wall-clock speedup of variance-controlled async RL vs synchronous at equal accuracy (~42h vs ~105h)

Name: wall-clock speedup of variance-controlled async RL vs synchronous at equal accuracy (~42h vs ~105h)
License: https://aidatacenterguide.com/terms#intellectual-property

2.5xobserved

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	One comparison: Qwen2.5-7B Base on long-context, tool-integrated multi-turn RL (SimpleTIR setting), scored on AIME 2025, with 2 B200 GPUs sampling and 2 training. The paper's Table 1 gives 420.2 versus 168.9 GPU-hours for the same task. Not a speedup to assume for async RL at fleet scale.
As of	2026-02
Source	Stable Asynchrony / VCPO (arXiv 2602.17616) · Figure 1 (AIME 2025: about 42 h versus 105 h to the best synchronous accuracy); Appendix F.4 (model and GPUs); abstract: 2.5x speedup in training time
Review	checking…review by 2027-04-30 · standard cadence
Recorded changes	last 2026-10-19 · 2 revisions tracked
Claim id	wall-clock-speedup-of-variance-controlled-async

Where the guide uses it

1.4 Post-Training, Fine-Tuning & RL: The Hybrid Middle

← Full numbers register — every date-stamped figure in the guide, with revision history.