unified memory llm weights offload gh200

1 article

GH200 for Inference: What NVLink-C2C Memory Really Buys

GH200 pairs 96 or 144 GB of HBM with up to 480 GB of LPDDR5X over 900 GB/s NVLink-C2C. Bandwidth math for weight spill, KV offload and buying decisions.

October 28, 2026• 15 min read
gh200 inferencegrace hopper vs grace blackwellgh200 480gb memory llm inference