Nvidia intends to set up a R&D center in China to lay out optical chip technology in advance

On December 15th, NVIDIA held China’s annual GPU technology conference online. Bill Dally, NVIDIA’s chief scientist, delivered a keynote speech at the conference for the first time. Dally introduced how NVIDIA’s AI accelerated processor technology empowers healthcare, driverless cars and robots. many industries.

Dally revealed to the first financial reporter: “NVIDIA is considering establishing a research and development laboratory in China. China has excellent universities and talent reserves. We hope to carry out technology research and development locally to serve the application of the local market.”

At present, NVIDIA has a large number of partners in the fields of artificial intelligence, autonomous driving, and biopharmaceuticals in China. However, Dally did not disclose the specific planning and timetable for the construction of NVIDIA’s R&D center in China. “It depends on the progress of our talent recruitment.” Dally told the first financial reporter.

Nvidia also showcased its ecosystem in China at the GTC conference, including Alibaba, Baidu, ByteDance, China Telecom, Didi, Inspur, Kuaishou, Ping An, Tencent and Xiaomi and other Chinese companies are Nvidia’s partners.

Nvidia’s goal is to double the AI ​​inference performance every year through improvements in GPU technology. At the GTC conference, NVIDIA announced the launch of an ultra-high-efficiency acceleration system, MAGNet, that can make AI reasoning capabilities reach an efficiency of 100 trillion times per second, an order of magnitude higher than current commercial chips. “Data transfer is the most energy-intensive link in today’s chips, and MAGNet uses a series of new technologies to coordinate and control the flow of information through the device to minimize data transfer,” Dally said.

In addition, Nvidia showcased the latest research advances in photonics to replace electrical links within existing systems with faster optical links. The technology, called “dense wavelength-division multiplexing,” promises to transmit Tb/s-level data on a chip just one millimeter in size, more than ten times the density of current interconnects. Dally revealed that Nvidia’s research team is working with Columbia University to explore how to use the technology used by telecom providers in their core networks to transmit dozens of signals over a single fiber.

Jin Xianmin, a professor at Shanghai Jiaotong University and director of the Integrated Quantum Information Technology Research Center, told Yicai.com: “NVIDIA is now focusing on the development of data centers. There is a huge transmission demand between various modules. Optical interconnection can greatly improve transmission efficiency.”

Compared with the existing technologies such as microwave communication, optical communication is the future development direction. Its applications include optical interconnection, optical communication, optical artificial intelligence processor, optical sensing, and lidar, etc., which can greatly improve the efficiency of data transmission. efficiency.

Jin Xianmin quoted data to the first financial reporter that the current global market size of optical modules has reached about 26 billion US dollars, and China has dominated the market, but the core technology of optical chips is still in the hands of international manufacturers. “The question now is how to combine the technology of photonic chips with practical application needs.” He told the first financial reporter.