Data-Centre GPUs
NVIDIA data-centre GPUs for training and inference — as individual cards, HGX baseboards or complete servers.
B200
Blackwell flagship for frontier training and large-scale inference, shipped on eight-GPU HGX baseboards.
1 configuration → NVIDIA · HopperH200
Hopper compute with 141 GB of HBM3e — 76% more memory than the H100, for serving large models.
2 configurations → NVIDIA · HopperH100
The workhorse of today's AI clusters — SXM for dense training nodes, PCIe and NVL for standard servers.
3 configurations → NVIDIA · AmpereA100
Still widely deployed for training and inference where cost per GPU matters more than peak speed.
3 configurations → NVIDIA · Ada LovelaceL40S
Data-centre card for inference, fine-tuning and graphics workloads in standard PCIe servers.
1 configuration →Every configuration, side by side
Per-GPU figures from NVIDIA's specifications. Pricing and lead times depend on quantity and whether you need bare cards, HGX baseboards or complete servers.
| Model | Architecture | Memory | Bandwidth | Form factor | NVLink | Max power |
|---|---|---|---|---|---|---|
| NVIDIA B200 SXM | Blackwell | 180 GB HBM3e | 8 TB/s | SXM (HGX B200) | 1.8 TB/s | 1,000 W |
| NVIDIA H200 SXM | Hopper | 141 GB HBM3e | 4.8 TB/s | SXM | 900 GB/s | 700 W |
| NVIDIA H200 NVL | Hopper | 141 GB HBM3e | 4.8 TB/s | PCIe, dual-slot | 900 GB/s | 600 W |
| NVIDIA H100 SXM | Hopper | 80 GB HBM3 | 3.35 TB/s | SXM | 900 GB/s | 700 W |
| NVIDIA H100 NVL | Hopper | 94 GB HBM3 | 3.9 TB/s | PCIe, dual-slot | 600 GB/s | 400 W |
| NVIDIA H100 PCIe | Hopper | 80 GB HBM2e | 2 TB/s | PCIe, dual-slot | 600 GB/s | 350 W |
| NVIDIA A100 80GB SXM | Ampere | 80 GB HBM2e | 2.04 TB/s | SXM4 | 600 GB/s | 400 W |
| NVIDIA A100 80GB PCIe | Ampere | 80 GB HBM2e | 1.94 TB/s | PCIe, dual-slot | 600 GB/s | 300 W |
| NVIDIA A100 40GB PCIe | Ampere | 40 GB HBM2 | 1.56 TB/s | PCIe, dual-slot | 600 GB/s | 250 W |
| NVIDIA L40S | Ada Lovelace | 48 GB GDDR6 ECC | 864 GB/s | PCIe, dual-slot | None | 350 W |
Building out a cluster?
Tell us the GPU count, the form factor and when it needs to be racked. We come back with availability and a landed price.
Talk to sales