Quiz: AI Data Center Networking & Fabrics (Part 8)

10 questions from Part 8 of the guide. Score instantly; every explanation links to the chapter that derives the answer.

In a synchronous AI cluster, the guide argues the network is best understood as…
the part of the computer that decides how fast every other part runsa cost line that scales linearly with GPU counta solved problem since 400G opticsplumbing between computers
Scale-up vs scale-out: what distinguishes the scale-up domain?
it carries only storage trafficaccelerators talk at memory speed over a switched fabric roughly an order of magnitude faster than the back-end networkit uses copper instead of opticsit is limited to two GPUs
How many GPUs share one coherent NVLink domain in a GB200 NVL72, and what does the roadmap point to?
8, growing to 16128, growing to 2561,024, shrinking for cost72, growing to 576 and 1,152
Two vendors quote collective-fabric efficiency percentages. What must be normalized before the numbers can be compared?
Only the switch ASIC generationProtocol, topology, direction and aggregation boundary, message-size/collective mix, endpoint count, and tuning stateNothing — an efficiency percentage is already normalizedOnly the advertised port line rate
The Ultra Ethernet Consortium (UEC) matters because…
its 1.0 spec (June 2025, 100+ members) standardizes an open AI-transport alternative to single-vendor fabricsit only covers opticsit licenses InfiniBand patentsit replaced the IEEE
What is the gating spec when choosing switch silicon, according to the guide?
port count alonethe switch OSthe SerDes generation (lane rate) — everything above inherits itaggregate Tbps marketing numbers
Shallow-buffer vs deep-buffer switch ASICs: which pairing is right?
shallow shared-buffer (Tomahawk 6, 102.4 Tbps, cheapest per port) vs deep-buffer VOQ (Jericho4, HBM packet memory, absorbs incast at a W+$ premium)both buffer in host DRAMbuffering no longer matters at 200G lanesshallow = Jericho with HBM; deep = Tomahawk
What evidence should set a back-end fabric's oversubscription ratio?
The lowest-cost switch quoteThe largest ratio used by another operatorThe workload label alone: training means 1:1 and inference means 3:1The traffic matrix and collective/request mix, topology, tail-latency or step-time SLO, failure headroom, and measured congestion tests
A two-tier fat-tree on 144×800G radix switches scales to roughly how many GPUs?
~1,000~10,368~4,000~100,000
Moving from a congested to a well-engineered fabric does what to GPU idle time, per the guide's figure?
increases it slightly but saves costnothing measurableroughly halves it — ~33% → ~15%eliminates it entirely

Check your understanding: Networking, Fabrics & Optics