What Is Compute Fabric?_
The compute fabric refers to the interconnection system used to link GPUs together for collective operations like distributed training or inference. In NVIDIA's NVL72 architecture, this is primarily the NVLink spine, a copper-based, low-latency network that stays inside the rack. It is distinct from the scale-out fabric (InfiniBand or Ethernet) that connects racks to each other.
Technical Details
The compute fabric uses NVLink copper cables to directly connect GPU-to-GPU within a rack, avoiding the latency of external fiber links. In the NVL72 configuration, this fabric forms a fully connected topology via NVSwitch using a spine-leaf arrangement, with each GPU able to communicate with others at high bandwidth. The cables use push-pull connectors (NVLink-specific high-density connectors) and are routed through structured cable trays to maintain airflow and serviceability. All cabling and termination must comply with the OEM's bend radius and pull-force specifications to prevent signal degradation.
How Leviathan Systems Works with Compute Fabric
For Leviathan Systems field crews, the compute fabric is the dense copper cabling run between GPU trays and the NVLink switch trays inside the rack. We terminate these cables with care, ensuring no kinks or tight bends, and label each run per the rack layout diagram to simplify troubleshooting.
Related service
Network Testing & Commissioning →