NVIDIA
Hopper

The GH100 is NVIDIA's Hopper architecture GPU silicon — the foundational chip that powers all H100 variants (SXM5, PCIe, NVL). Built on TSMC's custom 4N process with 80 billion transistors across an 814 mm² die, it introduced the Transformer Engine for accelerated LLM training and inference. The GH100 die is the largest and most complex GPU NVIDIA had shipped at its 2022 launch.

SXM5 form factor of H100. 80GB HBM3 with full NVLink 4 (900 GB/s). 700W TDP — the highest performance H100 variant for multi-GPU training.

Standard PCIe Gen5 form factor of H100 (no NVLink). 350W TDP — lower performance than SXM but easier to deploy. No multi-GPU NVLink interconnect.

PCIe variant of H100 with 94GB HBM3 (more VRAM than SXM version). Uses NVLink bridge for multi-GPU on PCIe.

SXM5 form factor of H200. 141GB HBM3e with full NVLink 4 (900 GB/s). Highest bandwidth Hopper variant for large model training.

NVL Gen5 form factor of H200. 141GB HBM3e but no NVLink — uses NVL Gen5 (128 GB/s) instead. Lower power (350W) for environments without SXM.

Grace + Hopper superchip with unified memory. Designed for giant-scale AI and HPC workloads.

China export variant of the H100 SXM. Meets US export controls with reduced NVLink bandwidth (400 GB/s vs 900 GB/s) while retaining 80GB HBM3 and full compute density. Built for AI training in compliant regions.

China compliance variant of the H100. Meets US export controls with reduced compute performance but retains full 96GB HBM3 memory and NVLink interconnect. Designed for large-model inference in the China market.