From the second half of last year to this year, AI chips have blossomed everywhere, and many domestic AI startups have launched their own AI chips.
But at the Xilinx Developers Conference on October 16, Xilinx President and CEOVictorPeng said that many start-ups have no funds to develop and mass produce AI chips because of the huge cost of research and development. AI startups should focus on innovative algorithms and architectures rather than designing chips. “How many startups have succeeded in doing ASICs (dedicated custom chips)?”
Advances in technology have made competition among different processor vendors such as CPUs, GPUs, FPGAs (Field Programmable Gate Arrays) and ASICs (Application Specific Integrated Circuits). As the leader of FPGAs, Victor said that Xilinx’s competitors are no longer the second largest manufacturer of FPGAs, ALTERA (acquired by Intel), but the processor business of NVIDIA and Intel.
Due to its high flexibility, FPGAs are considered to be an intermediate solution when the AI algorithm is not mature. The biggest advantage is that the hardware functions of the system can be modified by software like software. Compared with GPU and CPU general-purpose chips, it has higher performance and lower energy consumption.
Victor said: “With the explosive development of AI and big data and the slowdown of Moore‘s Law, the industry has reached a critical turning point. The cycle of chip design has been unable to keep up with the pace of innovation.”
Because of the AI vent, the stock price of Nvidia has risen sharply in the past two years. Just in February of this year, Google announced the opening of the TPU (tensor processor) service, joining the battle of AI chips. TPU is Google’s custom chip for machine learning and is an ASIC.
In an interview with reporters from China Finance and Economics, Victor said that the AI chip market will not only have one kind of chip architecture, but it is not optimistic about dedicated chips.
Compared with CPU and GPU, the biggest advantage of FPGA is its highly adaptive strain capability. “The GPU does have its own advantages for some applications and workload acceleration. In the machine learning world, the GPU does integrate some new module templates to speed up machine learning, but its performance is fixed for a fixed period of time. FPGAs can be accelerated for different workloads, and they perform much better than GPUs in this area, and can be applied to different networks during machine learning.”
For Chinese start-ups to build AI-specific chips, Victor bluntly, it is not a technical reason. Some insiders told reporters that some start-ups are mainly for financing.
“If you really let these companies create value in the high-tech field, you have to do things that others have not done, rather than doing things that several big companies are doing, which is a waste of resources and capital.” Victor told reporters. It‘s not that startups can’t be ASICs. “If you can do better than Intel, NVIDIA, and Xilinx, more startups should focus on specific areas and applications, rather than developing chips from scratch, because there are A much bigger business is doing it.”
Victor also made some responses to the newly acquired Chinese AI startup, Shen Jian Technology.
Victor told the First Financial Reporter that Shen Jian Technology has not yet brought substantial income, but hopes that the team and Xilinx can complete the integration as soon as possible to promote the company‘s business growth in the Chinese market.
On July 18, Xilinx announced the completion of the acquisition of Shenjian Technology, a domestic AI startup.
Victor told reporters that the reason why they value the deep insights lies in their network optimization, DNN and some architecture and practical technology.
After the acquisition, Shenjian Technology will focus on the FPGA field and will not develop chips by itself.