UALink Design & Verification
AI models are trained across pods of accelerators that must behave like one machine. UALink is the open-standard answer for the scale-up fabric inside the pod—delivering the memory-semantic communication these applications require.
network
memory-semantic
read/write
between accelerators
sub-μs
packet semantics
The problem UALink addresses.
Scale-up ≠ scale-out
Memory-semantic operations at far lower latency. Ethernet was not designed for this domain.
Ecosystem openness
An open standard lets vendors mix silicon the way PCIe and Ethernet did.
Architecture headroom
Pod sizes, topology, switch radix, and implementation decisions interact with the fabric spec.
| Characteristic | UALink | Ethernet UEC |
|---|---|---|
| Domain | Scale-up | Scale-out |
| Connectivity | Accelerator-to-accelerator | Node / rack / cluster |
| Communication model | Memory-semantic (load/store) | Network / transport (packets) |
| Typical distance | Intra-pod (short) | Inter-node (longer) |
| Primary challenge | Low latency, high bandwidth | Scale, traffic, reliability |
Where SNOVA contributes.
Scale-up architecture engineering
Endpoint integration, topology selection, and implementation choices that fit workload and silicon constraints.
Design
RTL design and integration for UALink protocol logic with clean interfaces and reuse in mind.
Verification
Protocol-level environments, compliance test planning, and coverage that targets real failure modes.
Adjacent-stack fluency
Ethernet/UEC and UCIe context so UALink verification is never an isolated activity.
UALink lifecycle.
review
architecture
design
environment
test plan
integration
interop
Related reading
Scale-Up vs Scale-Out: How UALink and Ethernet Divide the AI Fabric
A practical look at when to use UALink, when to use Ethernet/UEC, and how they complement each other.