The emergence of artificial intelligence (AI), machine learning, and large language models has made for turbulent times in data networking. The computational demands these technologies are placing on data networks are, by some reports, doubling every three to four months while the computational infrastructure is doubling in power every 18 months. Stakeholders along the entire data center value chain are scrambling to develop new solutions that will help improve latency, power efficiency, component density, and/or cost. Many are looking to integrated photonics for solutions. But integrated photonics itself is still evolving. Its inherent diversity and difficult path to scalability often raises as many questions as answers. (left) Nathan Tracy, President, Optical Internetworking Forum (OIF). TE Connectivity, Technologist. (right) Jeff Hutchins, Physical and Link Layer Vice Chair for Energy Efficient Interfaces, Optical Internetworking Forum. Courtesy of OIF. If one thing is certain, data center architectures will change on a fundamental level. New optical solutions will need to edge closer to core silicon components for AI to achieve its potential in cloud applications. But these solutions cannot be merely innovative, they must also be scalable and, more importantly, interoperable with other electronic and photonic network components. The thin path between breakout innovation and universal interoperability is traced by industry standardization bodies, such as the Optical Internetworking Forum (OIF). To better understand how standardization bodies prioritize problems, forge consensus among competitors, and choose the timing and scope of new standards, Photonics Media editors spoke to Nathan Tracy, OIF President, and Jeff Hutchins, OIF Physical and Link Layer Vice Chair for Energy Efficient Interfaces. Their conversation, edited for space and content, follows. PM: As new and bigger AI models emerge, the demand for computational power in the data center is quickly outstripping the power that traditional electronic solutions can provide. This demand for more intelligence, bandwidth, and interconnectivity has put a premium on breakneck innovation in photonics, which many see as the solution to improve latency, power, and other challenges. Trying to apply standards to this fast-changing landscape seems more like triage than surgery. How does an organization like OIF shape new standards for often unfamiliar new technologies in this environment? OIF: OIF is a member-driven, consensus-building organization. The membership represents a broad cross section of the industry including optical and electrical component suppliers, equipment innovators, test and measurement providers, and network operators — all serving multiple market applications. It is actually this diversity that enables OIF to identify and understand market directions and evolving requirements. With this as a starting point, for example, OIF realized that new energy-efficient architectures were becoming important in 2019. The membership developed a co-packaging framework paper that identified a number of different applications (including AI) that would benefit from interoperable interfaces that leverage co-packaged copper and optics modules. Our first Interoperability Agreement project was to use co-packaging for the 51.2T generation Ethernet switch as a starting point. Of the various options to choose from, the 51.2T Ethernet switch appeared to be the most challenging to get started with — that is, it’s a ‘super-set’ solution. We knew the industry wasn’t quite ready for co-packaging at this scale. But the effort helped the OIF and the industry learn what was needed to develop new solutions. As a result of this effort, a number of new products were developed such as low-power short-reach electrical interfaces, finer-pitch electrical connectors, denser optical connectivity, and external laser sources, which has since become the OIF’s ELSFP IA [External Laser Small Form-Factor Pluggable Implementation Agreement] project. PM: Can you give an example of how some of these new solutions addressed the need for more energy-efficient architectures? OIF: Sure, the Interoperability Agreement defined a form factor for an optical module that could be placed adjacent to the switching silicon instead of being at the face plate of the line card. What that did is reduce the length of the electrical channel from the silicon SerDes [serializer/deserializer] to the optical module SerDes, allowing it to need less equalization and transmit power. Putting optics next to an ASIC [application-specific integrated circuit] like this is really new for the industry. It hadn’t really been done before in this way, and it wasn’t like you could just go out and buy the things you needed to do this. So things had to be invented and created, and part of that was the sockets and the optical or copper module definition and the external laser small form factor. People had to work on all these new things that would allow this co-packaging ecosystem to be feasible. PM: As we noted earlier, the demand for bandwidth and interconnectivity as well as lower-energy solutions is quickly growing. How does OIF’s membership stay ahead of the curve? OIF: The co-packaging project was able to move very quickly as a result of highly focused, specific input from members, and we expect that to continue to be the case moving forward. These new requirements have near-term objectives — for example, about two years out — as well as longer-term horizons. This view helps the industry to plan investments to meet the timing of the different application needs over time. As that project wrapped up, we decided to engage our end-user members and see what would be needed next. As a result, we initiated the Energy Efficient Interfaces Framework [EEI Framework] project to understand the emerging applications for AI/ML [machine learning] and help the OIF choose our next projects. As mentioned, it is a rapidly changing area and we have seen the application requirements evolve over the short time since the EEI Framework project was initiated. PM: The scope and timing of a new standard are important considerations. You want to foster aims like interoperability, but if you issue a standard at too early a technology-readiness level, you might dampen innovation. Propose it too late, and your standard might complicate intellectual property. How do standardization bodies like OIF determine the scope and timing of a new standard to optimize user flexibility without shorting supplier creativity? OIF: Well, if you look at IEEE’s standards for DR4 and SR4 [optical transceivers], they don’t tell suppliers how to make them; they agree on specifications at the electrical and optical interfaces. Suppliers may use a variety of technologies like EA [electro-absorption] modulators, or Mach-Zehnder modulators, yet meet the overall specification at the interfaces. Some things are harder for some suppliers to achieve and easier for others. Likewise, our approach is to develop a high-level specification that allows a lot of different technologies to play. When we start one of these standards, we consider how far down into the system we need to standardize things. That still leaves room for innovation and people to use different technologies, because you never know until you get to a high-volume application what the winning solution is going to be. This is why the OIF efforts for both the Co-Packaging and the Energy Efficient Interfaces began with a framework-level project to study the applications and available technologies and look to prioritize when it makes sense to standardize. This process has served OIF well in the past. The application requirements from the end users also come with timing. This then generates feedback from component and equipment provider members regarding timing/performance trade-offs. Further, the requirements might note that it is acceptable to have a nonstandard solution available for the near-term with the goal to have a longer-term standard in place. This flexible approach allows the industry to innovate and experience cycles of learning. PM: Can OIF give an example of when a standard placed (or might place in the future) a premium on conserving the steady flow of new innovation? OIF: The OIF has done this multiple times over the years. One example is its standard for tunable lasers. The OIF initiated this effort when there were many different approaches for fabricating these devices. The goal for the first agreement was to standardize the management interface, which was very extensible and on an initial form factor. The optical specifications were left very wide open to accommodate all technologies. This approach has served the industry well as the tunable laser implementation agreements are still relevant 21 years after the first published tunable laser agreement. The main changes have been in the size of the laser form factor as technologies have progressed. Another example involves OIF’s project for external laser sources for co-packaging, which its co-packaging framework project identified as one of the important elements for co-packaging. OIF later initiated the ELSFP [External Laser Small Form Factor Pluggable] project, which defines specifications for mechanical, electrical, thermal, and optical connectivity but provides extraordinary flexibility in the choice of laser technologies, a very wide range of supported optical power levels, a variety of fiber configurations, and flexibility in the control interface. Despite the flexibility, it specifies enough common elements for the industry to rally around. It is expected that this concept will support the industry’s needs far into the future. PM: Does the inherent diversity of integrated photonics also make it more challenging to foster the multi-stakeholder participation and collaboration necessary to standards development? OIF: Yes, the diversity of technical approaches is always a challenge when developing standards. But you have to start with your application requirements. What do the end users need? What are their requirements, and what’s the minimum you have to specify to enable the industry to achieve those goals and still get a good degree of interoperability. You can always over-specify things. And this gets back to our first answer about OIF being a member-driven, contribution-based organization. We have technology providers, component integrators, equipment integrators, and end users; and they all have to reach a consensus on an appropriate level of standardization that won’t result in a single-point supply base or allow a solution to be so broad that it doesn’t deliver value. The Co-Packaging Framework generated several projects that are technology independent. For example, the development of the CEI XSR and XSR+ electrical interfaces are independent of technologies used for co-packaging. As mentioned earlier, the ELSFP was designed to be technology independent. And likewise, for the CMIS, the management interface used in co-packaging. However, the 3.2T co-packaged optical engine is a special case as the socket footprint was designed to mate with copper cable assemblies as well as optical engines. This end user requirement set the size of the optical engine module. Where possible, we strive to develop standards that enable most technologies to participate one way or the other. PM: Interoperability offers an obvious common denominator on which to build a standard for discrete components that need to have a common interface. Even so, the sheer diversity underlying integrated photonic solutions poses challenges elsewhere in the value chain, such as manufacturing and packaging. Are process design kits (PDKs) a sufficient de facto solution for governing what PIC designs are implemented downstream? Or do foundries need to take a bigger role in standards for optical interconnect? OIF: While PDKs are a wonderful starting point for designs, designers often will further optimize the elements for their particular application. In the short term, limiting customers to PDK elements may limit innovation. Probably more important for the foundry is to settle on standardized process steps and allow users to innovate their designs while leveraging PDK elements. Courtesy of iStock.com/Just_Super PM: Integrated photonics poses further challenges for testing. Can you talk about how those challenges might evolve as optical interconnect becomes more deeply integrated? What shape will standardized solutions need to take? OIF: Features to aid testing of the electronics can be integrated into digital circuits such as pattern generators and loop back features. Likewise, photonics can integrate a variety of components, such as lasers, splitters, grating couplers, and photodetectors, to aid testing. Perhaps more importantly for the EEI project to discuss is new methodologies to standardize and to test the new Energy Efficient Interfaces. This work is challenging and expected to break new ground. But integration does present challenges to testing. If the interoperable part is a QSFP [quad small form-factor pluggable] transceiver module, then we can come up with a way to test that module, and we can come up with a way to test the switch that’s going to be plugged into it. When that module and switch are shipped to an end user and meet each other for the first time, we know that they’ll work together because we defined a way to test that separable interface and it’s based on the electrical signaling that’s going to come out of the switch and into the module.When those components become part of an integrated optics solution, that optical subassembly is going to be inside of the equipment where it’s not as easily tested or replaceable. The framework methodology that we use identifies multiple solutions, multiple challenges, pros and cons, and trade-offs to this challenge. If I did that on my own, I would have a very different list of pros and cons than if I did that with you sitting beside me because I’m an electrical guy and you’re an optics guy, right? By having all the OIF members in the room together as we create this list of trade-offs, we each discover that what might be a showstopper for one technology is not a showstopper for another technology. We learn ways to overcome these challenges through each technology’s strength. And that’s really what has to happen here because we are trying to do things that are very hard, that aren’t testable, and that may not be manufacturable. But it’s through that process that members can approach these challenges together to find an interoperable solution that meets end users’ needs.