Preprint / scholarly article
Tail-Limited Useful Compute for Large-Scale AI Training
- Published
- DOI
- 10.5281/zenodo.18051330
Abstract
The preprint defines Tail-Limited Useful Compute, a contract-based telemetry framework that converts tail latency into fail-closed bottleneck certificates and anytime-valid lower bounds on useful progress for large-scale AI training.
Keywords
- tail latency
- useful compute
- AI training
- telemetry contract
- fail-closed verification
- confidence sequences
- bottleneck certificates
- stragglers
- distributed training
- auditability
- throughput floors