projected MTTF for a 16,384-GPU vs 131,072-GPU synchronous job, computed as 1/(nodes × observed failure rate) from Meta's A100 research-cluster data
1.8 hr → 14 minmodeled
| Value kind | modeled — Interpret this value according to its displayed kind, scope, source, and as-of date. |
|---|---|
| Scope | Projection applying MTTF = 1/(nodes × failure rate) to the failure rate Meta observed on jobs over 128 GPUs in its A100 research clusters (11 months, more than 150M A100 GPU-hours). The paper gives 1.8 hours at 16,384 GPUs and 0.23 hours at 131,072 GPUs. Not a measurement, not an H100 or Blackwell-era observation, and separate from Meta's 16,384-H100 Llama 3 run. |
| Caveat | A 1/n projection from Meta's A100 research-cluster rate, not a measurement or H100 behavior. Meta's observed 7.9 h at 1,024 GPUs sits below this line. |
| As of | 2024-10 |
| Source | Meta A100 research-cluster model. This is a hypothetical scale result, not observed H100 behavior or a dated deployment forecast. · Figure 7 and Section III: MTTF = (N_nodes × r_f)^-1 projection by job size |
| Review | checking…review by 2027-05-01 · standard cadence |
| Recorded changes | last 2026-10-16 · 3 revisions tracked |
| Claim id | projected-mean-time-between-failures-for-a-16 |
Where the guide uses it
← Full numbers register — every date-stamped figure in the guide, with revision history.