Meta RSC-1 revised analysis: what ETTR of about 0.9 requires on the largest feasible run (about 12,000 GPUs, two-thirds of the cluster)
≈10 s checkpoint writes, or failure rate ≈1 (vs 6.50) per 1,000 node-daysmodeled
| Value kind | modeled — Interpret this value according to its displayed kind, scope, source, and as-of date. |
|---|---|
| Scope | Meta's analytical expected-ETTR model (classical checkpointing, Daly-Young-optimal intervals) applied to RSC-1's observed failure rate. A projection for a hypothetical run, not an observation. It replaces the 16,000-GPU 0.70 to 0.93 scenario in the first version of the paper, which the revised version removed. |
| Caveat | Modeled by Meta for a hypothetical run, not observed. |
| As of | 2025-02 |
| Source | Meta, Revisiting Reliability in Large-Scale ML Research Clusters, arXiv 2410.21680v2 (2025-02-06) · Section III, Looking towards the future; Fig. 10 (v2) |
| Review | checking…review by 2027-10-16 · standard cadence |
| Recorded changes | last 2026-10-16 |
| Claim id | meta-rsc1-ettr-0-9-conditions-largest-feasible-run |
Where the guide uses it
← Full numbers register — every date-stamped figure in the guide, with revision history.