by FormFactor
The rapid buildout of AI data centers is putting enormous pressure on the interconnects moving data among GPUs, accelerators, switches, and storage, and adding more racks and processors only gets you so far. The traffic between them also has to move faster, with lower latency and without sending power consumption through the roof. Increasingly, that means optics.
The transition from NRZ to PAM4 signaling has made 224 Gb/s per lane practical, with 448 Gb/s now coming into view. Those lane rates support optical transceivers operating at 1.6 Tb/s today, with 3.2 Tb/s not far behind. Reaching those speeds is not simply a matter of designing faster devices. Packaging and test become increasingly difficult as well, particularly for the photonic integrated circuits (PICs) at the center of these systems. Their electrical and optical functions are closely intertwined, so neither can be characterized effectively in isolation.
Why Photonic ICs Are Different
Conventional RF devices are characterized electrically and optical devices optically, but a PIC requires both to form an astonishing array of functions. For example, a single die can contain photodetectors, modulators, waveguides, optical couplers, lasers, transimpedance amplifiers, drivers, and high-speed electrical interfaces. Performance depends on the interaction among these elements, which means engineers have to evaluate the optical and electrical portions of the device together during development and, ultimately, production.
At the same time, electrical bandwidths continue to rise while optical coupling structures demand increasingly precise alignment. Finding problems before packaging therefore becomes particularly important.
That puts much of the burden on wafer-level test. Packaging represents a significant portion of the cost of many photonic devices, and discovering a bad die after packaging is an expensive way to improve yield. Testing at the wafer level can identify failures earlier and provide valuable process information, but it also brings RF, optical, mechanical, and automation requirements together in the same test environment. That combination is what makes PIC characterization unusually challenging.
The Expanding RF Measurement Challenge
The RF side of the problem begins with bandwidth. The move to 224 Gb/s and eventually 448 Gb/s PAM4 increases the frequency range that must be characterized. In addition, harmonic content also influences signal integrity and the accuracy of device models, so meaningful characterization has to extend well beyond the fundamental frequency. Consequently, measurements that once ended below 70 GHz increasingly extend to 130 GHz, and harmonic characterization is moving toward 200 GHz.
One way to reach those frequencies is to combine several narrower-band measurement systems, but doing so introduces additional calibration steps and makes correlation between setups more difficult. A broadband system that maintains a common reference plane across the entire range can eliminate much of that complexity.
That requirement has driven coaxial interfaces and measurement hardware to millimeter-wave frequencies. Coaxial standards are now available at 168, 220, and 250 GHz, and wafer probes have followed them upward in frequency.
When Repeatability Becomes the Limiting Factor
At these frequencies, instrument bandwidth is only part of the measurement problem because repeatability is just as important. Small differences in probe landing position or overtravel may have little effect at lower frequencies, but at higher frequencies they can produce measurable changes in the result. Mechanical stability, touchdown consistency, and contact quality all begin contributing to uncertainty.
A probe that lands slightly differently from one device to the next can alter the measured response enough to affect the resulting device model. That becomes a serious problem when hundreds or thousands of measurements are being compared across devices, wafers, or test sessions.
For this reason, semi-automated and automated probing has become increasingly important in high-frequency characterization. Controlling the touchdown mechanically removes one source of operator-to-operator and measurement-to-measurement variation. The objective is not only to gather more measurement data at higher frequencies but also to ensure the measurements correlate well enough to be useful.
Calibration Has to Keep Up
Calibration techniques are evolving along with measurement frequency and device complexity. Multi-line TRL, for example, uses multiple calibration structures to maintain accuracy and characterize uncertainty over a broad frequency range. It requires more effort than simpler calibration approaches, but the additional complexity becomes increasingly worthwhile as bandwidth grows.
Differential interfaces add another layer to the problem. Many optical engines rely on differential electrical signaling to achieve high data rates, and crosstalk becomes increasingly troublesome as frequency rises. At that point, the interaction is not limited to the contacts within an individual probe because coupling between adjacent probes can also affect the measurement.
Multi-port calibration techniques, including 32-term and 64-term approaches, can correct for crosstalk that would otherwise degrade differential measurements. So, calibration is becoming less of a procedure performed after the measurement system is assembled and more of an integral part of the system design.
The Optical Alignment Challenge
Wafer-level optical measurements usually require positioning a fiber at a coupling structure on the die, either to launch laser light into the device or collect light coming from it. Alignment tolerances can be measured in microns or even fractions of a micron. Making such small movements is possible, but knowing where the fiber is relative to the optical structure is much more difficult.
PICs employ several coupling methods, including vertical couplers and edge-coupled interfaces. Some of these structures are difficult to view once the wafer is positioned for test. Trenches and V-grooves can place the coupling point where a conventional microscope provides little or no direct visual feedback.
Newer optical probes attack this problem by incorporating fiducials and alignment references into the fiber assembly. Machine-vision systems can locate these references and establish the fiber position without requiring a direct view of the actual coupling interface.
Combined with precision positioning software, this approach can provide sub-micron alignment even when the optical interface itself is obscured. As channel counts increase and optical coupling tolerances continue to shrink, repeatable automated alignment becomes increasingly important.
Measurement Technology Moves to the Foreground
While optical engines, transceivers, and new photonic architectures get most of the attention, the measurement technology required to develop and manufacture them tends to stay in the background, and that’s becoming a lot more difficult to do.
The progression in optical systems from 1.6 Tb/s to 3.2 Tb/s is pushing RF and optical characterization beyond what either discipline can handle independently. Broadband RF measurements, precision optical alignment, calibration, mechanical repeatability, and automated wafer probing increasingly have to operate as parts of the same test system.
PICs will carry a growing share of the bandwidth required by AI infrastructure. As their complexity and operating speeds increase, the ability to characterize them accurately at the wafer before the cost of packaging has been added will become as important as advances in the devices themselves.