Co-packaging photonics and electronics poses challenges

 

Beat the co-package heat

The research community and industry are asking questions about how to assemble these different technologies—photonics and electronics—together in the same heterogeneous package. Two of these key questions involve heat: How do we cool the package? And how do we maintain a certain temperature for optimal performance?

Photonics is a thermally sensitive technology, and “if the temperature of the photonic chip changes, its index of refraction changes,” explains Bergman. “We need to be able to adapt the photonic chip to temperature shifts—because temperature changes come from both the surrounding environment and the electrical side of the interface.”

In a three-dimensional (3D) package, heat generated on the electrical side can affect the performance of the optical chip, so it’s crucial to understand the thermal environment and compensate/design for it.

“We approach this in several ways,” Bergman says. “One is closed-loop circuitry that maintains the operation point of the photonics—even if the temperature changes. Another approach is to design the photonics to be as athermal as possible.”

These approaches inherently compensate for changes in temperature. “One material changes its temperature in one direction with the index of refraction, while the other one changes in the opposite direction—and it provides an inherent robustness we can design into the photonics,” Bergman adds.

Electronics, you’re the problem

Zooming out a bit, deep inside systems today “we have very good electrical connectivity in 3D, and the electrical chips between memory and the graphics processing units (GPUs),” says Bergman.

While this connectivity is good in terms of energy consumption and bandwidth, the problem with electronics is that when you need to move the data across the system—it costs a lot of in terms of energy, and the bandwidth can drop by as much as two orders of magnitude.

Bringing photonics right into the chip like an interface can potentially flatten communications across the entire system to eliminate the two-orders-of-magnitude taper that currently exists within systems completely.

“It will greatly accelerate the execution time of the application and the way we architect the systems to begin with,” says Bergman. “Potentially putting photonics within systems isn’t just a technology replacement, it will allow us to accelerate the performance of AI systems by orders of magnitude of what we’re capable of today—while maintaining the energy consumption. We can bend the curve on the energy consumption.”

An inflection point

We’ve reached an inflection point where photonics is still more costly than today’s electronic interconnect infrastructure—because manufacturing and the entire semiconductor ecosystem is much more mature than photonics.

“Companies want to commercialize and deploy these systems but, while energy and performance are important, cost ultimately rises to the top in real systems and projects. So we’re in a catch-22 situation: Can we bring photonics to a full manufacturing modality so there’s large volume to eventually drive costs down?”

Bergman is optimistic we’re getting there, because vendors on the computing side either have or are exploring a program in co-packaged optics/photonics and so it’s indeed on the horizon. “But we’re not quite there yet,” she says.

On the horizon

Photonics will clearly enable extreme-scale computing and beyond. For the electrical version of connectivity’s scaling will mean “we’d essentially need a nuclear power plant for the system’s power consumption,” says Bergman. “This is why it’s important to end the energy consumption curve and enable the scalability of future systems.”

Bergman and colleagues are also working on integrating flexibility into communication systems beyond interconnects to ensure switches are also wavelength-selective to enable systems to adapt to the nature of the communication of a particular application.

“This flexible, adaptable interconnectivity is another exciting area we’re working on, and another one is increasing the bandwidth beyond the wavelength domain—exploring the spatial/modal domain, so for every wavelength you can also have orthogonally independent spatial modes within that color or wavelength,” Bergman says. “When it increases, it’s a wave that increases the bandwidth density even further.”

FURTHER READING

1. A. Rizzo et al., Nat. Photon., 17, 781–790 (Jun. 2023).

2. A. Rizzo et al., IEEE J. Sel. Top. Quantum Electron., 29, 1–20 (Feb. 2023).