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…
plumbing between computersthe part of the computer that decides how fast every other part runsa cost line that scales linearly with GPU counta solved problem since 400G optics
Scale-up vs scale-out: what distinguishes the scale-up domain?
it uses copper instead of opticsaccelerators talk at memory speed over a switched fabric roughly an order of magnitude faster than the back-end networkit is limited to two GPUsit carries only storage traffic
How many GPUs share one coherent NVLink domain in a GB200 NVL72, and what does the roadmap point to?
8, growing to 1672, growing to 576 and 1,152128, growing to 2561,024, shrinking for cost
Effective throughput under collective load: how do InfiniBand and tuned Ethernet compare in 2026?
IB ~98–99% vs scheduled/tuned Ethernet ~95%IB ~80% vs Ethernet ~99%both plateau near 70%Ethernet cannot run collectives
The Ultra Ethernet Consortium (UEC) matters because…
it replaced the IEEEits 1.0 spec (June 2025, 100+ members) standardizes an open AI-transport alternative to single-vendor fabricsit licenses InfiniBand patentsit only covers optics
What is the gating spec when choosing switch silicon, according to the guide?
aggregate Tbps marketing numbersthe SerDes generation (lane rate) — everything above inherits itthe switch OSport count alone
Shallow-buffer vs deep-buffer switch ASICs: which pairing is right?
shallow = Jericho with HBM; deep = Tomahawkshallow 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 lanes
Training vs inference fabrics: what oversubscription does the guide recommend?
1:1 non-blocking for training; 2:1–3:1 acceptable for inference3:1 for training; 1:1 for inferenceoversubscription is obsolete8:1 everywhere to save cost
A two-tier fat-tree on 144×800G radix switches scales to roughly how many GPUs?
~1,000~4,000~10,000~100,000
Moving from a congested to a well-engineered fabric does what to GPU idle time, per the guide's figure?
nothing measurableroughly halves it — ~33% → ~15%eliminates it entirelyincreases it slightly but saves cost

Check your understanding: Networking, Fabrics & Optics