產業_SEMICON2026矽光子論壇_UMC簡報_20260831

PDF 原檔：產業_SEMICON2026矽光子論壇_UMC簡報_20260831_original.pdf

原始內容

Breaking through the DCI data rate barrier

Wenchi Ting, Ph. D. Vice President Corporate Technology Office UMC

Outline

• 1 | Bandwidth scaling laws
• 2 | Challenges from higher data rates
• 3 | Modulator selection
• 4 | Coherent technology comes to rescue?
• 5 | WDM vs single wavelength
• 6 | Summary

UMC

Copyright © UMC

2

Is there a data rate scaling law equivalent to Moore's law for CMOS scaling?

• You bet there is. In fact there are two such laws I managed to find.
• Edholm's Law of Bandwidth: Telecommunication bandwidth doubles approximately every 18 months.
• Nielsen's Law of Internet Bandwidth: Bandwidth capacity roughly doubles every 21 months.
• What does this mean for DCI's (data center interconnects): Roughly every a year and a half to two years, the capacity/module doubles.

per Gemini

Challenges from higher data rates

• In this presentation we focus on the scaling trend of DCI's date rate/channel, not per module, and on the photonic part of the transceivers.
• Critical components in an electrical-optical transceiver:
• The modulator
• The DSP/SERDES
• EIC: TIA and driver
• The photo diode
• 200 Gbps/channel already happened. The most common configuration:
• PAM4
• Si MZM (DFB laser)
• Ge PD
• SiGe BiCMOS TIA/driver
• Advanced logic DSP/SERDES
• IM/DD, single-wavelength
• 400 Gbps/channel in demo status, facing many challenges, as almost every component seems shaky at such data rate.
• Si modulators out of steam
• Ge PD bandwidth very marginal
• Bipolar transistors need to reach 500 GHz f T for TIA/driver
• DSP/SERDES extremely difficult to design

Courtesy of Cisco

UMC

Can this scaling trend go on forever?

• The scaling trend is not likely to continue in a brute-force manner. Both electronics and optics face stiff challenges at 400 Gbps/channel and above.
• SERDES facing CMOS performance limitation.
• On optical side both photo diodes and Si modulators face bandwidth limitations.
• 400 Gpbs/channel (PAM4) requires > 110 GHz electrical 3-dB bandwidth!
• 300 Gbps/channel a 'bridging' data rate spec
• Paths to higher data rates (400 Gbps/channel and above):
• PAM4 -> PAM8
• Adopting components with higher bandwidth
• Coherent transmission
• WDM: multiple wavelengths per channel
• A combination of the above

Modulator selection

• Modulator - maybe the most critical device of the photonic technology.
• Si and SiGe modulators:
• MZM:

Courtesy of Photonic Spectra

Table 1. State-of-the-art high bandwidth all-silicon MMs.

Modulator selection

1. MRM:

Jayatilleka et al., Intel, 2022

UMC

• Small foot print
• Low drive voltage
• High intrinsic electrical bandwidth (due to low resistance)
• Extremely sensitive to temperature
• Demanding on process control
• Trading off extinction ratio against speed
• Not a promising candidate for 400 Gbps/channel and above

Modulator selection

1. SiGe EAM:

Verbist et al., J. Lightwave Technology, 2016

Courtesy of Gemini

Modulator selection

• Other candidates:
• InP EML (laser+EAM) and MZM
• Thin-film LiNbO 3 (TFLN) MZM
• Organic
• Graphene
• Thin-film LiTaO 3
• …
• InP and TFLN are the two most promising solutions, as these are much more mature than the other contenders.

TFLN modulators

• UMC Group's TFLN program started on 2021.
• 6' in mass production, 8' in sampling
• LiNbO3 's excellent optical and E/O properties coupled with mature TFLN substrate infrastructure (up to 8' in substrate size) making TFLN the leading modulator candidate for 400 Gbps/channel and above.

Zhang et al ., Optica, 2021

TFLN pros and cons

• Strengths: low loss, high bandwidth (> 110 GHz), good thermal stability, superb linearity.
• Weaknesses: large, expensive (compared with Si), integration with SiPh requiring either heterogeneous integration or advanced packaging.

Kharel et al ., Optica 2021

Copyright © UMC

TFLN is here to stay

• InP facing severe shortage, adding to TFLN's momentum.
• TFLN the best modulator solution for Coherent Lite given its superb linearity. It is the best solution for 400 Gbps/lane IMDD too.
• Integrating TFLN with PIC is a challenge as well as an opportunity. This is in hot pursuit.
• Advanced packaging schemes and micro transfer printing are both promising candidates.
• UMC is developing the integration technology, leveraging our expertise in PIC, TFLN, and advanced packaging.
• Rather diverse applications of TFLN are being explored, across a number of wavelengths. Innovations taking advantage of LiNbO 3 's unique optical properties are happening at an unexpected pace. This is perhaps only the tip of the iceberg.

Ge Photo Diodes

• Good enough for 200 Gbps (bandwidth ~ 50+ GHz) but marginal at 400 Gbps/channel (110+ GHz).
• May still make it to 400 Gbps/channel with process/layout/design optimization.
• High uncertainty whether Ge PD can ever go beyond 400 Gbps.
• InP very expensive and harder to integrate.
• Alternative materials being explored: GeSn, SiSn, etc., but still early on the learning curve .

Coherent technology comes to the rescue?

• Traditionally, coherent technology was used for telecom only, not for datacom.
• Expensive, bulky, and energy hungry.
• Modulating amplitude, phase, and polarization.
• Carrying huge data load over very long distances.
• Coherent Lite is now entering datacom as a viable solution for :
• Connecting servers that could be a Km away within the same data center.
• Providing relief to the bandwidth challenge.
• Additional optical loss (~ 3 dB) from OCS helps justify Coherent Lite over IM/DD.
• Coherent Lite could be different from a telecom coherent technology:
• O-band
• Shorter reach (2~20 Km)
• Single polarization
• Hardware simplified
• Lower power
• Lower cost

Coherent technology comes to the rescue?

https://techtimes.dexerials.jp/en/photonics/coherent-lite-pic/

https://fundaai.substack.com/p/deepcoherent-lite-reshapes-ai-optical

WDM

• Squeezing multiple  's into the same channel.
• DWDM (8 or more  's/fiber) and CWDM (fewer than 8  's/fiber).
• Actively pursued for scale-up CPO applications to handle the large amount of electronic channels.
• Multiple single- lasers are easier to implement but comb-lasers may be a better solution in the long run due to better power efficiency.

So what is likely to happen at 400 Gbps and beyond, for DCI?

• Adoption of higher-performance modulators (TFLN/InP) and improvement in photo diodes are key factors affecting deployment of plain-vanilla (non-WDM, IM/DD) DCI's with 400 Gbps/channel performance and beyond.
• A combination of WDM, PAM8, and even coherent transmission may be more 'elegant' solutions than brute-force bandwidth scaling. And these are being actively pursued.
• The scaling law in DCI's data rate/channel is likely to slow down as transceivers becoming more complicated to enable WDM and coherent coding schemes, and/or to incorporate non-Si (TFLN, InP) modulators.
• As a result, future connectivity scaling will increasingly rely on heterogeneous integration and advanced photonic platforms, rather than brute-forcing wider bandwidth from each component.

Conclusion

• Winning solutions for future DCI need to be balanced in cost and performance, likely leveraging optical components made by different materials and integrated together seamlessly through advanced packaging schemes.

[讀者資訊已遮蔽] [讀者資訊已遮蔽]

Thank you!

圖片清單（已驗證 2026-10-23）

逐張 Read 驗證（≥40KB 者）。p12／p17 為整頁渲染圖（文字頁），非資料圖。