Silicon photonics is the key to unlocking the full potential of AI

Rate this post

A large language model enabling generative artificial intelligence (AI) is fueling investment and increasing competition in the field of silicon photonics, a technology that integrates and widely deploys silicon-based integrated circuits (ICs) and optical components. Volume of data more efficiently.

NVIDIA, TSMC, Intel, IBM, Cisco Systems, Huawei, NTT and imec, a Belgium-headquartered interuniversity microelectronics center, have joined the race with top-ranked designers and makers of ICs, AI systems and telecommunications equipment.

These and other organizations have been working on silicon photonics for many years, some (including Intel and NTT) for nearly two decades.

On September 5, at the Silicon Photonics Global Summit in Taiwan, Douglas Yu, TSMC’s vice president in charge of pathfinding for system integration, told the Nikkei newspaper that “If we can provide a good silicon photonics integration system … we can solve both critical problems. Energy efficiency and computing power for AI. This is going to be a new paradigm shift. We may be at the beginning of a new era.”

In the invitation to the summit, SEMI (Microelectronics Industry Association) noted that “Silicon Photonics has become a buzzword in the semiconductor industry” due to its high bandwidth, high-speed data transmission, wide transmission distance, low power consumption and utility in advanced networking, computing architectures. , cloud computing, data centers, autonomous vehicles and smart transportation systems.

In other words, at the very leading edge, the high-tech industry will benefit from silicon photonics in ways that improve device and system performance by reducing energy consumption.

According to Semi, “The global silicon photonics market is projected to grow from a valuation of $1.26 billion in 2022 to reach a value of US$7.86 billion by 2030 with a compound annual growth rate (CAGR) of 25.7%.” of

About a year ago, in September 2022, DigiTimes reported that NVIDIA and TSMC had launched a joint R&D project called COUPE, which stands for Compact Universal Photonic Engine. The goal of the project is to integrate multiple AI processors (GPUs) using NVIDIA’s Silicon Photonic (SiPh) technology.

DigiTimes wrote, “Sources revealed that the SiPh chip and CMOS process go through co-packaged optics (CPO) technology integration, which can connect multiple advanced GPUs with chip-on-wafer-on-substrate (CoWoS) 2.5D IC packaging. . .”

The combination of low-latency optical data transmission and advanced packaging technology significantly reduces signal loss, making it possible to build “ultra-large GPU sets,” DigiTimes wrote.

Silicon photonics is an old technology with huge new cutting-edge potential. Image: Twitter

The technology won’t be ready until the “SiPh ecosystem matures,” which helps explain why Taiwan’s leading outsourced semiconductor assembly and test (OSAT) company ASE Technology and Japan’s Advantest were also prominent at the Silicon Photonics Global Summit.

At the Semicon Taiwan trade show held after the Silicon Photonics event, TSMC told the press that the lack of CoWoS packaging capacity is likely to limit NVIDIA’s ability to supply AI processors until the end of next year. By then the capacity should double.

Intel defines silicon photonics as “the combination of two of the most important inventions of the 20th century – the silicon integrated circuit and the semiconductor laser. It enables faster data transfer over longer distances than conventional electronics while leveraging the efficiency of Intel’s high-volume silicon manufacturing.”

Intel explains that “the power of optics and the scalability of silicon… optical transceivers are optical interfaces for Ethernet switches, routers, and transport networking equipment, providing connectivity for large-scale cloud and enterprise data centers.”

Dylan Patel of SemiAnalysis wrote that “Intel is the world’s largest manufacturer of silicon photonics. They are the market leader in the manufacture of optical networking transceivers … Intel’s scale and the integrated nature of their solutions make them the industry leader … (failure rate) two orders of magnitude better than the competition.

Intel plans to quadruple its advanced IC packaging capacity by 2025 with a new 3D packaging facility in Malaysia. In late August, Intel’s corporate vice president for supply chain and operations, Robin Martin, told Tech Wire Asia that “Malaysia will eventually become Intel’s largest production base for 3D chip packaging.”

In addition to expanding its own production of optical transceivers and other silicon photonics products, Intel is providing the technology to others. Last March, China’s Fast Photonics Technologies announced plans to build optical transceivers based on Intel technology.

Fast Photonics’ headquarters and factory are located in Shenzhen. It also has a sales and technical support center in San Jose, California, near Intel’s headquarters in Santa Clara.

In May 2023, Japan’s national telecommunications carrier and telecommunications technology developer NTT announced plans to build a series of increasingly sophisticated photonics-electronics conversion devices over the next few years.

These devices, which eliminate the need to convert signals from optical to electrical and back again, should pave the way for drastic reductions in power consumption of telecommunications networks and data centers.

NTT aims to increase energy efficiency by 100 times, increase transmission capacity by 125 times and reduce end-to-end delay (latency) in mobile and optical networks by 200 times by 2030.

Along with Intel and Sony, NTT has established the “Innovative Optical and Wireless Networks Global Forum to accelerate the adoption of new communications infrastructure that will bring together an all-photonics network infrastructure with silicon photonics, edge computing and wireless distributed computing.” “

Huawei and imec added silicon photonics to their joint research on optical data link technology in 2014. Huawei then acquired Caliopa, a developer of silicon photonics optical transceivers spun out of IMec and Ghent University.

But the Huawei-imec collaboration was terminated and Huawei was placed on the US Commerce Department’s entity list and banned from shipping ASML’s EUV lithography systems to China in 2019. ASML has a close relationship with imec.

Huawei continues its own research, which is important to its telecommunications equipment business and efforts to avoid sanctions imposed by the US government. In a video posted on YouTube in October 2022, it announced that: “Huawei intends to build a photonic chip to circumvent US chip restrictions.”

He adds that photonic chips outperform silicon-based ICs, which would be nearly impossible to fabricate at sub-2nm process nodes… Moreover, the (photonic) chip manufacturing process is completely different… It will use brand new manufacturing technology and lithography machines. No need at all…”

The West does not yet dominate the field of silicon photonics. Image: FormFactor

“The most important point,” the video concluded, “is that research in this field is still in its early stages, and European and American countries have not yet established a monopoly.”

This assessment appears to be at least a decade ahead. This is optimistic but plausible, and with the recent launch of the 5G Mate 60 Pro smartphone with its 7nm processor despite US restrictions, Huawei’s intentions should not be discounted.

On August 23, Bloomberg published an article under the headline “Huawei is building a secret network for chips”. According to the article, the US Semiconductor Industry Association (SIA) has warned that Huawei has acquired at least two fabs and is building three more, backed by $30 billion in Chinese government funding.

If the reports are accurate, it would be foolish to believe that Huawei is simply doing this to import standard semiconductor manufacturing equipment under different names. Huawei and China have every incentive to invest heavily in silicon photonics.

Follow this author on Twitter: @ScottFo83517667

Leave a Comment