Monolithic tunable lasers can adapt statically and dynamically.
Key Takeaways
The move to co-packaged optics (CPO) holds the promise of putting photonic ICs (PICs) and electronic ICs (EICs) into the same package.
Lasers typically don’t join the party in the package, bringing their light in through the faceplate of a server rack. But a new approach could move the lasers into the system, although still not into the package.
By building laser arrays monolithically, companies can simplify the alignment challenges associated with separate lasers and the optical components. Hundreds of lasers are envisioned. Making those lasers software-programmable allows tuning for any random combination of wavelengths. By eliminating many other components, this approach increases system reliability.
Lightmatter recently announced such a light engine that shares technology with its earlier Passage product, which acts as a photonic interposer. Tools from Cadence and Synopsys enabled the photonic chip design, and those companies are also supporting Passage with interface IP for these products.
Eliminating long copper lines in a server
Fibers carrying signals from server to server traditionally terminate at the faceplate of the server. The pluggable module — which includes the laser — contains all the necessary elements for converting the optical signals into electrical ones. Copper then carries signals to the processor or other chips in the package. The signals remain serial, being converted back to parallel when they reach their destinations.
Fig. 1: In a conventional configuration, the electrical portion contains drivers and amplifiers for the optical signal and a SerDes to convert to and from the typical parallel configuration used in an ASIC or SoC. The ASIC and SerDes may be on the same chip (as indicated by the dashed box), in which case the interconnect would be on-chip rather than using UCIe. That UCIe interface is only needed for a standalone SerDes die. Source: Bryon Moyer/Semiconductor Engineering
The challenge with this arrangement is that the long copper lines, typically driven by SerDes, consume a lot of energy. The amount of energy can be reduced by moving the optical elements closer to where the signals are used — farther into the board. Then, a fiber can carry the photonic signal from the faceplate to the point of use.
“CPO moves components closer together,” said Calvin Cheung, vice president of engineering and business development at ASE Group. “There’s a limit to the bandwidth that pluggables offer. So to be able to continue the roadmap and keep performance, you have to migrate to something like CPO to bring the components closer together.”
The laser stands alone
Notably, the laser typically remains in the pluggable module (known as an external laser small form-factor pluggable, or ELSFP), due to reliability concerns.
“In the current lingo of the industry, you have EMLs [electro-absorption-modulated lasers] in a lot of these pluggables,” said Nick Harris, CEO of Lightmatter. “These EMLs are directly modulated lasers that reside inside the transceiver.”
The laser module is one of the least reliable components in such a system. For that reason, it remains in a convenient pluggable form, enabling easy and inexpensive replacement. If designers were to move it into the board, a failure might necessitate replacing the entire expensive board.
“In a lot of cases, with CPO, the laser and the modulator are separated,” continued Harris. “The modulator is located inside the CPO (or the NPO [near-package optics]), and the laser is now a separate thing.”
This change is likely to have an impact on silicon foundries, as well. “Conventionally, the photonic dies are manufactured in a set of foundries, and then electrical dies are manufactured by standard CMOS processes by several companies,” said Priyank Shukla, director of product management for interface IP at Synopsys. “Then, a packaging house packages it together.”
Bringing electronic and photonic dies together in a single package motivates a more integrated process to build both electronic and photonic technologies in a single foundry. “TSMC is taking the lead, offering the whole platform,” Shukla said. “It will manufacture photonic engines and electrical ICs.”
Fig. 2: CPO moves most of the optical components and electronic chiplets in the same package. Notably, the laser remains in the faceplate for reliability reasons, making it easier to replace. Source: Bryon Moyer/Semiconductor Engineering
“Some of the latest generation of pluggable optical modules could be considered as CPO,” said Suresh Jayaraman, senior director, package development at Amkor. “There are other implementations where the optical transceiver and photonic IC, including a micro-lens array for fiber-attach, are integrated in the same package.”
WDM means multiple lasers
Multiplexing colors in a single fiber can be an effective way to raise bandwidth without adding fibers. Multiple signals, each modulated on a different wavelength, can run through the same fiber in what’s called wavelength-division multiplexing (WDM). The downside is that each of the colors being multiplexed requires its own laser.
If those lasers are independent, then they must be carefully managed to ensure that one laser doesn’t drift more than the others. If spaced too closely, two colors could overlap, corrupting the signals intended for each color. The easiest way to deal with this is to minimize the number of colors so they can be spaced far apart, with guardbands allowing for drift.
But the wavelength spacing between colors ends up being wasted because it can’t be used effectively. It’s utilized only if a laser happens to drift into that region, and that doesn’t raise bandwidth. Additional colors would be possible if all lasers could be made to move together when affected (by environmental factors such as heat), or if each laser could be independently locked to a stabilized wavelength.
Digging into laser reliability
The notion that lasers are inherently unreliable may place the blame on the wrong component. The laser assembly may be the culprit. “The laser diode isn’t the thing that fails,” said Harris. The assembly consists of a laser followed by other optical components, including lenses and multiplexers.
These components aren’t built monolithically. They’re glued to the surface that the laser is attached to.
The assembly includes a lens-isolator-lens chip, which effectively multiplexes the different colors. Epoxy, which is subject to contamination and outgassing, can affect reliability. “Sometimes it actually outgasses onto the lenses and blocks the fiber. That assemblage of components has very low reliability.”
Looked at another way, eliminating those added pieces could restore laser-module reliability.
Fig. 3: Typical setup today. The laser light travels through air upon exiting the laser, so lenses and other components are required. Source: Lightmatter
An additional challenge with conventional lasers lies in manufacturing. Each laser and its follow-on components must be carefully aligned to minimize any light loss or distortion. If multiple colors are desired, then multiple laser assemblies are necessary, and they must be aligned with one another. This can result in yield issues, and when done manually, it is error-prone.
Laser bars are one way to provide multiple colors with less of an alignment challenge, but the lasers in the bar are not independent. This runs into a tuning problem. “That’s because the way you tune them is you heat the substrate, and all the lasers tune at the same time,” said Harris. “These lasers are so hot they heat each other up and then they drop in efficiency, so the brightness drops a bunch.”
Multiple lasers with no other components
Lightmatter has announced one possible solution to this challenge with its Guide laser module. This approach contains both photonic and electronic elements that configure and stabilize the lasers while functioning as a light source.
To fabricate the unit, the company builds a monolithic array of identical lasers — potentially a hundred or more — where each has a local laser-cavity heater and a feedback loop to stabilize the color. Multiplexers can then direct the laser light into the appropriate fibers to send into the rest of the system. Existing plugs for these packages can support many fibers.
Building these assemblies monolithically eliminates the manual alignment. They’re constructed in a way that ensures alignment. It’s analogous to what happened with inertial measurement units (IMUs) when accelerometers only measured acceleration in one dimension. Getting two dimensions required aligning two such sensors at precisely 90° from each other. The advent of bidirectional monolithic sensors eliminated that alignment need.
Harris said that Guide has benefited from the company’s experience with Passage. “In Passage, we have chips with 1,000 micro-ring modulators,” said Harris. “Micro-rings are cavities, and they’re extremely small and sensitive, and we’ve learned how to control them to an incredible degree. We’re leveraging that exact same kind of technology with Guide.”
A software-defined laser
Color assignments are tied to the boot-up process. Software can write color values to each laser, which form part of the firmware used when booting.
“The chip has its own microcontrollers,” explained Harris. “You boot it up and say, ‘I want this laser to be this color, or I want that laser to be that color.’ You give it the instructions, and then the chip will communicate with the laser diodes and set up the colors.”
Fig. 4: Software assigns the colors at bootup so that any laser can be any color. Source: Lightmatter
Also, laser color can be reassigned during operation. That allows for redundancy. If a laser fails during manufacturing test or operation, another one can be programmed to the same color, with the multiplexer now guiding the new laser into the fiber.
“If there are any failures at time zero, we can swap in another laser,” said Harris. “If something fails at burn-in, we can swap in another laser. If something fails while it’s training, we can swap in another laser.”
Fig. 5: If one laser fails, whether early or late in life, another can be swapped in as long as a spare is available. Source: Lightmatter
Unlike the ELSFPs, the laser light in Guide exits into a semiconductor waveguide rather than air. It therefore requires no lenses to move the light forward since the waveguide itself provides the necessary confinement. That eliminates the unreliable components, making it inherently more reliable than a traditional ELSFP. That means one can mount it directly to the board rather than the faceplate. A single chip can provide many light sources for use in multiple places on the board. The light moves out of the system via detachable fiber connectors.
Fig. 6: Instead of the laser residing in a pluggable, it can move onto the board closer to where it’s needed. Source: Bryon Moyer/Semiconductor Engineering
Here, the laser is still not co-packaged with other chiplets. That scenario remains unlikely given the concentrated heat inside a package. That heat can play havoc with a laser, so keeping it away from hot components is preferable. But there are still plenty of places it could go on the board without moving all the way to the faceplate.
The independent heating of each laser addresses the challenges presented by independently drifting lasers and light bars. And it keeps the many lasers from coupling to each other through heat.
Passage gets UCIe and SerDes
Although Guide isn’t directly tied to Lightmatter’s Passage product, it is expected to complement it. Passage also contains both electronic and photonic components, and chiplets will need to talk to each other. A standard interposer would simply connect two chiplets electrically using a protocol such as UCIe or BoW, but Passage allows optical interconnections.
In order to allow chiplet makers to remain agnostic as to what lies across the chiplet interconnect, Passage will integrate UCIe and 224-Gbps SerDes from Cadence and Synopsys. Any conversion from UCIe-based signals to optical would occur on Passage, relieving chiplet makers of the need to provide an optical interface.
Although the CPO concept is motivated by removing the power-hungry SerDes arrangement shown in Figure 1, it’s really about getting rid of the long copper lines running to the faceplate. The SerDes driving such lines must be high-power ones to deliver the signals with high integrity.
“If this SerDes drives a long channel — let’s say 40dB of channel loss or 19 inches of backplane — this is a long reach SerDes, which burns more power,” said Synopsys’ Shukla.
With CPO, a SerDes is still necessary to adapt the inherently serial optical signals to the parallel buses typically found on-chip. It’s just that it doesn’t require the high power of the older ones.
“You need to have a SerDes because an accelerator has parallel data,” continued Shukla. “This SerDes is a very short channel SerDes. It’s just driving the co-packaged channel. So it will have a 3 dB – 5 dB loss. This is extremely efficient.”
As a first project, they’ve integrated a simple electrical die as proof of concept. “The present generation will demonstrate that electrical die can drive the optics reliably. And as this ecosystem evolves, you will have a real accelerator integrated in this package,” Shukla added.
Collaboration between EDA firms and system providers is likely to continue in future designs. “As AI capacity continues to expand dramatically to accommodate unprecedented demand and workloads, scale-up and scale-out are transforming AI infrastructure,” said Boyd Phelps, senior vice president and general manager of the silicon solutions group at Cadence. “Our collaboration with Lightmatter demonstrates our commitment to the evolution of advanced interconnects. By integrating our high-speed SerDes and UCIe IP into this new CPO platform, we’re helping our customers build more scalable, power-efficient AI systems.”
Keeping it under wraps
Curiosity motivates a desire to understand exactly how this all happens. Lightmatter claims that it’s not easy and that the team benefited from what it learned from Passage. But that’s as far as it goes. “We will never explain how we do it,” said Harris.
But it also issues a challenge to the rest of the photonics industry. If Lightmatter is successful with this CPO configuration, others will want to try, too. The company characterizes its approach as a transition to high-volume automated manufacturing as compared to the manual efforts necessary today, similar to how integrated circuits moved from small-scale ICs (anyone remember SSI?) through MSI, LSI, and VLSI. Indeed, the company refers to its approach as very-large-scale photonics, or VLSP.
CPO is challenging, and it’s been slow to move into high volumes. Laser engines such as Guide should make it easier to increase the scale of photonics in systems that require it.
—Laura Peters contributed to this report.
Related Articles
Co-Packaged Optics Reaches Power Efficiency Tipping Point
But blazing fast data speeds come with significant manufacturing challenges.
Linear Pluggable Optics Save Energy In Data Centers
New OIF electrical standards will enable interoperability, adding another option for faster and more efficient data movement.
Chiplet Fundamentals For Engineers: eBook
A 65-page in-depth research report on the next phase of device scaling.
 |
 |
 |
 |
 |
 |
 |  |
 |  |
 |
 |
Leave a Reply