Why Optical Interposers Are The Hidden But Essential Component For The AI Revolution

By Dr. Suresh Venkatesan, Chairman and CEO, POET Technologies

The term “optical interposer” is escalating in consequence for the semiconductor and optoelectronics industry. It was among the buzz phrases at the Optical Fiber Communications (OFC) Conference in March of this year and the recent China International Opto-Electronics (CIOE) Exhibition in Shenzhen, in large part because the AI revolution has created an insatiable demand for data transfer. The desire for more bandwidth, lower latency, and better connectivity is why there is an urgent need for companies in our industry — whether they be decades-old giants or ambitious startups — to adopt an interposer-based approach.

Historically, an electrical interposer has been used as a means to combine special-purpose electronic chips on a single platform to perform multiple functions, rather than making a single general-purpose chip that is larger and less efficient. Electronic chips on an electrical interposer communicate with each other using the high-speed metal layers embedded in the interposer platform. An optical interposer adds the ability to connect photonic chips that emit or detect light to one another using an embedded waveguide layer so that both electronic and photonic chips can communicate with each other at the highest speeds possible. Thus, an optical interposer forms the foundation for combining electronic and photonic components in an optical transceiver – a device that converts electronic signals carried on copper wires into light signals carried on fiber optic cable.

Compared to copper wire carrying electrons, fiber optic cable carrying photons (i.e., light) is a far better means to move data. Light can carry ten times the amount of data without the resistance-produced heat of electrons flowing through copper, and it does so at the speed of light, the fastest anything in the universe can move. An optical interposer populated with tightly integrated electronic and photonic components reduces power consumption, eliminates noise from copper wire bonds, uses fewer components, and can be built with lower labor costs than optical transceiver modules produced any other way.

While once a boutique industry, optoelectronics is increasingly being pursued as a critical part of conventional chip design. Artificial intelligence is driving the urgency for energy efficiency more than ever because it requires so much more data to be communicated at higher speeds than what traditional electronics can support. As Juniper Networks points out in its recent industry report, despite the efforts of semiconductor infrastructure vendors to design more efficient products “the rigorous demands of AI model training are pushing power requirements up — from 10kW to over 100kW per rack in some cases. This creates enormous power and cooling demands.”

When Nvidia unveiled its famous Blackwell AI graphics processor, it highlighted how the breakthrough can reduce “energy consumption by up to 25 times compared to its predecessor.” Experts say that Blackwell requires “approximately two 800G transceivers per GPU, the need for efficient, high-bandwidth communication is becoming more critical for AI, positioning silicon photonics and PICs [photonic integrated circuits] as essential components in the AI-driven future. The biggest driver of the development of PIC transceivers is AI.”

Former Google Chairman Eric Schmidt pointed out that big technology companies are planning for significantly greater investments into hyperscale data centers that will be built at a combined estimated cost of $300 billion and with the specific purpose of supporting the growth of AI manufacturing and software. Also in August, LightCounting published a research report that showed record-setting capital expenditures by Alphabet and Microsoft, and other major investments by tech giants, that are intended to focus on AI-related expansion projects.

Significantly, a sizeable amount of industry spending continues to go to optical connectivity — a vital element of the AI growth story that many people are still unaware of, even if they see the burgeoning market for photonic-integrated transceivers. LightCounting’s recent quarterly industry report shows that “sales of 400G and 800G ethernet transceivers for deployments in AI clusters propelled Q2 2024 sales of optical transceivers to $3.1 billion” compared to $1.9 billion in Q2 2023.

It’s important to note that the ambitious plans so many of the world’s engineers have for AI cannot be achieved without dramatic upgrades in connectivity. Faster data rates must be achieved to develop the fantastical applications that the AI industry envisions. And those speeds cannot live in a lab. They must originate from real-world commercial solutions that are scalable and cost efficient while meeting that prerequisite for low energy usage. As many in the industry professed at OFC 2024, the optical interposer — the indispensable piece from which optical engines and pluggable transceivers are built — is the new way forward. It makes optical connectivity goals feasible and addresses a market for datacom links that is expected to grow from $4.8 billion to $10.9 billion in 2028.

At POET, we foresaw years ago how attractive optical interposers would be. In fact, we conceptualized the POET Optical Interposer™ in 2017 and have had products based on it since 2022. As others try to create or fine tune their optical interposers, our team is far ahead with a platform that has quickly matured and is showing more and more adaptability as we design customized products for our partners — which include Foxconn Interconnect Technology (FIT) and Luxshare Tech. What draws those giants of electronics manufacturing and data center infrastructure to POET’s optical interposer is the architecture that addresses several industry pain points. Two of these include:

• Wafer-scale passive assembly: The high-volume production requirement by AI clusters cannot be cost-effectively scaled by traditional assembly techniques. POET’s optical interposer-based optical engines are assembled and tested at wafer scale. The electronic and photonics components are connected using the built-in high-speed metal traces on the interposer. The photonic devices passively connect using the optical waveguides built into the interposer. That wafer-scale passive assembly process eliminates wire bonds, the requirement for lenses and isolators, and simplifies the system design.
• Cross-talk reduction: The use of high-speed metal traces to connect electrical and optical components instead of wire bonds enables the best performance for high-speed chip-scale design. Combined with the heat spreaders, the metal traces lower the cross-talk which can distort signal output from lasers. When industry observers talk about the need for low-loss laser performance that is what they mean. The lasers must perform at a consistently high level of efficiency to reduce latency and maintain the integrity of the data transfer speed required by each application.

As the appetite for more data speed grows, companies don’t want to feel at a loss about how to reach 1.6Tbps and 3.2Tbps. They need a platform technology that can smoothly support and guide that growth without necessitating the retrofitting of existing architecture or more costly capital expenditures.

While the opto-electronics industry has certainly realized the importance of the optical interposer, soon the various markets for computer hardware and software will see the benefits of the technology, too. The next leap in computing, which will be intrinsically tied to the growth in AI, is certain to have optical interposers at the core of new designs. Even so, many observers may still overlook the technology but an increasing number of engineers who develop chip-level solutions are understanding that the results that these discreet devices provide are nothing short of eye-opening.