inference/research
draftlow · 0.60updated 2026-10-15

NVIDIA GH200 Grace Hopper (96GB HBM3)

nvidia hopper + grace (GH100/Grace) 2023
This record is a draft. Not yet verified. Provisional. Treat every number on this page as provisional.

Fields

id
accelerators/nvidia-gh200
Vendor
nvidia
Architecture
hopper + grace (GH100/Grace)
Release year
2023
Form factors
socs
VRAM (GB)
96
Memory type
hbm3 + lpddr5x unified
Memory bandwidth (GB/s)
4,000
Bandwidth basis
4 TB/s HBM3 figure as used in NVIDIA GH200 materials; TRT-LLM page lists 'GH200 96GB HBM3 (480GB LPDDR5X)'. Recorded as nominal HBM bandwidth pending datasheet transcription.
Unified memory
yes
Consumer part
no
notes
Grace+Hopper superchip: 96 GB HBM3 on the GPU plus 480 GB LPDDR5X on the Grace CPU, coherent over NVLink-C2C (900 GB/s). The big coherent pool is the LLM-relevant property. TRT-LLM vendor throughput numbers recorded in data/benchmarks/trtllm-gh200-*.

Notes

Grace+Hopper superchip: 96 GB HBM3 on the GPU plus 480 GB LPDDR5X on the Grace CPU, coherent over NVLink-C2C (900 GB/s). The big coherent pool is the LLM-relevant property. TRT-LLM vendor throughput numbers recorded in data/benchmarks/trtllm-gh200-*.

Cross-references

Interconnect (1)