JOHN WILLIAMSON AI/ML AND PHOTONICS
UNEQUAL ATTRACTION? At this point in time, though, the
pace of transition to higher bandwidths. “New data centres supporting AI/ ML are going to be utilising a lot higher bandwidth as we go forward,” emphasises Carter.” And, the transition between the different speeds – going from 800G to 1.6T and to 3.2T - is going to be a lot faster than we’ve traditionally seen.” Gupta has an arresting metric here. “Generative Large Language Models are expected to grow logarithmically in the number of parameters, adding to the need for more memory and faster, low latency interconnects,” he comments. “For example, Generative Pre-trained Transformer (GPT) 4 will potentially have trillions of parameters as opposed to its predecessor GPT3’s 175 billion.” The responsiveness and co-ordination of the optical component industry may also be due an overhaul. “AI is going to really drive high bandwidth, high density optical interconnects. In the networks that are being deployed to do AI it’s a big challenge right now with the way that the component industry does things today,” argues Carter. “Things will have to change.” Greater degrees of component integration and optics co-packaging is on the PIC front burner to further reduce power consumption and minimise losses from the use of separate packages. “PICs are currently used mostly in traditional optical transceivers and installed in fabric switches such as Ethernet and Infiniband and so on. Most deployed PICs are 4-channel or 8-channel,” points out Mehta. “Progression to higher levels of integration, coupled with manufacturing in the mature silicon foundry and Outsourced Semiconductor Assembly And Test (OSAT) ecosystem is critical to enable optical interconnects to begin to compete with copper links on power and cost.” Mehta also contends that silicon photonics PICs offer the most effective pathway to high density integration of optical devices in silicon using the widely used silicon foundry and OSAT advanced packaging. Kozlov concurs. “A big trend that we
see is the transition from pluggable optical devices to co-packaged optics” he reports. “I think that in the future it will all converge on integrated solutions, and probably silicon photonics will be the dominant platform – an integrated platform for a variety of materials.” “In the near future, photonics technologies will be used to develop optical interposers – think optical network on a “chip” to enhance GPU networks and memory disaggregation,” predicts Gupta. “The next usage would be photonic compute chips.” For its part Intel is investing heavily in next-generation photonics wafer process technology, which targets >15% power reduction and >40% die area reduction with respect to its current devices. Urricariet discloses that, inter alia, it is also adding performance enhancements such as 200G/lane with high yield, increased laser power, improved optical coupling efficiency, and improved modulator figure of merit. COMMERCIAL BRAKES? The greater commercialisation of PICs has not been without its complications. Increasing integration has meant PIC designers approaching foundries for custom process technology and packaging solutions. Gupta suggests custom solutions are not commercially interesting for high volume foundries and there has been the lack of a common optical test approach. “The ‘customisation’ has been the biggest challenge on the commercialisation of PICs due to the hesitancy of large-scale foundries, OSAT and test vendors to engage in a fragmented market” he maintains. Reliability is another potential obstacle to much wider PIC deployments. “In addition to meeting the aggressive performance requirements in bandwidth density, energy efficiency, latency, and cost, emerging optical I/O solutions also need to offer very high levels of reliability, especially for applications that are not pluggable or easily replaceable in the field,” asserts Urricariet.
reciprocal attractions of AI/ML and PICs may not be equal. While acknowledging that a number of players are breaking new ground and using AI/ML to produce advanced optical element designs that may be counter-intuitive to the human mind, Kozlov judges that it’s early days yet for the commercialisation of such innovations. “I think at the moment it’s probably more that PICs are helping AI and ML systems to become more powerful systems than vice-versa,” he says. “Optics is aiding AI probably more than vice- versa.”
PIC ‘N’ MIX Different PIC platforms have their
particular functional, operational and application features. Although a number of others, including Lithium Niobate and Gallium compounds, are also in the mix, three of the most prominent platforms are SiPho, Indium Phosphide (InP) and Silicon Nitride (SiN). Among other characteristics, InP scores in the realisation of active optical components, SIN has a wide spectral reach and SiPho can leverage the scale economies of the huge global CMOS semiconductor industry. There seems to be some agreement that some of the various platforms complement, rather than compete, with each other. “SiPho needs InP because silicon doesn’t generate light,” observes Kozlov. “There’s always an InP laser next to a SiPho chip. Or even right on a SiPho chip.” “In our case they complement. We need the InP to do the gain medium inside the PIC in silicon,” agrees Carter. “There are companies doing InP circuits, but I think the main challenge there is going to be cost. AI networks in particular, with the volumes that are shipping, are going to be very cost- sensitive.” PIC-TURE THIS What’s up next on the PIC horizon as it relates to AI/ML? Increases in speed look to be required, both in the bandwidth capabilities of photonic elements and the
Manish Mehta VP, Marketing and Operations, Optical Systems Division, Broadcom Incn
Christian Urricariet Senior Director, Product Marketing, Silicon Photonics Product Division, Intel Corp
Dr. Vladimir Kozlov Founder and CEO, LightCounting
Dr. Adam Carter CEO, OpenLight Photonics
Vikas Gupta Senior Director, Product Management, GlobalFoundries
12
www.opticalconnectionsnews.com
| ISSUE 37 | Q2 2024
Made with FlippingBook interactive PDF creator