GlobalFoundries expands SiGe capacity for Marvell optics

GlobalFoundries and Marvell have expanded their manufacturing relationship through a multi-year agreement to increase silicon-germanium capacity at GlobalFoundries’ Burlington, Vermont facility for high-speed optical networking components.

The additional production will support Marvell devices used in next-generation pluggable optical transceivers, near-packaged optics, and co-packaged optics as AI data-centre networks move towards higher lane rates and greater aggregate bandwidth. Neither company has disclosed wafer volumes, financial terms, or the size of the capacity increase.

The agreement is therefore a manufacturing commitment around an established specialist process rather than the launch of a new semiconductor platform. GlobalFoundries has produced SiGe technologies in Burlington for more than a decade and positions the process alongside its silicon-photonics offering for high-speed optical connectivity.

Silicon germanium occupies an important part of the optical link because transceiver electronics have to operate at very high analogue bandwidth while maintaining gain, noise, linearity, and power efficiency. SiGe bipolar devices can provide high-frequency performance while retaining compatibility with established silicon manufacturing infrastructure, making the technology useful in optical drivers, receivers, and related analogue functions.

Those electronics sit alongside the photonic devices rather than replacing them. Silicon photonics can guide, modulate, and detect light, while high-speed electronic ICs are still required to generate, amplify, receive, and condition signals at either side of the optical interface.

GlobalFoundries has been expanding both technologies as optical functions move closer to switching and compute silicon. The Marvell agreement adds customer-backed capacity to that strategy, tying investment at Burlington to the expected production requirements of a major data-infrastructure semiconductor supplier.

The pressure comes from the structure of large AI systems. Accelerator clusters operate as distributed machines, so overall performance depends on moving data between processors, memory, storage, and network switches rather than increasing arithmetic throughput inside one package alone.

Electrical connections remain effective over short reaches, but loss and power consumption become more difficult to manage as data rates and physical distance rise. Optical interconnects move more of that transport into the optical domain, while leaving demanding analogue electronics at the electrical-optical boundary.

Marvell’s requirement spans several packaging approaches. Conventional pluggable modules place optics at the front panel and offer straightforward replacement, while near-packaged and co-packaged architectures shorten the high-speed electrical path by moving optical conversion closer to the switch ASIC.

That migration tightens the requirements placed on the semiconductor components surrounding the photonic engine. Power density, signal integrity, thermal behaviour, and packaging become more difficult as the optical interface moves into a confined environment close to high-power networking silicon.

SiGe manufacturing therefore forms part of an interconnect chain that includes photonic ICs, laser sources, drivers, receivers, DSPs, substrates, packaging, and fibre coupling. Increasing foundry output at one layer does not remove constraints elsewhere, but it reduces the risk that specialist analogue capacity becomes a bottleneck as optical-module volumes rise.

The agreement also illustrates why established specialist semiconductor processes remain strategically important while much industry investment concentrates on leading-edge digital logic. An accelerator may be manufactured on one of the most advanced available nodes, while its high-speed optical interface depends on analogue and photonic technologies optimised for very different electrical characteristics.

GlobalFoundries has not published the additional wafer output Marvell will receive or when every increment of new capacity will come online. The multi-year commitment nevertheless provides evidence that optical-interconnect demand is translating into foundry-capacity decisions rather than remaining solely on product roadmaps.

As AI systems become larger, the useful performance of the processors increasingly depends on the network surrounding them. Marvell’s decision to secure more Burlington SiGe capacity places manufacturing of those high-speed analogue interfaces alongside accelerator supply as part of the infrastructure scaling problem.