This article dives into GlobalFoundries’ SCALE optical module—a co-packaged optics (CPO) platform built to speed up AI and hyperscale data-center interconnects. It’s based on GF’s silicon photonics. SCALE aims to beat industry Optical Compute Interconnect Multi-Source Agreement (OCI MSA) specs, pushing high-bandwidth, bi-directional data over fiber using both CWDM and DWDM. Here’s a look at its main capabilities, design choices, and what all this means for AI infrastructure and manufacturing.
What SCALE means for the data center and AI workloads
AI models keep getting bigger, and so do throughput demands. Data centers now need interconnects that deliver more bandwidth with lower latency and less power per bit.
SCALE puts GlobalFoundries right at the crossroads of silicon photonics and modular interconnects. The platform goes beyond copper, offering native DWDM channels and advanced photonic devices.
It tries to blend cutting-edge optics with manufacturing practicality. The goal? Scaling up without giving up on serviceability or yield.
Core capabilities of SCALE
SCALE taps into GF’s established silicon photonics to pull off some impressive features. Here’s what stands out:
- Bi-directional DWDM operation with native 8λ and 16λ support, packing more bandwidth per fiber.
- CWDM and DWDM support for flexible channel spacing and better use of the optical spectrum.
- Detachable fibres that keep insertion loss flat across the CWDM spectrum—good for scaling and easier servicing.
- Broadband detachable fibre approach that moves from 4λ per direction to 8λ and beyond, all while supporting known-good-die testing.
- Fully qualified photonic devices like 50 Gbps and 100 Gbps micro-ring modulators, coupled ring resonators, and integrated photodiodes.
- Through-silicon vias (TSVs) for high-speed signaling and power, making the packaging both compact and dense.
- 2.5D/3D stacking readiness with copper pad pitches shrinking from 110 μm to below 45 μm for tight integration.
- Electrical ICs on single-digit node processes paired with optics, balancing compute and optical efficiency.
Let’s pause on detachable fibres and the flat-loss CWDM spectrum. These features let you deploy and upgrade quickly while keeping things serviceable and testable. You get to scale optical channels without being stuck with a single, fragile fiber setup—huge if your AI workloads are always evolving.
Architectural features enabling scalability
SCALE doesn’t just stop at photonic devices. It brings in some smart architectural moves that really matter for big AI deployments.
- High-speed optical signaling rides on TSV-enabled substrates, which keeps data paths short and power delivery efficient.
- Support for 2.5D/3D stacking boosts density in co-packaged layouts and shrinks the compute node footprint.
- Detachable fibre options keep manufacturing flexible, make field maintenance easier, and let you scale up channel counts as needed.
- A design philosophy that matches broadband optics with single-digit node electrical ICs, so you get compute power without losing optical punch.
GF takes a pretty broad, flexible fibre approach with flat insertion loss across the CWDM spectrum. This makes it possible to scale from just a few channels to dozens per direction. If you’re running AI accelerators or modern data centers, that means less electrical re-wiring and more usable bandwidth per fiber. Sounds like a win, doesn’t it?
Why this matters for AI infrastructure and data centers
AI models keep asking for faster training cycles and lower latency inference. They also need more efficient data movement.
Optical scale-up interconnects are quickly becoming a key piece of the puzzle for architectural efficiency. SCALE brings together OCI MSA-aligned specs, multi-wavelength DWDM channels, and integrated silicon photonics.
This approach opens up a real path to higher bandwidth density per fiber. It also helps reduce cable clutter and makes maintenance less of a headache.
GF embeds high-performance photonics with CMOS-compatible electronics. They pack it all into a denser package, aiming to speed up deployment for AI-ready networks.
They’re also looking to keep things scalable, so future AI advancements won’t hit a wall. GF’s background in silicon photonics and focus on detachable fiber architectures gives data center operators and hardware folks a practical way to scale up optically.
The focus on 8λ–16λ bi-directional operation stands out. Plus, there’s support for DSL-friendly CWDM/DWDM and TSV-enabled substrates.
That all points to a broad, flexible path for optical interconnects across AI, HPC, and hyperscale environments. SCALE mixes proven photonics with scalable packaging to meet the growing bandwidth and flexibility needs of modern AI infrastructure.
Here is the source article for this story: GlobalFoundries Advances Co-Packaged Optics for AI Data Centres