published scale-out oversubscription examples include 1:1, 2:1–3:1 and a reported 7:1 deployment; they are workload-specific observations, not training/inference topology defaults
1:1 vs 2:1–3:1observedcontested
| Value kind | observed — Reported measurements, counts, and specifications keep the precision and scope stated by their source; an exact specification is not treated as a range. |
|---|---|
| Scope | Named reference designs and deployments with different traffic, topology, placement and service objectives. Meta's 7:1 is the aggregation layer above 3,072-GPU pods, which are non-blocking inside. |
| Caveat | Derive oversubscription from the measured traffic matrix, collective/request mix, topology, failure headroom and SLO; validate it on the target fabric. |
| As of | 2025 |
| Source | SemiAnalysis Neocloud Playbook (newsletter.semianalysis.com/p/ai-neocloud-playbook-and-anatomy); Juniper AI-cluster design; Meta, The Llama 3 Herd of Models (arXiv 2407.21783, 2024-07-31) for the 7:1 case · Meta, The Llama 3 Herd of Models, section 3.3.1: pods of 3,072 GPUs with full bisection bandwidth (no oversubscription); the aggregation layer above them is oversubscribed 1:7. |
| Review | checking…review by 2027-01-16 · fast cadence |
| Recorded changes | last 2026-10-16 · 3 revisions tracked |
| Claim id | training-non-blocking-vs-inference-2 |
Where the guide uses it
- 1.7 The Requirements-and-Consequences Matrix
- 8.1 Network Fundamentals & AI Traffic Characterization
- 8.5 Scale-Out Topology, Sizing & Oversubscription
← Full numbers register — every date-stamped figure in the guide, with revision history.