On September 10, Huawei introduced its own standard for optical communication technology in a challenge to Nvidia and Broadcom, as the Chinese tech giant aims to have a greater say in an emerging area of chipmaking that could shape the future of artificial intelligence computing.
Dubbed near-package optics (NPO), the approach governs how main chips, optical engines, and other components are packaged together to boost data transmission speeds. Huawei said NPO is more feasible than the co-packaged optics (CPO) method backed by US chipmakers Nvidia and Broadcom, in part because it leverages existing chipmaking supply chains.
The Chinese tech titan demonstrated its 7.2T NPO solution for the first time at the China International Optoelectronic Exposition (CIOE) on September 10, the most important optical tech expo in the country.
CPO technology situates optical components close to a chip’s electronic components directly at the packaging level. This shortens data transmission delays and reduces signal loss while enabling higher-bandwidth connectivity in smaller form factors. Better interconnect technologies have been identified by many leading AI chipmakers as key to overcoming one of the biggest bottlenecks in building advanced AI computing systems.
Huawei’s NPO approach follows a similar concept of bringing photonic and electronic components closer together, but does so at the printed circuit board level. The company said this can simplify manufacturing and take advantage of a more mature, readily available supply chain.
NPO is also a more technically mature technology and is “pluggable,” which means if one piece is defective, it can be easily fixed, while in the CPO approach, it would affect the whole chip, according to Huawei.
“NPO is a more practical choice compared with CPO,” Man Jiangwei, a chip and optical executive from Huawei’s advanced optoelectronics laboratory, HiSilicon, told a media roundtable on September 10. “However, the current NPO solutions are fragmented and we need to work together for a standard solution.”
Huawei has accumulated knowledge in the optical interconnect technology field, and its new module is the first NPO solution in the industry to operate at 7.2 terabits per second, according to Man. It will be introduced in the company’s superPoD supercomputer solutions in the future, and includes a laser source, silicon photonics, and other chips, the executive added.
Huawei is working, through the Optical Internetworking Forum (OIF) of the US, to set technical standards for the NPO approach.
Tiger Ninomiya, who is in charge of energy efficient interfaces and c-packaging interoperability demos at the OIF and serves as director of standard of optical technologies with Amphenol, told Nikkei Asia that the potential new standard could help the industry to grow.
“We identify NPO as an evolutionary step forward toward CPO in terms of addressing the needs with technology that is more practical today, while also addressing manufacturability, ecosystem risk and interoperability,” Ninomiya told Nikkei Asia.
The newly proposed NPO standard will need to undergo review before it can become an OIF standard. Meanwhile, a global standard for co-packaged optics has already been established by industry members through an implementation agreement under the OIF.
The Chinese tech giant said it has been engaging with cloud computing players as well as equipment, module, and component suppliers from the US, Japan, China, Canada, and elsewhere to gather support for the NPO standard initiative. Chinese players such as Innolight and Eoptolink play important roles in the global optical communication and AI data center supply chain.
The proposal for a new standard comes ahead of a planned high-level meeting between US President Donald Trump and Chinese President Xi Jinping in late September. It also came shortly before the start of Huawei’s criminal trial in the US, where it is facing allegations of stealing trade secrets from US companies and violating sanctions against Iran. Huawei denies the allegations.
Data communication technologies have taken a bigger role in the AI infrastructure buildout as the focus shifts from improving the performance of individual processing chips to boosting the overall communication and interconnectivity capabilities of server racks and clusters.
In addition to Nvidia and Broadcom, Taiwan Semiconductor Manufacturing Company, and Intel are among the players working to develop CPO. AI chip developer Marvell has said high-speed connectivity is essential for enabling next-generation computing that is ever more powerful.
Nikkei Asia earlier reported that pluggable transceivers and optical fibers are becoming the latest bottlenecks for AI infrastructure. According to Yole, the optical module market is expected to reach USD 112 billion by 2031, representing a compound annual growth rate of 30%.
Both Nvidia and Broadcom’s CPO solutions using both companies’ switch chipsets entered production in the second half of this year.
This article first appeared on Nikkei Asia. It has been republished here as part of 36Kr’s ongoing partnership with Nikkei.