ams OSRAM has demonstrated an 850nm thin-film microVCSEL platform for highly parallel optical interconnects, combining dense addressable emitter arrays with silicon through-vias and a multi-core fibre connection system developed with BizLink.
The company has demonstrated error-free 32Gbps NRZ operation at approximately 0.25pJ per bit from the optical emitter platform. Reliability testing has also exceeded 2,000 hours under elevated junction-temperature and current-overstress conditions without a recorded failure.
The architecture uses top-emitting vertical-cavity surface-emitting lasers arranged at a 25µm pitch. Individual emitters can be addressed separately, allowing aggregate bandwidth to increase by operating a large number of optical channels in parallel.
ams OSRAM describes the approach as “wide-and-slow”. The term is relative: a 32Gbps lane is not slow in conventional communications terms, but it is less demanding than pushing a much smaller number of serial channels towards progressively higher individual data rates.
Spreading traffic across more lanes can reduce the electronic complexity required around each channel. Drivers and receivers can operate at lower individual speeds, potentially reducing equalisation, latency, and power consumption.
The difficulty moves elsewhere. Highly parallel links require correspondingly dense optical packaging, fibre alignment, connectors, testing, and manufacturing processes capable of handling large numbers of channels without yield falling sharply.
That is where the partnership with BizLink becomes important. Its connectorised multi-core fibre demonstrator combines fibre-array integration and optical packaging with ams OSRAM’s two-dimensional addressable VCSEL arrays.
The system is intended to show how a dense collection of optical emitters can be connected into practical fibre infrastructure without reproducing a conventional separate connector arrangement for every lane.
The thin-film microVCSELs are integrated on silicon substrates using through-silicon vias. TSVs provide electrical connections vertically through the substrate and support compact three-dimensional packages in which emitters, photodetectors, and CMOS electronics can be positioned more closely together.
ams OSRAM is developing that concept into fully integrated 3D photonics stacks combining microVCSEL arrays, micro-photodiode arrays, and mixed-signal CMOS.
Shortening the electrical path between those elements could reduce parasitic losses and allow the optical engine to sit closer to the processor package. That becomes increasingly valuable as AI systems push more data between accelerators and the power cost of electrical interconnects rises.
The company’s manufacturing route has an unusual origin. The underlying highly addressable micro-emitter architecture draws on technology developed for EVIYOS digital automotive lighting, where millions of individually controlled optical elements have already been manufactured for demanding vehicle applications.
Automotive lighting and data communications are not interchangeable, but some manufacturing disciplines overlap. Wafer-level processes, CMOS integration, fine-pitch arrays, advanced packaging, and high-volume reliability controls are all relevant when moving from prototype optical devices into volume interconnect production.
That existing production experience gives ams OSRAM a starting point that differs from a photonics programme developed entirely in the laboratory. The company can draw on processes that have already been required to meet yield and quality targets across very large numbers of devices.
The optical-interconnect application nevertheless creates new problems. Data links demand precise modulation, timing, signal integrity, and fibre alignment rather than simply controlled optical output.
Manufacturing cost may eventually be determined as much by packaging and test as by the VCSEL die itself. Thousands of optical channels have to be aligned, electrically contacted, verified, and connected with sufficient repeatability to make the complete engine economical.
The 850nm wavelength helps because it is already widely used for short-reach multimode-fibre communications. Existing fibre and component ecosystems can therefore provide part of the infrastructure around the new emitter architecture.
The 2,000-hour reliability result is another useful step but does not yet represent qualification of a complete commercial optical engine. Thermal cycling, mechanical stress, fibre attachment, connector contamination, driver electronics, detector matching, and long-term field conditions will all influence final product reliability.
Nor has ams OSRAM announced a date for volume manufacture of the complete AI interconnect platform. The current announcement demonstrates the emitter array, silicon integration, operating efficiency, reliability progress, and a workable multi-core fibre interface.
That puts the programme beyond a single-device experiment but short of a finished production system. The industrial question is now whether the manufacturing methods developed around addressable automotive light sources can deliver the yields, packaging density, and cost required when similar micro-emitter arrays are used to move data.
If they can, the significance may lie less in the individual 32Gbps result than in the ability to replicate hundreds of such channels economically. AI infrastructure is forcing the optical industry to consider bandwidth as a manufacturing-density problem rather than simply a race for higher speed through each individual lane.
