Controlling Light: Is Silicon Photonics an Emerging Front in U.S.-China Tech Competition?

Introduction

Semiconductors are likely the most important enabling technology of the twenty-first century. To date, the semiconductor “arms race” between the United States and China has been largely characterized by a sprint to etch ever smaller feature sizes onto the surface of silicon wafers—the more transistors that can be placed on a single chip, the greater the processing power and energy efficiency. The greater the power and efficiency, the more accurately, quickly, and cost-effectively artificial intelligence (AI) algorithms can be trained and operated.

However, as the end of Moore’s law approaches, and the physical limits of process sizes are reached, the semiconductor industry is looking to ways other than increasing transistor density to enhance performance—especially as advances in AI are driving demand for increased computing power. One promising emerging technology, silicon photonics, has the potential to reduce latency while increasing efficiency by enabling the fabrication of photonic components on silicon using standard semiconductor manufacturing processes. Some in Beijing believe it could alter the contours of U.S.-China competition over control of advanced semiconductors.

Breaking the Blockade?

Silicon photonics is an innovation that has enabled the fabrication of photonic components directly onto a silicon material base using standard semiconductor manufacturing processes. Photonics, in contrast to electronics, make use of photons (light) rather than electrons to carry information. Their integration alongside electronics holds promise to create large-scale computing systems with higher bandwidth and improved energy efficiency that go beyond the physical limitations of traditional electronic chips.

In Beijing, some observers also believe that silicon photonics has the potential to help China break the so-called U.S.-led technology blockade. Chen Wenling, a prominent economist at the government-backed Chinese think tank, the China Center for International Economic Exchange, identified silicon photonics in a March 2023 discussion on counter-containment as a technology that China’s advantage in could allow it “to change lanes and overtake”—a phrase associated with China leapfrogging ahead of the United States in emerging technology areas: 

“Mainland China is preparing to build a photonic chip production line, which is expected to be completed in 2023, which means that China will be at the forefront of the world in terms of photonic chips, and even completely change the chip technology route. Photonic chips have many technical advantages. Its calculation speed is faster and its information capacity is larger, which will be more than 1,000 times higher than the current silicon-based chips,” (author’s translation).

The “photonic chip production line” referenced by Chen is reported to be built by the Beijing-based firm, Sintone. According to Chinese reporting, the president of Sintone has noted that China has the capabilities to produce photonic chips domestically because their production does not require extreme ultraviolet lithography machines—a piece of advanced semiconductor manufacturing equipment that China cannot access due to U.S.-led export controls. However, the status of the photonic chip production line remains unclear.

A February 2023 Politburo study session provides more evidence, albeit highly circumstantial, that Beijing may be taking an interest in photonics as a means to circumvent Western technology controls. Xi Jinping chaired the study session and focused on “basic research for self-reliance in science and technology”—a phrase associated with efforts to reduce China’s vulnerability to foreign technology controls. The speaker at the study session was the president of Beijing University, Gong Qihuang. Gong is a physicist by training and specializes in optics and has several publications related to silicon photonics and AI. Unfortunately, it appears no readout of the session was made public.

The following month, Gong’s Beijing University colleague, Yao Yang, the dean of the National School of Development, spoke at the Bo’ao Forum where it was reported he stated that U.S. semiconductor restrictions amount to the United States “shooting themselves in the foot,” because photonic chips will eventually make electronic chips obsolete. In an interview in September, Yao echoed Chen’s comments, likewise arguing that photonic chips represent an opportunity for China to “change lanes to overtake.” Most recently, Yao published an article in which he pointed out that China’s leading position in photonic chips positions it to “break through first” in this emerging technology. 

Photonics also receives mention in the 14th Five-Year Plan—China’s most authoritative economic policy document. This iteration of the plan, which is notable for its emphasis placed on achieving self-sufficiency in core technologies and reducing reliance on foreign technologies and imported resources, calls for the establishment of “a number of national laboratories with a focus on quantum information, photonics and micro and nano electronics, network communications, artificial intelligence . . . and other major innovation fields,” (italics added).

From Electrons to Photons?

Unlike what Yao Yang suggested, it does not seem likely that photonic chips are going to replace electronic chips—at least not anytime soon. Instead, the relationship between photonics and electronics is better understood as symbiotic. However, some recent developments do seem to indicate that silicon photonics could present China at least a partial path to the leading edge of semiconductor manufacturing that does not require the most advanced semiconductor manufacturing equipment (SME)—a weakness in China’s domestic manufacturing ecosystem that has been leveraged by the United States with recent export controls.

The most immediate application of silicon photonics is in the form of optical interconnects. That is, replacing the copper wiring in circuits with photonics to speed the transmission of information between processors and/or memory, reducing the input/output bottlenecks currently plaguing AI computing. Where data transfer rather than data processing is the bottleneck, the integration of optical interconnects could increase performance of a computing system beyond that of a system with more advanced electronics, but without optical interconnects.

This appears to be what Lightelligence, a U.S.-based optical computing company that has received Chinese funding, achieves, or claims to achieve, with its recently unveiled AI accelerator: Hummingbird. Hummingbird uses optical interconnects to connect electronic chips fabricated by the Taiwan Semiconductor Manufacturing Company at a 28-nanometer (nm) process—far from the current leading edge and well within China’s domestic semiconductor manufacturing capabilities. In doing so, the firm claims (but does not publish) latency and efficiency metrics surpassing those of competitors in certain AI tasks.

Another application of silicon photonics is in the more nascent field of optical computing. In optical computing, photonic processors perform computations using light rather than electrons. Although the types of computations these photonic processors are capable of performing are currently limited, optical computing is showing particular promise for conducting matrix multiplication operations. This type of computation happens to account for over 90 percent of inference operations in the neural networks which form the foundation of large language models, and generative AI more generally, currently driving the most spectacular advances in AI as epitomized by OpenAI’s ChatGPT.

In 2021, before the release of Hummingbird, Lightelligence also unveiled an optical computing system called the Photonic Arithmetic Computing Engine, or PACE. PACE co-packages photonic and electronic integrated chips to achieve advertised processing speeds of 25–100 times faster than NVIDIA’s industry-leading GPU in certain compute-intensive applications.

Previewing his firm’s technology at the Emtech China 2020 Global Emerging Technology Summit in Suzhou, Lightelligence CEO Yichen Shen detailed an advantage of his firm’s use of photonic chips (author’s translation):

“Photonic chips represent a technology direction that can be industrialized faster. One of the main reasons is that photonic chips are not particularly dependent on manufacturing processes and the technology is not restricted. Therefore, we can use 28 nm electronic chips to produce the effect of 7 nm electronic chips faster. This is why Xizhi Technology [Lightelligence’s Chinese name] has recently received the attention of the Chinese Academy of Science and is implementing some very important projects.”

Based in part on Shen’s statements it seems conceivable that breakthroughs in optical computing could allow for the construction of a computing system that uses photonic processors supported by legacy, or at least not the most advanced, electronics that performs certain AI-related tasks as well or better than an all-electronic system using the most advanced electronics. While Lightelligence does not list the process nodes at which PACE’s electronics are fabricated, Lightmatter—Lightelligence’s U.S.-based and funded competitor—is reportedly using 12 nm electronics to support its photonics, which are fabricated at a 90 nm process, and claims to be achieving superior computing performance in certain AI-related tasks compared to Nvidia’s A100, which uses a 7 nm process.

More recently, it was reported that researchers at Tsinghua University have developed a photonic integrated chip that is claimed to achieve performance speeds 3,000 times faster and 4 million times more energy efficient than a “top-of-the-line graphics processing unit” in certain AI tasks related to computer vision. The chip, referred to as the All-Analogue Chip Combining Electronics and Light, or “ACCEL,” was also reportedly manufactured by SMIC—China’s leading semiconductor manufacturer—using a 180 nm CMOS process, which is a decades-old fabrication process. While ACCEL’s current capabilities appear to be narrowly focused and its commercialization timeline unclear, the researchers did posit that future iterations may have broader applications including in large language models.

Conclusion

To the extent that silicon photonics underpins and enables advances in optical interconnects and optical computing, this emerging technology could conceivably alter the counters of U.S.-China competition over semiconductors and AI. The recent series of U.S.-led export controls have sought to cut China off from the most advanced SME needed to manufacture logic chips below the 16–14 nm process, while also preventing the China’s import of the most advanced chips needed to train and operate the most advanced AI algorithms. However, recent advances in silicon photonics seem to suggest China could indigenously manufacture, even without the most advanced SME, computing systems that perform better in certain important AI tasks than their fully electronic counterparts.

That said, despite their advertised performance, the current capabilities of photonic processors should not be overstated, as their capabilities remain narrowly focused. This narrow applicability contrasts with the general-purpose nature of their electronic counterparts. In addition, numerous technical barriers to the widespread adoption of silicon photonics still exist, and an optical computer would also require software development in operating systems and applications to optimize their capabilities. Taken together, the reality of optical computing is possibly years, if not decades away. With the current rate of progress in AI—the size of large language models are doubling every 3.5 months—any delay could be consequential. Likewise, leading semiconductor firms in the United States, as well as those in allied and partner nations, are beginning to commit more resources to silicon photonics, making Chinese leadership in the field far from certain.

Nevertheless, the case of silicon photonics is an important reminder that the United States should not assume that because China cannot access the most advanced SME it will have a permanent ceiling on its domestic semiconductor manufacturing capabilities. New technologies and architectures have the potential to redefine what constitutes a leading-edge chip, potentially eroding the impact of the current controls or reshaping competition in ways that are difficult to forecast. While the U.S.-led export controls are likely setting back China’s capabilities in the manufacture of traditional chips, cutting off certain routes to the leading edge, the export controls could also inadvertently incentivize China to devote more resources to emerging technologies that will play an important role in next generation semiconductors, especially as the physical limits of Moore’s Law are approached and advances in AI are increasing demand for compute.

Matthew Reynolds is a fellow with the Asia Program at the Center for Strategic and International Studies in Washington, D.C.