11 Feb 2025
AI data centers and high-performance computing expected to drive huge increase in demand over the coming decade.
The market for photonic integrated circuits (PICs) is predicted to grow by close to an order of magnitude over the next ten years, and reach an annual total of $54 billion by 2035.
That’s according to the latest report from the market consultancy IDTechX, which says that the biggest driver of that demand will be for PIC transceivers used in artificial intelligence (AI) data centers.
“The rise of AI has spurred an unprecedented demand for high-performance transceivers capable of supporting the massive data rates required by AI accelerators and data centers,” stated Sam Dale, a senior technology analyst at IDTechX.
“Silicon photonics and PICs are at the forefront of this revolution, with their ability to transmit data at speeds of 1.6Tbps and beyond.”
AI-driven future
Dale cites Nvidia's latest H200 server units as evidence of the performance requirements, with each GPU said to require multiple 800 Gb/s transceivers, and 3.2 Tb/s transceivers expected to arrive by 2026.
“The need for efficient, high-bandwidth communication is becoming more critical for AI, positioning silicon photonics and PICs as essential components in the AI-driven future,” writes Dale.
In a report co-authored with Yu-Han Chang, he identifies key players in the industry including the collaboration between Intel and Jabil, optics and photonics giant Coherent, and Infinera, who are all actively using PICs within their transceivers.
“InnoLight, a China-based transceiver company, hit 1.6 Tb/s of transfer speed in their latest transceivers in late 2023,” he adds. “Coherent, which has its own indium phosphide (InP) wafer fab facilities, is also developing higher-performance transceivers for 1.6 Tb/s-plus applications.”
Intel’s silicon photonics division, which in late 2023 transferred its transceiver business to Jabil, is said to have shipped more than 8 million PICs since 2016, showing the relative maturity of the technology.
Right now, the PIC market is believed to be worth about $6 billion annually, and already dominated by applications in data centers and high-performance computing.
Non-AI applications
Dale suggests that other PIC applications will soon emerge within that dominant sector, for example in high-bandwidth chip-to-chip interconnects, advanced packaging approaches that connect multiple processor units, and in co-packaged optics.
“These technologies are paving the way for next-generation computing,” Dale suggests, adding that more complex optical components, for example Mach-Zehnder interferometers, could unlock higher speeds than is currently possible.
The proliferation of PICs in datacoms should also help to make other applications viable in the longer term, with the IDTechX report predicting that by 2035 there will be roughly an $11 billion market for non-datacoms use.
The most significant other applications are expected to be telecommunications and lidar, with other uses set to include quantum technology and sensors.
“Certain PIC materials, such as silicon nitride, can be used for a range of different sensors, from gas sensors to ‘artificial noses’,” states the report. “The healthcare sensor industry may be able to take advantage of the miniaturization of optical components into PIC devices, which could see applications in point-of-care diagnostics or wearables.”
PIC development could also help with the emergence of frequency-modulated continuous-wave (FMCW) lidar, which offers major advantages over conventional time-of-flight lidar but relies on complex photonic integration and a relatively powerful and coherent laser source.
“PIC-based FMCW lidar has the potential to transform the automotive and agricultural industries with applications in drones and autonomous vehicles,” writes Dale.
Lithium niobate
In terms of materials, the analyst says that although silicon- and silica-based PICs currently dominate thanks to their light propagation properties, this may change in the future.
Because silicon does not emit light directly, it is usually combined with InP for the source and photodetector elements. The enormous existing silicon integrated circuit manufacturing industry means that situation is unlikely to change too much, although other materials do offer some key advantages.
“Thin-film lithium niobate (TFLN), with its moderate Pockels effect and low material loss, is emerging as a strong contender for applications that require high-performance modulation such as quantum systems or potentially high-performance transceivers in the future,” suggests Dale.
Last year two startups - the Harvard spin-off HyperLight and Swiss firm Lightium - attracted a combined $44 million for their TFLN-based photonic chip technologies, citing key advantages including high optical linearity, a broad transmission range from visible to mid-infrared wavelengths, and nonlinear optical frequency conversion.
The IDTechX team also notes that more exotic materials like barium titanite (BTO) and rare-earth metals are being explored for their potential in quantum computing and other cutting-edge applications.
© 2025 SPIE Europe |
|