The Definitive Guide toAI Data Centers
Ask the GuideAboutAccount
GuideGlossaryAll-reduce

All-reduce

A collective operation that sums gradients across all GPUs and shares the result, the dominant traffic in training.

Current numbers

~2xeffective all-reduce bandwidth from in-network reduction (SHARP) vs non-SHARP configas of 2025 · register ↗
870-928 GB/sin-domain all-reduce busbw on GB200 NVL72 (~saturates the 900 GB/s/GPU NVLink5 unidirectional rate; NVLS lets ring-convention busbw read slightly above it); scale-out gate set as % of thisas of 2025 · register ↗

← All terms