ZeRO · Zero Redundancy Optimizer
A technique partitioning optimizer state, gradients and parameters across GPUs to remove memory redundancy in training.
Current numbers
800 Gb/sBlueField-4 DPU throughput; 64 Arm cores, ~6x BlueField-3 compute; zero-trust east-west enforcement at line rate
SP 800-207NIST Zero Trust Architecture — 'never trust, always verify'; no trust from network location
Measure resume, model-ready and TTFT separatelyserverless GPU time-to-first-token (H100): warm-pool vs scale-from-zero; snapshots claim ~10x cold-start gains