the days of intellectual core FPGA The business will continue to bring feedback from the market and bring valuable iteration opportunities for general-purpose, standard, high-performance computing chips. This is also the overall strategic timeline that CEO Li Yunpeng has repeatedly emphasized.
总结 Summary: As a system company like Apple, creating a complete ecosystem from chips to applications
At this point, the grand blueprint of Tianma Zhixin’s chip dark horse has been fully launched. However, perhaps it is not just the black horse of the chip industry.
As CEO Li Yunpeng said, Tianji Zhixin is not just a chip company, but a system company. For the concept of a system company, Li Yunpeng said that there is a company that does both mobile APP (calendar, weather, stocks...) and The APP store and operating system also make cell phone chips. This company is called Apple, it is also a system company, but Apple is a type C system company, and days number smart core is to B type system company.
In the Silicon Valley team's lead system and chip architecture lead, Nanjing's software team to do the software application of machine learning and big data platform, Shanghai's hardware team is based on the existing software market and ecology down to the base layer chip solution , thus forming a complete ecological chain from chip to system to application.
Although aiming at the construction of the entire ecological chain, Dong Tao said that the company does not have much competition with domestic Internet giants and is essentially a different company. Due to the existing problems in the field of AI chips, there is not yet a universal algorithm. For example, algorithm-based AI chips are closely related to application scenarios. One of the difficulties faced by many startup companies is that when solving a computing problem in a certain scenario, users in this area often need support, but users do not see the solution. It is difficult to provide effective support in the case of the plan. This becomes a chicken or egg problem.
'And this is the advantage of Tianxu Zhixin, relying on existing products and users, we can develop solutions with independent intellectual property rights, so the success rate is very high,' Dong Xu said.