Optical is the future, but getting there is harder than it looks.
Key Takeaways
Optical fiber carries signals faster and more efficiently than copper, making it the material of choice for the world’s telecommunications networks. Over shorter distances — between racks and between boards in data centers — designers would like to miniaturize optical components to unlock the energy savings fiber can provide.
AI workloads differ from other HPC workloads in the amount of data movement they require. An LLM query might transmit the same amount of information as a conventional search query between the end user and the data center (“north-south traffic”), but it requires much more data movement within the data center (“east-west traffic”). GPU clusters, as well as the individual GPUs, perform relatively simple multiplication and addition on very large data arrays. But the required bandwidth limits overall performance, and it is a major contributor to power consumption.
Silicon photonics startup Enosemi, which was acquired last year by AMD, estimated in a white paper that leading-edge high-performance ASICs spend as much as half of their total power on data movement.
Mike Hogan, GlobalFoundries’ chief business officer, pointed to four key metrics that determine interconnect efficiency:
At short distances, the relatively large size of optical components limits the bandwidth density. To capture the energy savings photonic interconnects offer, optical components need to scale, along with the electronic circuits they connect.
To that end, designers are considering three possible concepts. The first is pluggable components, which are relatively large, modular, and easy to integrate. Because of their size, though, they require relatively lengthy connections to their control electronics. Pluggables are best suited to connections between boards, between racks, or across the data center.
Second is co-packaged optics (CPO), which as the name implies, integrates discrete optical elements with electronic control circuits, typically connected by wire bonds. Third, optical I/O (OIO) modules combine optical and electronic ICs into a unit that effectively behaves as a single device.
Photonic building blocks
Regardless of size, the basic building blocks of a photonic interconnect are simple. First, they need a light source. At data center distances, this typically is an indium phosphide diode laser. Lasers, by their nature, must withstand higher currents and higher temperatures than passive optical components, and are therefore likely to be the least reliable element in a photonic circuit. Identifying known good lasers before packaging is important for integrated modules. But the ability to replace defective lasers easily is a significant advantage of pluggable components.
Fig. 1: Building blocks for silicon photonics. (a) Single polarization grating coupler, (b) microring modulator, (c) coupling gap for MRM, (d) dual microring resonator. Source: [1]
From there, a modulator breaks the laser’s continuous emission into a stream of data bits. The modulator also defines the data transmission rate. At this time, according to work presented at December’s IEEE Electron Device Meeting by imec’s Joris Van Campenhout, materials like lithium niobate (LiNbO3) can achieve bandwidth in excess of 100 GHz with high efficiency and low losses. However, their large footprint and the risk of lithium contamination are undesirable for direct integration with silicon.[2]
Researchers in Singapore used microtransfer printing to transfer patterned lithium niobate modulators to silicon.[3] Microtransfer printing uses a PDMS stamp to pick-and-place individual devices from a growth wafer to a destination substrate, avoiding the contamination risks that would arise from direct growth on silicon. Still, CPO and especially OIO applications often turn to silicon resonators instead. These use heaters to modulate the refractive index of a doped silicon ring.
NLM Photonics CEO Brad Booth noted that silicon is not a very efficient modulator. The search for alternatives is ongoing. NLM’s solution incorporates a chromaphore into a small molecule organic glass. Processing aligns the chromaphore dipoles, which can be switched via adjacent silicon components.
Once the modulator has defined it, the signal propagates through a waveguide to a photodetector. Silicon is an excellent low-loss waveguide at the wavelengths of interest. Integrated modules, both CPOs and OIOs, use either silicon or organic interposers, which may incorporate photonic components. Germanium works well as a photodetector.
Finally, the circuit needs some form of coupling to carry light from the laser into the waveguide, and from the waveguide to the photodetector. Couplers are responsible for transferring light between materials with different refractive indices, and it’s important to note that coupling losses can be a large fraction of the overall system losses. Depending on the situation, a coupler might simply nudge an optical fiber up against a light source, use polymers to span the gap, or any of a number of other alternatives. Regardless of the design, the coupling should not introduce optical defects, and it should provide for a gradual refractive index transition. A comprehensive review by Keuntae Baek and colleagues at Hanyang University discusses coupling issues in detail.[4]
While these building blocks form an end-to-end optical path, they are not “smart.” The optical circuit still requires control electronics to operate the modulator, to process the signals from the photodetector, and so on. Designers can improve performance and reduce power consumption by reducing the distance between the control electronics and the optical elements.
Electrical-optical integration
Reducing interconnect distance is the goal of heterogeneous integration, in general. Many of the issues encountered by integrated optics are common to other types of advanced packages. The individual components within a heterogeneous package typically use well-understood technologies. The challenge is to connect them in a cost-effective manner.
Combining electrical and optical components presents some new issues, as well. For example, researchers at CEA-Leti embedded a waveguide and other silicon photonic components in a silicon interposer. Such an interposer needs an optical path to lasers and photodetectors on the surface. It also needs through-silicon vias to connect to modulators and, potentially, to the bottom of the package.
But mechanical stresses from these elements can cause optical distortions, leading to optical losses, so design tools for the interposer and the combined package must be able to model both optical and electrical effects of thermal and mechanical stress. “We had to define our own design rules to mix optical and electrical components,” noted Jean Charbonnier, R&D project leader.
For modulators, heaters control the resonant frequency. The group was able to save significant energy by thermally isolating them.[5]
Other problems can crop up, as well. While germanium for photodetectors can be grown directly on silicon, the thick germanium layers they require are not comparable to the thin SiGe nanosheets used in advanced transistors. In this application, epitaxial germanium deposition can account for up to 40% of the overall circuit cost.[6]
In research applications, indium phosphide lasers are often grown separately and transferred to the interposer by microtransfer printing. Handling small numbers of lasers at once provides maximum flexibility and helps ensure that only known-good lasers are transferred. For commercial-scale integration, S. Matsuo and colleagues at NTT demonstrated bonding of an InP wafer to a silicon wafer, followed by in situ growth of an InGaAsP laser.[7]
The electronic portion of the circuit can be manufactured in a conventional CMOS fab and attached to the interposer by copper hybrid bonding. Here, too, the thermal and mechanical stress on the optical elements needs to be accounted for.
While much work remains to develop fully integrated photonic interconnects that will scale with the silicon circuits they support, there isn’t much debate about the ultimate goal. “Across compute vendors, network leaders and silicon suppliers, there is strong alignment,” Hogan said. “Scaling east-west traffic requires optical interconnect.”
References
 |
 |
 |
 |
 |
 |
 |  |
 |  |
 |
 |
Leave a Reply