Semiconductor Industry Closes in on 400 Gb/s Photonics Milestone

The optical links that help connect the computers densely packed inside data centers may soon experience pivotal upgrades. At least two companies, Imec and NLM Photonics, now say they have either achieved 400-gigabit-per-second per lane data rates, the next key goal for data centers, or have these speeds within sight. Moreover, instead of relying on exotic new technologies, the devices of both teams are based on silicon.

Nowadays, racks of servers within data centers communicate with each other over tens or hundreds of meters using optical transceivers that encode electronic bits onto beams of light and decode them on the other end. The transceivers sit at the ends of cables that each pack in 8 optical fibers. Right now, the typical data rate for such transceivers is 100 Gb/s per lane, with the industry rapidly ramping up to 200 Gb/s, says Cedric Bruynsteen, a researcher at IDLab, an Imec research group at Ghent University in Belgium.

However, “the explosive growth of AI training clusters and other compute-intensive applications is driving an urgent need for greater bandwidth, higher performance, and improved efficiency,” Bruynsteen says. As such, 400-Gb/s-per-lane transceivers “will represent a new milestone.”

Researchers are exploring a range of technologies to meet this need. For instance, semiconductor manufacturing giant Taiwan Semiconductor Manufacturing Company (TSMC) is working with Sunnyvale, Calif.-based Avicena to produce microLED-based interconnects. “Any of these technologies could end up winning,” says Clint Schow, professor of electrical and computer engineering at the University of California, Santa Barbera. “It’s the Wild West right now.”

Silicon photonics was generally not thought capable of scaling to 400 Gb/s per lane due to energy efficiency and other factors, says Lewis Johnson, chief technology officer and co-founder of NLM Photonics in Seattle. As such, researchers are exploring other platforms, such as indium phosphide (InP), barium titanate (BTO), and thin-film lithium niobate (TFLN), he notes.

However, these new platforms have drawbacks of their own when it comes to optical interconnects. For instance, Bruynsteen says that InP is fundamentally limited by smaller wafer sizes and higher manufacturing costs. Johnson adds that BTO and TFLN both require expensive manufacturing modifications.

Now Imec and NLM both show it might be too early to dismiss silicon. “Silicon still has plenty of headroom, even for the most demanding high-speed applications,” Bruynsteen says.

Imec’s new wafer delivers 448 Gb/s per lane, which the company says is a first for any silicon-based electro-absorption modulator.Imec

Imec’s New Modulator

Researchers at Imec have developed a silicon germanium electro-absorption modulator. When a voltage is applied, the semiconductor absorbs more light, letting the device control the intensity of optical signals passing through it.

The new device can deliver a 448 Gb/s per lane data rate, a first with any silicon-based electro-absorption modulator, Imec notes. “Electro-absorption modulators have always been an intriguing component because they uniquely combine low power consumption, compact footprint, and high-speed operation,” Bruynsteen says. “Bringing all of these benefits together in one device can be described as the holy grail of modulator design.”

It’s possible 448 Gb/s is not the technology’s top speed. “We’ve now reached the point where the test equipment is the limiting factor,” Bruynsteen says. Higher-frequency measurement tools could help explore how fast the new device’s data rates could get, he notes.

The new device performs best in the conventional (C) band—infrared rays about 1550 nanometers in wavelength that are typically used for long-distance optical communications. However, most data center links today operate in the O-band, centered at about 1310 nanometers, Schow notes. This is because the O-band experiences less chromatic dispersion, in which different wavelengths travel at different speeds in a material, causing light pulses to spread out and become distorted. Still, chromatic dispersion should not be a problem for this application, given the relatively short distances involved, he adds.

The new roughly 300-square-micron device allows Imec “to leverage the scalability and cost benefits of standard CMOS manufacturing,” Bruynsteen says. This “may be one of Imec’s strongest points here,” says Schow, who did not take part in research at either Imec or NLM. The Belgium-based research organization detailed its work at the European Conference on Optical Communication in Copenhagen in September.

Currently Imec is sharing their new device with partners to explore its potential in AI training clusters and other high-performance environments, Bruynsteen says. “Our next goal is to validate the device under realistic data center conditions, such as elevated operating temperatures and a wide range of optical power levels, to ensure stable and reliable performance,” he says.

Hybrid Silicon-organic Photonics

In contrast to Imec, NLM Photonics employs silicon-organic hybrid photonics. Each of their new chips possesses eight Mach-Zehnder modulators, which split light entering them down two separate arms. The chip can electrically alter the optical properties of one of these paths, shifting its phase. When these beams recombine, any phase shift can alter the intensity of the resulting light. The silicon-organic hybrid material NLM uses requires less voltage to alter its optical properties than regular silicon.

Third-party tests found NLM’s chip is capable of 224 Gb/s data rates per channel. The company now aims “to demonstrate a 400 Gb/s per channel link with our partners, showcasing practical performance scaling,” Johnson says. NLM detailed its findings at the Photonic-Enabled Cloud Computing Industry Summit in October.

NLM claims its eight-channel chip delivers 10 to 15 times more efficient operations than conventional silicon photonic modulators, thanks to exceptionally low operating voltages: The new device operates at a drive voltage of 1 V or less, whereas comparable silicon photonic modulators operate at drive voltages of 2.5 to 3.5 V. The NLM chip is also smaller than those made with competing technologies—17 square millimeters versus 25 to 50 square millimeters.

Although NLM’s device uses organic materials uncommon in photonics manufacturing today, it does so late enough in the manufacturing process that it does not lead to costly modifications.

“Our near-term focus is working toward manufacturability at scale,” Johnson says. “We’re developing automated processes for… work that’s essential for integrating our organic electro-optic materials into existing foundry workflows without disrupting established production lines.”

Schow notes the most likely criticism of NLM’s work will center on the question of how stable the organic materials will prove over time, “but polymers have gotten quite a bit better over the years.”

Scott Hammond, the director of process development at NLM Photonics, holds the company’s newest wafer.NLM Photonics

Johnson says that NLM has “documented excellent materials-level stability results.” It’s shown both long-term thermal stability in excess of 120 ℃ and encapsulation technology capable of withstanding the 85 ℃, humidity, and heat testing requirement for telecom hardware. NLM is also developing next-generation materials in-house to offer enhanced thermal stability for more demanding processing conditions, as well as in applications beyond data communication, such as quantum computing, he says.

Of this pair of advances, “Imec’s is probably more prudent,” Schow says. “It’s a device available now at 300 mm, without any question of manufacturing issues.” That said, “polymers are a contender for a ‘beyond-silicon’ innovation. As we develop faster and faster links, it’s never clear what the winner will be until it is, especially with next-generation materials.”