A $100M compute study says layer-loop beats model-loop for large MoE pretraining
teortaxesTex · x · 2026-07-21
The first author says the team spent hundreds of millions of dollars in compute validating the classical Universal Transformer, or “model-loop,” setup.
They found that model-loop does not scale well to billion-parameter MoE models. In later-stage pretraining, their “layer-loop” approach outperformed model-loop, stayed ahead through the end of training, and was friendlier to infrastructure and pipeline parallelism.
They also argue that, at equal theoretical FLOPs, the looped approach can beat non-loop baselines, and that model-loop places parameters and gradients farther apart, which may slow training as scale grows.
Related event: Loopie Looping Transformers Match Larger Models at Fraction of Cost(7 posts)→
More from Infra
- Musk: robot RAM demand will dwarf everything else, and 100% of superintelligence compute goes to deep space — elonmusk · 2026-10-27
- RTX 5090 hits $7K as PC gaming community goes into 'denial' over AI boom — PAstynome · 2026-10-27
- Wan2GP lands on Pinokio: 1-click AI video on 6GB GPUs — cocktailpeanut · 2026-10-27
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27