NVIDIA Hopper
Industry standard (2022-2024)
Hopper GPU Family
NVIDIA Hopper is the industry standard for AI training, featuring HBM3 memory, the Transformer Engine for LLM workloads, and NVLink 4 interconnects (900 GB/s). The H100 is the most widely-deployed AI GPU in cloud computing today.
6 GPU models in this family
GPU Comparison
| GPU Model | VRAM | Bandwidth | Interconnect | Best For | Starting Price | TDP | Year | |
|---|---|---|---|---|---|---|---|---|
| H200 PCIe PCIe Gen5 | 141 GB HBM3e | 4.8 TB/s | PCIe Gen5 (128 GB/s) — no NVLink | LLM inference | — | 350W | 2024 | View |
| H200 SXM SXM5 | 141 GB HBM3e | 4.8 TB/s | NVLink 4 (900 GB/s) | LLM training (70B+ params) | — | 700W | 2024 | View |
| GH200 Superchip (Grace CPU + Hopper GPU) | 144 GB HBM3e (or 96 GB HBM3) | 4.8 TB/s (HBM3e) | NVLink 4 (900 GB/s) | Large-scale AI training | — | 1,000W | 2023 | View |
| H100 NVL PCIe Gen5 (dual-slot) | 94 GB HBM3 | 3.9 TB/s | NVLink 4 (900 GB/s) — PCIe card with NVLink bridge | AI training | — | 350W | 2023 | View |
| H100 PCIe PCIe Gen5 | 80 GB HBM3 | 3.35 TB/s | PCIe Gen5 (128 GB/s) — no NVLink | AI inference | — | 350W | 2022 | View |
| H100 SXM SXM5 | 80 GB HBM3 | 3.35 TB/s | NVLink 4 (900 GB/s) | AI training | — | 700W | 2022 | View |