GlobalFoundries Thomas Barber: CPO, Silicon Photonics, 300mm, SiGe, OCI, NRZ

Summary

Thomas Barber lays out GlobalFoundries’ silicon photonics strategy as part of its broader “essential silicon” focus for data centers, spanning communications, power, and memory. He traces GF’s photonics roots back to IBM Microelectronics, noting a decade-long development arc from early waveguide/modulator/photodiode integration to mass-volume 25 Gb/lane shipping, then highlights the company’s early move to 300mm manufacturing in Malta as a major capacity and cost advantage. Barber argues the real market constraint is copper: as data rates rise, copper reach collapses, making optics necessary for modern AI scale-up networks. He walks through the progression from pluggables to NPO to CPO, emphasizing that the main win is power efficiency rather than raw speed. He also explains OCI’s decision to favor NRZ and wavelength multiplexing over PAM4, and why GF’s silicon germanium process matters for the very high-speed drivers and TIAs needed in these systems.

Notes

• GlobalFoundries frames silicon photonics as part of its “essential silicon” portfolio for data centers.
• Barber positions photonics as strategically aligned with GF’s broader data-center business.
• GF’s photonics work traces back to IBM Microelectronics and Chartered Semiconductor acquisitions.
• IBM Microelectronics started the photonics effort in the mid-2000s.
• By the mid-2010s, GF had integrated waveguides, modulators, and photodiodes into CMOS.
• About 10 years of work led to mass-volume 25 Gb/lane products.
• GF says it is the largest pure-play silicon photonics foundry by revenue.
• Strategic customers and market leaders were central to that revenue lead.
• GF moved silicon photonics production to 300mm in Malta, New York.
• GF says the AMF team in Singapore broadened the portfolio beyond its core silicon photonics set.
• Moving from 200mm to 300mm yields 2.25x more die per wafer.
• Most silicon photonics production is still on 200mm wafers, according to Barber.
• Copper reach shrinks sharply as speed rises.
• At 100 Gb/s, direct-attach copper reaches about 2 meters.
• At 200 Gb/s, copper reach falls to about 1 meter.
• At 400 Gb/s, copper reach falls to about 0.5 meter.
• Optics become necessary once scale-up spans multiple racks.
• Google and Huawei are cited as examples of multi-rack scale-up networking.
• Barber describes a migration from pluggables to NPO to CPO.
• Pluggables are around 35 dB loss from rack edge to ASIC.
• NPO cuts loss to roughly 15-20 dB.
• CPO brings loss down to about 6 dB.
• Pluggables are described as consuming about 20-25 pJ/bit.
• NPO can get to about 10 pJ/bit by dropping DSP.
• CPO can get below 5 pJ/bit with lower-power SerDes.
• The system-level goal is to minimize picojoules per bit.
• Pluggable power is held back by long PCB traces and high-bandwidth PAM4 signals.
• CPO reliability concerns center on permanence and serviceability.
• Meta’s long-running tests reportedly found better reliability for CPO than pluggables.
• Dust in connector interfaces is a real-world pluggable failure mode.
• Integrated assembly can eliminate connector dust issues.
• OCI MSA is a multi-source agreement that narrows standard options into a product spec.
• OCI is aimed specifically at scale-up computing.
• OCI uses NRZ at 50 GHz instead of PAM4.
• OCI reaches 200 Gb/s per fiber by using four wavelengths.
• Barber says NRZ’s native bit error rate is about a million times lower than PAM4.
• NRZ is simpler to receive and lowers power.
• GF’s Scale platform packages electrical-to-optical translation in a known-good form factor.
• Scale combines an electronic IC, a photonic IC, and a detachable optical connector.
• GF’s module is intended to be assembled like a wiring harness at the end of server buildout.
• GF says it can manufacture the photonic IC in-house.
• GF can pair photonics with FINFET, FDX, silicon germanium, or customer-supplied advanced nodes.
• GF also handles assembly and test for micro-optics.
• The resulting product is a known-good, fully tested, high-reliability module.
• GF’s coupling uses an implanted micro-mirror instead of a grating coupler.
• The micro-mirror approach preserves full O-band flexibility.
• Known-good die requires two-sided test flows.
• Optical engines must be tested at much higher bandwidth and channel count than standard transceivers.
• Normal pluggable transceivers top out around 1.6 Tb/s.
• Barber argues scale-up optical interconnect can enlarge the market by 5x-10x.
• The enemy is copper, not other foundries.
• Silicon germanium is used for high-speed modulator drivers and TIAs.
• 200 Gb/s PAM4 systems need roughly 55 GHz bandwidth; linear systems closer to 70 GHz.
• Transistor speed needs to be about 5x bandwidth, implying 350-400 GHz for 200 Gb/lane.
• The 200 Gb/lane electrical path creates a major silicon germanium performance requirement.