We're revolutionizing
how data moves.
The Photonic Fabric™
As large AI models with reasoning capabilities and new paradigms like inference-time compute proliferate, traditional data center infrastructure is strained to its limits. The fundamental unit of compute is evolving from a few XPUs in a server to tens of XPUs in a rack to a cluster of thousands of XPUs across multiple racks. Traditional copper-based interconnects are reaching their limits and are unable to scale without incurring excessive power and costs.
The Photonic Fabric™ is a revolutionary optical interconnect technology for scale-up networks that shatters these limitations and enables continued scaling in AI data center compute by delivering very high bandwidth at nano-second latencies with in-network high-performance, high capacity memory and in-network computing.
The Photonic Fabric is the only technology platform addressing the full stack of connectivity, switching, and packaging challenges. By leveraging our expertise in photonics, advanced ASICs, high Speed AMS Design, 2.5D packaging, and software, Photonic Fabric provides customers with solutions that help them design the AI infrastructure for the next decade.
Photonic Fabric technology is available in multiple product forms:
Connectivity (PFLink™) in chiplet or licensable IP
Low-latency high bandwidth scale-up switch (PFSwitch™)
Packaging (OMIB™)
Package-to-Package (XPU-XPU) Scale-Up Network
Photonic Fabric is the only full-stack scale-up networking solution
XPU-to-XPU interconnectivity (PFLink™)
Low-latency, high bandwidth scale-up switch (PFSwitch™)
PFLink supports multiple customer adoption options – Chiplet or IP
Chiplet based on UCIe-A D2D interface with 14.4 Tbps (Gen1) bandwidth
Supports AXI, UAL and other proprietary protocols
Compatible with standard packaging flows
PFLink reach > 50+ meters for multi-rack clusters
PFSwitch – High Bandwidth Low Latency Scale-Up Switch
System-in-package (SiP) consisting of a high bandwidth low latency electronic switch with high bandwidth optical ports
Supports all-to-all communication between XPUs to enable them to act a virtual “Super XPU”
Supports large scale-up domain sizes for 1000s of XPUs
Enables continued scaling of AI models as well as new paradigms like reasoning models & inference time compute
Unlocks large context sizes and batch size scaling improving efficiency
In-Network High Performance Memory
In-Network Computing
In-package (die-to-die) Interconnect (OMIB™)
OMIB™ based on our Photonic Fabric technology platform offers a revolutionary approach to integrate and interconnect multiple reticle-sized die in a package
Scalable to support very large package sizes to integrate 4 or more reticle sized die
Compatible with industry standard packaging flows from multiple OSATs
Allows continued scaling of AI compute within package while reducing package TDP