Which traffic stays inside a node on NVLink?
Why B is the answer
GPU-to-GPU traffic between local GPUs
NVLink joins GPUs in the same system.
In detail
- NVLink joins GPUs in the same system.
- Traffic between nodes uses the fabric.
Use case
An illustrative scenario, not a real incident.
Two GPUs in one system exchange data over NVLink, so it never reaches the network.
See it in different ways
flowchart TD A["GPU"] -->|"Inside a node: NVLink"| B["GPU"] B -->|"Between nodes: NIC and fabric"| C["GPU in another node"]
mindmap
root((AI factory))
Scale up
NVLink
Scale out
Rail-optimized fabric
Building block
Scalable unit
Management
Out of bandflowchart TD
Q{"Same system?"}
Q -->|"Yes"| A0["NVLink"]
Q -->|"No"| A1["Compute fabric"]Key point: NVLink carries local GPU traffic.
Why the other options are wrong
- A. User traffic from the internet
- That comes through the front end.
- C. Storage traffic to a parallel file system
- That uses a storage network.
- D. Management traffic to each BMC
- That uses the management network.