From the second half of last year to this year, AI chips blossomed everywhere. Many AI start-ups in China have introduced their own AI chips, but Victor Peng, president and CEO of Cellings, the world's largest FPGA manufacturer, poured cold water on them, saying that start-ups should not start making AI chips from scratch.
On October 16, Victor told reporters at the Selins Developers Conference that many start-ups do not have the money to develop and mass-produce AI chips because of the huge cost of research and development. AI start-ups should focus on innovative algorithms and architectures, not on chip design. How many start-ups have succeeded in doing ASIC?
Grasp AI opportunities
Technological advances have made competition between CPU, GPU, FPGA (Field Programmable Gate Array) and ASIC (Application Specific Integrated Circuit) processors increasingly fierce. As a leader in FPGA, Victor says that Cellings'competitors are no longer Altera, the second-largest maker of FPGAs (acquired by Intel), but the processor businesses of Invid and Intel.
Because of high flexibility, in the current AI algorithm is not mature and fixed, FPGA is considered as an intermediate solution, its greatest advantage is that it can make the hardware function of the system can be programmed like software. Compared with GPU and CPU chips, it has higher performance and lower energy consumption.
Xilinsi thinks that FPGA chip will be the highlight. To better meet the challenges of the AI era, the company launched its first new category platform, Versal ACAP (Adaptive Compute Acceleration Platform), on June 16.
Versal's product portfolio is based on TSMC's 7 nm FinFET technology and is a platform that combines software programmability with hardware acceleration and flexibility in specific areas. With the explosive growth of AI and big data and the slowdown of Moore's Law, the industry has reached a critical turning point, Victor points out. The cycle of chip design is no longer able to keep pace with innovation. Versal supports all types of developers to speed up their overall application by optimizing software and hardware, while providing immediate flexibility.
Because of the AI draught, the price of yvda has risen sharply in the past two years. In February, Google announced that it would open its TPU (Tensor Processor) service to the public and join the AI chip war. TPU is Google's customized chip for machine learning, and it is ASIC.
In an interview with First Finance and other reporters, Victor said that the AI chip market will not only have one chip architecture dominated the world, but not favor dedicated chips.
Compared with CPU and GPU, the biggest advantage of FPGA is highly adaptable strain capacity. GPU does have some advantages in accelerating certain applications and workloads. In machine learning, the GPU does integrate some new module templates to speed up machine learning, but its performance is fixed over a fixed period of time. FPGA can be accelerated for different workloads, which is much better than GPU, and can be used in different networks during machine learning.
For Chinese start-ups to get together to make AI chips, Victor is not technical. Insiders told reporters that some start-ups are mainly for financing.
If we really want these enterprises to create value in the high-tech field, we must do something that others have not done, rather than do what several large enterprises do, which is a waste of resources and capital. Victor told reporters that it's not that start-ups can't do ASIC, if they can do better than Intel, Invid and Salings, but that more start-ups should focus on specific areas and applications, rather than developing chips from scratch, because there are so many companies doing it.
In July 18th, xilinsi announced the acquisition of Shenzhen AI technology. As an influential AI chip startup in China, the acquisition of Deep Learning Technology makes many people feel very sudden and unexpected. Some people in the industry also said that Shenzhen Technologies may encounter a bottleneck in business development. At the same time, this incident shows us the fact that the chip is not so simple.
Salil Raje, vice president of software and IP products at Cerings, told First Financial Reporter in an interview that most AI chip companies will be eliminated. AI will change dramatically over the next few years, so flexible hardware is needed, and AI-specific chips will be used in a particular vertical area, such as the Smart City camera chip, but not as a general purpose chip.
Speed up integration with deep science and technology team
At the same time, Victor has also made some responses to the newly acquired AI start-ups in Shenzhen. He told First Financial Reporter that the technology has not yet brought substantial revenue, but hope that the team and Salings can complete the integration as soon as possible, to promote the company's business growth in the Chinese market.
Deep learning technology was founded in 2016, focusing on neural network pruning, deep compression technology and system-level optimization, in recent years also committed to the development of AI chips.
Victor said the company would report to Salil after the acquisition because Shenzhen Technologies is primarily an engineering department, and that the market and other business sectors of Shenzhen Technologies would be taken over by Salings and further expanded. He said that Sallings would help to learn more about technology and serve customers around the world, and would also use its technology and do joint marketing for businesses.
Victor told reporters that the main reason for the importance of deep lessons lies in their network optimization, DNN and some of the architecture and practical technology. However, as Deep Learning focuses on R&D and does not have much revenue, it will not bring substantial revenue to Salings in the short term. But Victor also said that through post-merger collaboration, Salings would recommend technology-savvy products to more customers, boosting the latter's revenue, which in turn would enhance Salings'overall competitiveness.
After being purchased by Xilinx, Shenzhen technology will focus on the FPGA field, and will no longer develop its own AI chips. Salil told reporters that they realized that the chip (ASIC) chip after streaming, in some new neural network layer performance is not good, very inefficient, which is why we do not intend to let them develop ASIC chips.