HWBRIEF
A "factory" here means a purpose built cluster of accelerators running long training jobs, and Cadence says five competing protocols now decide how those machines talk.
Cadence's newest positioning paper argues that an "AI factory" is architecturally distinct from a general-purpose cloud or HPC data center. A factory here is a purpose-built cluster of GPUs or other accelerators running long training jobs, where the binding constraint is no longer raw compute but how fast each accelerator can pull weights from memory and trade gradients with its neighbors.
Cadence names five contenders: UALink (a consortium-backed accelerator-to-accelerator link), NVLink (NVIDIA's proprietary scale-up fabric), Scale-Up Ethernet (SUE, an Ethernet-based proposal for the same job), Ethernet for Scale-up Networking (ESUN), and Ultra Ethernet. Each promises a different mix of bandwidth, latency, and vendor independence, and none has won.
Cadence is shipping pieces of this rewire: a 12.8 Gbps HBM4 IP subsystem for the memory tier, a UCIe chiplet IP subsystem validated on TSMC 3nm for the multi-die side, and a documented partnership with NVIDIA on industrial-scale systems. What the paper cannot settle is the standards war itself.
The consortium math matters more than the brochure. UALink now counts dozens of members, and that consolidation pressure is the real test of whether scale-up networking fragments or converges before the next training-cycle buying season.