GlobalFoundries, Tower Semiconductor, and possibly some companies in China are poised for a wave of silicon photonics demand driven by AI and other applications like quantum computing.

The emerging tech may give second-tier foundries a competitive advantage and alter the chip competition between the U.S. and China. In Beijing, some believe silicon photonics can help the nation skirt a U.S.-led ban on exports of EUV tools that have blunted China’s progress making advanced chips, according to a report by the Center for Strategic and International Studies.

The silicon photonics market is small, but forecasts show a 42% compound annual growth reaching $850 million by 2029, according to market research firm Yole Group. The main growth driver is demand for the latest 800G transceivers and increased fiber-optic network capacity to move huge amounts of data.

China’s state-funded JFS Laboratory has announced a “milestone” in the development of silicon photonics, which could help the nation overcome technical barriers in chip design and strengthen domestic supply chains amid U.S. sanctions, the South China Morning Post reported in October.

Foundries producing silicon photonics do not need the most advanced process technology to make chips that nevertheless can perform high-speed data transfers on low power. The new tech generates light instead of electrons to accelerate data transmission and cut energy consumption in AI data centers.

AI’s machine-learning datasets are becoming so large that light transmission has become essential for data transfer in servers, according to Yole.

Starting in March 2023, hyperscale data center clients like Google and Amazon, together with top AI chipmaker Nvidia, began to ramp up demand for 800G transceivers, Martin Vallo, a photonics analyst at Yole, told EE Times in an interview.

“This surge led to a continuous increase in orders and replenishments, fundamentally shifting industry trends,” he said.

These 800G transceivers, doubling the previous 400G generation, deliver ultra-fast, high-bandwidth communication in AI infrastructure and large AI/ML clusters.

GlobalFoundries

As the need for higher data rates and lower power becomes key to the growth of data centers, silicon photonics is emerging in the lead, GlobalFoundries senior director Vikas Gupta told EE Times in an A lot of the photonics parts for data centers today have standardized plugs insertable into a switch or a GPU board. That pluggable form is shifting to silicon-based co-packaged optics, where the conversion interface is closer to compute.

“Because these interfaces are sitting closer to the electronics, you can potentially get lower power consumption because now [system designers] don’t have to account for all of the signal integrity losses that are occurring in the copper traces, even on the board level,” Gupta said. “Silicon photonics has the capability and the opportunity to reduce the losses in data centers significantly while maintaining or increasing the data rates that are required by the next generation of data centers.”

The typical metric for energy efficiency is picojoule per bit.

“Co-packaged optics using silicon photonics has [the] ability to go below five or even one picojoule per bit,” Gupta said. “Our customers have already demonstrated five picojoules per bit on silicon. We do expect that we will be getting down much below five picojoules with the next generation.”

GlobalFoundries has several customer tape-outs that are close to completing the prototyping phase, Gupta added.

“We’re at the point where at least from the transceiver side, datacom side, there are several customers that are ready to go into production,” he said.

GlobalFoundries has quantum computing customers including PsiQuantum and Diraq who are buying photonics. PsiQuantum has made photonic chips on a standard 45-nm silicon-nitride photonics process at GlobalFoundries’ Malta, N.Y., fab.

Even so, the quantum computing company had to install its own tools in the GlobalFoundries fab to start production.

“Silicon as a material runs out of steam for things like modulators, and we expect that there will be other materials like thin-film lithium niobate or barium titanate or the need, for example, of attaching indium phosphide to a silicon photonics die for either lasers or as a semiconductor optical amplifier,” Gupta said. “Once you start talking about these disparate materials, which are non-silicon based, there will clearly be some retooling requirements that would be specific to silicon photonics. Once we start talking about heterogeneously attaching silicon to different materials for the silicon photonics, that would require specific tooling. Whether that specifically occurs within the fab or post fab, that is dependent on the technology. At this point, there’s a significant use of tooling that occurs within the fab for silicon photonics.”

Assembly and test companies are one of the bottlenecks in the supply chain.

“OSATs [outsourced assembly and test] are realizing that this is a significantly growing market, and they need to get into the market right at the at the bottom floor,” Gupta said. GlobalFoundries partnered with Fabrinet in silicon photonics packaging in 2022.

Global R&D organization imec says it has solved many of the problems involving packaging.

OpenLight

OpenLight, a company spun out of Juniper Networks and Synopsys in 2022, designs silicon photonics chips and licenses its related royalties and IP.

“One good thing about silicon photonics is that you don’t have to go down process nodes,” OpenLight CEO Adam Carter told EE Times in an interview. “We’re on a BiCMOS process at Tower Semiconductor at 45 nanometers and we don’t need to chase it. We don’t need to go any further than that.”

The startup founded in 2022 has six customers that have ported their designs into Tower using Openlight’s process design kit, according to Carter. The company has a total of about 15 customers after starting with two when Carter joined in January 2023 as CEO, he said.

The company is building an ecosystem of silicon photonics manufacturing partners that includes design software providers like Synopsis and packaging companies like Jabil.

“It’s our process,” Carter said. “We monitor the bond yields and things like that and develop different recipes depending on what process technology is being bonded onto.”

OpenLight has three photonics processes, two of which have been set up at Tower.

OpenLight offered its first standard process when the company started up two years ago. The next two processes are for automotive sensing and data centers. The company will soon offer a distributed feedback laser (DFB) for the data center process.

“The DFB laser is going to be really what will drive the AI intake inside the data center space,” Carter said.

TSMC and Intel

The world’s top foundry, TSMC, is also making photonics chips.

The company’s roadmap is for COUPE in pluggables in 2024, followed by COUPE on substrate in a chip on wafer on substrate (CoWoS) co-packaged optics solution in 2026, TSMC told EE Times in an email exchange. The COUPE acronym stands for compact universal photonic engine, which is TSMC’s name for the tech it uses to stack silicon and photonics.

The company said it is exploring COUPE on CoWoS interposer.

In June, Intel’s Integrated Photonics Solutions Group demonstrated what the company called the industry’s most advanced and first-ever fully integrated optical compute interconnect chiplet co-packaged with an Intel CPU and running live data.

Intel said its photonics chips were packaged in pluggable transceiver modules, deployed in large data center networks at major hyperscale cloud service providers for 100-, 200- and 400G applications. The chipmaker is developing 200G lane photonics chips to support emerging 800G and 1.6T applications.