Optoelectronics
AI’s unrelenting power demands drive the need for optics
By Adam Carter, CEO at OpenLight I
t would be hard to underplay the impact AI and machine learning has on the electronics industry, driving rapid expansion in compute infrastructure for even larger and more power-hungry data centres that even recent geopolitical and economic turmoil seems unable to derail. Yole Group estimates that in 2024 alone over $200B has been invested in generative AI (1)
for a number of applications.
To enable the most advanced architectures, the need for bandwidth between elements within the cluster needs to keep pace and, in some instances, even exceed the advances in processing power. At the same time, there is increased pressure to reduce power consumption and latency. Optical technologies, particularly Silicon Photonics and its integration with materials such as Indium Phosphide (InP), will play a crucial role in addressing these escalating demands for high- performance interconnects within AI-driven data centres.
Back-end vs. front-end: tale of two architectures
AI is not only driving the need for more bandwidth but also influencing how data centres are built. We now must consider both the front end – which interfaces with the outside world – and the back end – which provides interconnectivity between compute elements.
The front end of data centres resembles ‘traditional’ Ethernet centric architectures and is expected to grow more slowly compared to the massive explosion in the back-end infrastructure, which is what is now setting the pace on adoption of high-performance technologies, including 1.6T transceivers.
Scaling up and out: the relentless pursuit of bandwidth
Companies big and small in the US, Asia and Europe are developing novel architectures and solutions to solve the challenges of back- end connectivity. AI drives a need to both vertically scale up (adding more resources) as well as horizontally scale out (increasing the size of resources), as line speeds increase to
44 June 2025
greater than 200G/lane. With this shift, the widespread use of pure electrical connections such as Direct Attach Copper cables becomes difficult as signal integrity concerns mean that cables are shorter and physically bigger. While solutions such as adding retimers and equalizers can help, they add power, cost and latency, spurring an interest in using optics for shorter connections between chips on a board. The most obvious candidates for low-cost optics, VCSELs (Vertical-Cavity Surface- Emitting Lasers) are also reaching the limits of their capabilities and are not expected to scale beyond these data rates either.
The need for optics
Optics have always been essential for links longer than a few metres. However, the way optical transceivers have been made from individual piece parts with a DSP to equalize and recover the optical signal does not achieve the scale, cost, or power consumption targets
Components in Electronics
required by the industry. Thankfully, silicon photonics have been quietly waiting in the wings, biding its time until the demand caught up with its capabilities, and that time has arrived. Silicon photonics, circuits that steer and modulate photons, leverages the billions in dollars invested in electronics, providing a path to wafer scale production and tests that drastically change the economics of building transceivers. An 8- or 12-inch silicon wafer has room for many hundred individual transceivers, which are manufactured and tested at the same time using the same Automated Test Equipment (ATE) systems, designed to automate the testing of traditional electronic devices. These ATE systems are widely used today to achieve economies of scale. Photonic Integrated Circuits (PICs), being physically small and made of silicon with similar electrical and thermal properties as electronics, also lend themselves to novel architectures such as CoPackaged Optics (CPO) or optical interposers, allowing
designs that can grow well beyond the 4 and 8 lane applications supported by OSFP and similar pluggable form factors.
The silicon modulator bottleneck: material science to the rescue? Despite its advantages, silicon photonics faces a significant limitation: silicon itself is an indirect bandgap material, meaning it cannot natively generate or detect photons. Whilst germanium doping can deliver excellent detectors, the most important function of a datacom PIC is the modulator. Silicon modulators in either Mach-Zehnder or micro ring configurations have been shown to be able to support data rates up to 200GB/s with the industry standard PAM-4 modulation scheme, but the consensus is that to go beyond this, a change in material is needed. Suitable materials for achieving higher bandwidths range from established solutions such as Lithium Niobate and compound
www.cieonline.co.uk
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56