Optical technology is well established for long-haul communications, but the distances it serves are shrinking — especially in the data center.
Vertical-cavity surface-emitting lasers (VCSELs) already drive short fiber links. But efforts are underway to further scale them down to provide more connections through waveguides than fiber can provide.
“We have seen the transition from long haul to metro-to-local area networks and then into the data center,” said Suresh Jayaraman, senior director for package development at Amkor Technology. “Fiber has moved from the edge of the card to on-board, and more recently onto the package.”
Much research is underway to move optics closer to servers in the data center. Today, the technology typically is employed for rack-to-rack communication, while most intra-rack wires are copper. Replacing the copper with fiber remains a development opportunity.
Integrating lasers with silicon
Of particular interest is where the fibers will attach to processors. Standard pluggable formats may give way to linear pluggable optics (LPO) and co-packaged optics (CPO). But those are still in development.
Three laser attributes dominate concerns over how to integrate optical communications — reliability, temperature sensitivity, and energy consumption. Reliability has been a top concern, and although it’s improved, developers are still concerned about soldering a laser onto a board in case it goes bad and must be replaced. This is the central motivator for pluggable formats.
“The good news is that the lasers are getting pretty robust, pretty reliable now compared to what they were in the past,” said Mitch Heins, senior product manager at Synopsys. But designers remain wary.
Lasers are notoriously sensitive to changes in temperature, and that can be challenging when precise wavelengths are necessary. (Note that in electronics we typically define speed in terms of frequency, or Hz, whereas in optics we tend to speak of the wavelength, typically in nm.)
Finally, energy is important. The obvious energy consumer is the laser when operating, but a related aspect of power is the laser threshold. That’s the minimum current necessary to stimulate lasing. That current is the ante for laser power, and the higher the threshold, the more current will be constantly running.
Efforts therefore focus on establishing lasers with high reliability, low temperature sensitivity, and a low threshold.
Incumbent lasers
Today’s optics often rely on so-called distributed feedback lasers (DFBs). These are well-behaved structures with thresholds in the mid-to-low milliamp range — which is lower than some other types of lasers.
This technology represents a scaling down of long-haul optics for shorter distances. The lasers typically operate in one of two bands, C or O. The former is centered at 1,550nm wavelength, while the latter is at 1,310 nm. The O band aligns with an optimal low-loss frequency through the fiber channel. The C-band range doesn’t have quite the same low loss, but it’s acceptable, and it lends itself to erbium-doped fiber amplifiers (ERFAs) that help signals cross under the oceans.
In the data center, and within a rack, loss is important, but the distances don’t require ERFAs, so the C band adds no value (although it’s certainly usable). Whichever band is employed, DFBs provide a clean single-mode laser source that can serve for interconnects.
“DFB lasers are ‘better’ — i.e., with narrower linewidth, larger bandwidth, higher optical output power, and they can operate in O and C bands — but they’re also more expensive,” said Jigesh Patel, principal engineer, photonic systems and circuits at Synopsys. “So they’re typically used for long reach over single-mode fibers. DFB lasers are also better suited for coherent fiber-optic communication (which some say will eventually move inside data centers) and wavelength-division-multiplexed fiber-optic systems (partly because of single-mode operation with narrow linewidth and partly also because of it supporting much larger operational wavelength band).”
Much of the optical research underway drives the development of silicon photonics, where photons do more than just travel from here to there. They perform computing along the way, just as electrons do in electronics. As such, new optical developments that don’t further that technology might be bypassed. And yet, effective interconnect should be possible without meeting the needs of photonics.
Temperature control drives system configurations
Of the challenges developers face with lasers, the big one is managing temperature. “You have to do something either to control the temperature of the device in the rack, which is tough because you’re controlling something on the order of one degree or so,” said Dick Otte, CEO of Promex. “We’ve seen cases in which people are trying to control the temperature of optical devices to as fine as one-tenth of a degree centigrade.”
Absent good control, noise margins are insufficient for signaling formats such as PAM4. “Variations in temperature cause dimensional variations in the laser, which change the frequency enough to get things out of band and reduce the signal-to-noise ratio,” Otte said. “When you start utilizing things like PAM4 and PAM16 as modulation schemes to maximize the amount of data, those schemes start to fail here.”
Although CPO remains a goal, whether to include the laser remains an open question. Most experts lean toward keeping the lasers out of the package that houses other electronics and optics for easier temperature control. Laser proximity to a high-temperature chip, such as a processor, is of particular concern — especially as activity rises and falls, leading to temperature changes.
“Lasers have mostly stayed outside the package due to area and power concerns,” said Jayaraman. “Laser performance and reliability (lifetime) also improve if it is decoupled from the package, even though optical coupling loss and loss of edge bandwidth densities are issues that need to be addressed with remote laser sources.”
Steve Klinger, vice president of product at Lightmatter, agreed. “Lasers are relatively temperature sensitive, so having them physically separated from the rest of the silicon and being able to locate it elsewhere in the system is advantageous.”
It may even make sense to bring lasers together in a single, carefully controlled module. Fiber can still bring the light to the co-packaged circuits, avoiding the current situation where optics are converted to electrical SerDes to drive chips — an energy-hungry approach.
“The lasers come in on some number of the fibers that are at the edge of the package, and they’re separate [from the package] for two reasons,” said Klinger. “One is the serviceability of that laser. The second is just the sheer amount of laser power that is needed to power a very large scale. The number of lasers lends itself toward integrating the laser in a separate physical module, and that module plugs into and out of the system.”
Work continues, however, on other ways to house lasers, which largely rely on III-V semiconductors. The typical co-packaged scheme is to create a cavity in silicon in which to insert the laser, assuming it’s an edge-emitting model like most lasers. By embedding it in a die, the light coming from the edge can enter a waveguide.
Fig. 1: Edge- vs. surface-emitting lasers. The top image shows an edge-emitting model embedded in silicon. The bottom depicts a surface-emitting laser (in this case, backside) atop the silicon with a mirror or grating coupler to send the light into the waveguide. Source: Bryon Moyer/Semiconductor Engineering
Further work continues to grow III-V materials epitaxially atop silicon. If such a laser were edge-emitting, efforts would be necessary to route the emitted light down into the silicon.
“There is a lot of active research around integrating III-V materials onto silicon substrates, such as monolithic integration, heterogeneous integration, and transfer printing,” said Jayaraman. “However, none of these appear to be ready for high volume.”
Heins agreed. “People are working on all sorts of solutions, whether you grow III-V on silicon or whether you cut a hole in the silicon and drop a laser die into it.”
Meanwhile, work continues on quantum-dot (QD) lasers, which have less temperature sensitivity but lower output power. QD lasers are largely found in displays, where their color purity and tunability make them particularly attractive.
VCSELs may have a new application
VCSELs are easier-to-build (i.e., cheaper) lasers that are found in a variety of systems, such as optical mice, as well as shorter data-center connections. “If the intra-rack type short-reach connectivity uses multimode optical fibers, VCSELs are the optimal choice,” said Patel. “They’re easy to manufacture for the wavelengths that multi-mode fibers use and cost less. 850nm is the choice at which VCSEL manufacturing technology is the most mature.”
VCSELs have some issues when considered in the context of silicon photonics, however. Those issues may not matter for the interconnect role, and further developments have improved some of their other characteristics.
To start with, their wavelength isn’t O band or C band. Many are centered at 850nm, with some new ones at 980nm. Efforts are underway to create VCSELs with longer wavelengths, but nothing is in commercial production yet.
“VCSELs that operate in the O band are desired, but difficult to manufacture with good yield,” explained Patel. “Remarkable progress has been made in the last three years, not only to make O‑band VCSELs happen, but also with much wider bandwidth than in the past. And newer modulation formats, such as PAM4 and DP-PAM4, can achieve much better spectral efficiency than the bandwidth of the laser device itself. Both high-bandwidth O‑band VCSELs and PAM4 have re-invigorated interest in using VCSELs in the O band with single-mode fibers, something that used to require DFB lasers before.”
Some VCSELs apparently also exist at 1,550nm, but they’re not mature yet. Even so, as long as the photodiode at the end of a waveguide that converts the optical signal back to electrical can respond efficiently, and as long as there’s no interoperability question with other optical components, the specific wavelength may matter less for interconnect.
Their output is relatively wideband. This is a problem for photonics, which employs components such as resonators that need precise wavelengths. Again, for pure interconnect, this may not matter.
VCSEL power is too low for photonics, although it can be adequate for interconnects. But it’s slow to modulate if using direct modulation as compared to the speeds necessary for photonics. Direct modulation involves running the electrical communication signal through the laser itself to create a modulated optical signal. It contrasts with generating a pure laser beam and then modulating the beam afterwards.
“If you want to directly modulate them, they’re slow, and that’s the big reason why VCSELs have fallen out of favor,” said Tapa Ghosh, founder and CEO of Volantis Semiconductor.
They also exhibit multiple modes and polarizations. These are characteristics that might otherwise help with long-haul communication, where many signals are aggregated and multiplexed into a few long-distance lines, but they are not useful for extremely short point-to-point connections.
VCSELs in the data center typically drive fibers. As indicated by their name, VCSELs emit their light perpendicular to the surface, in contrast to edge-emitting lasers such as DFBs. This can make integration easier, with vertical light from backside-emitting lasers reflected into waveguides using mirrors.
A big win here is power. Whereas DFBs have milliamp thresholds, VCSELs have thresholds in the hundreds of microamps.
Can VCSELs be tamed?
These characteristics are a mixed bag, some good, some unhelpful. Volantis, an optical startup, said it has engineered an improved 980nm VCSEL that’s single-mode and single-polarization. The company is combining it with a wafer that acts as an optical interposer, with grating couplers to admit the laser light into waveguides and photodiodes to convert back to the electrical domain.
Fig. 2: VCSELs used for an optical interposer. Volantis’ first-generation implementation has a 2D array of VCSELs that couple onto a waveguide using a grating coupler (GC). In their second generation, they plan to use transfer-printed VCSELs that can sit under the chips they interconnect. PD = photodiode; SiN = silicon nitride. Source: Volantis
“These devices can be very small,” said Ghosh. “You’re not doing standard bonding here, so you can use backside-emitting devices.”
Among the various issues the company says it has addressed are:
Back-reflections, which have plagued some VCSELs;
Eliminating the need to attach fibers;
Issues with modes and polarization;
Thermal drift;
Infant mortality;
Reliability and yield through redundancy;
Simpler assembly processes, and
Power — the laser bandwidth is 18 GHz, which is lower than the 30 GHz or so bandwidth that VCSELs are capable of, but more energy-efficient.
The temperature tolerance is greatly improved. “A VCSEL [in our system] has about a 1nm shift across its temperature operating range,” said Ghosh. “There are no modulators that care about the specific frequency wavelength coming in. We’re driving a 10 to 20nm broadband channel [the waveguide], so that 1 to 2nm wavelength shift doesn’t affect us.”
Volantis is starting with a 2D array of VCSELs coupled into silicon for its first generation, but the second generation should be transfer printed based on lasers built on a separate wafer. After building those lasers, “you’ll do an etch process to create a very thin lever,” said Ghosh. “And in the transfer printing process, you have an array of little pillars, and they bond temporarily onto the [lasers], break that tiny lever, and then lift the laser off and transfer it to the target wafer.”
Volantis claims efficiency of 139 fJ/bit, which is a big benefit for the data center. This contrasts with DFBs, which communicate in the low picojoules-per-bit range. “If you make single-mode VCSELs with small apertures, you can get high energy efficiency and extremely high temperature tolerance,” said Ghosh. “There are some tricks you need to do to get this, but you get almost all the advantages of quantum dots with VCSELs.”
The form-factor is unusual, with a full wafer that is similar to Cerberus, but custom routed for an application. Typical rack installation would be vertical (on edge) in a 32U box.
The speed limitation is countered by high levels of parallelization — 5,000 to 10,000 signals at 25 Gbps — to make up the difference. This is where waveguides are essential, because a rat’s nest of fiber at that level would be unworkable.
If this proves to be effective, it could boost VCSEL adoption. If the wafer-scale approach is essential to achieving power-efficiency goals, however, then the application space is likely to be limited. But if developers can leverage improved VCSELs in smaller installations, they could add to the options available for heterogeneous integration.
Even if this particular VCSEL doesn’t play out as Volantis hopes, it seems likely that new VCSELs with longer wavelengths are likely to start populating the data center in the future. We just don’t know when.
Related Articles