What is a rail-optimized topology designed to reduce?
Why A is the answer
Hops for traffic between matching NICs
Same-numbered NICs share leaf switches.
In detail
- Same-numbered NICs share leaf switches, which reduces hops.
- This suits collectives.
Use case
An illustrative scenario, not a real incident.
Traffic between matching NICs stays on one leaf, so it crosses fewer hops.
See it in different ways
flowchart TD L1["Leaf for rail 1"] --- A["Server A NIC 1"] L1 --- B["Server B NIC 1"] L2["Leaf for rail 2"] --- C["Server A NIC 2"] L2 --- D["Server B NIC 2"]
mindmap
root((AI factory))
Scale up
NVLink
Scale out
Rail-optimized fabric
Building block
Scalable unit
Management
Out of bandmindmap
root(("Rail-optimized design"))
n1["Leaf 1"]
n2["NIC 1 of each server"]
n3["Leaf 2"]
n4["NIC 2 of each server"]Key point: Rails reduce hops.
Why the other options are wrong
- B. The number of GPUs in each server
- It does not change GPU count.
- C. The amount of power used by storage
- It is not about storage power.
- D. The time taken to rack each server
- It is not about racking time.