Synchronisation Barrier

Appears in 1 paper

A point where all P GPUs pause and wait for the slowest GPU to finish.

As used in Paper 19 — Ring Attention with Blockwise Transformers for Near-Infinite Context →

A point where all P GPUs pause and wait for the slowest GPU to finish. In Ring Attention, barriers occur between rounds (after each GPU completes blockwise attention). Synchronisation overhead is typically <1% with fast networks.