問題文
An architect is asked why the reference designs for AI training clusters are described in terms of a repeatable scalable unit rather than as a single flat capacity number. Which reason is correct?
選択肢
- A scalable unit guarantees that any job placed inside it will never traverse a spine switch, because all traffic within one unit is switched locally by a single device regardless of the number of racks in the unit, so the spine layer can be sized from the unit count alone.
- A scalable unit is the smallest number of nodes that a subnet manager can address, and the reference designs quote capacity in these units so that every deployment begins with enough nodes for the management entity to enumerate the fabric in a single pass.
- A scalable unit fixes the ratio of compute racks, leaf switches and uplinks, so capacity grows by replicating a validated block instead of re-deriving the cabling and the switch count each time.
- A scalable unit is a licensing construct that determines how many accelerator hours the cluster may consume in a month, and the reference designs express capacity that way so that procurement can compare quotations from different vendors on a single commercial dimension.