Illustrative 10.11 worked example: break-even full cost per provisioned minute for a prewarmed split prefill/decode pool that passes 97 of 100 requests in the failure test, against an aggregated pool costing $1 per minute that passes 95
≈$1.021 per minute ($97/95; ≈2.1% premium)modeled
| Value kind | modeled — Interpret this value according to its displayed kind, scope, source, and as-of date. |
|---|---|
| Scope | All inputs are the worked example's assumptions: a 60 s replay of 100 requests with 100 output tokens each, joint quality/TTFT/TPOT/completion limits, and $1 per minute of provisioned capacity for the aggregated pool, warm reserve included. No provider performance, incident or price is asserted. |
| As of | 2026-10-19 |
| Source | Guide worked example, Chapter 10.11 (October 2026); the linked NVIDIA Dynamo documentation supplies the pool-separation method, not the numbers. · Chapter 10.11, worked example: pooled or split serving after a decode-worker failure |
| Derivation | Healthy accepted rates 9600/60=160 and 9700/60=161.67 tokens/s; costs 1/0.0096=104.17 and 1/0.0097=103.09 $/M. Failure split: 9000/60=150, $111.11/M, 90% attainment fails 95%; aggregate: 9500/60=158.33, $105.26/M, 95% passes. Flip: 1.01/97 < 1/95; 1.01/0.0097 = 104.12 $/M. Break-even at 97 accepted requests: cost ≤ 97/95 = 1.0210526 $/min (≈2.1% above the aggregated pool's $1/min); at 95 it must be ≤ 1. |
| Review | checking…review by 2027-04-19 · standard cadence |
| Recorded changes | last 2026-10-19 · 2 revisions tracked |
| Claim id | guide2-t210-serving |
Where the guide uses it
← Full numbers register — every date-stamped figure in the guide, with revision history.