SPECIFICATIONS & BUYING DECISIONS
Compare the H100 80GB PCIe card with the H200 NVL card, using per-GPU memory, power and interconnect specifications. These are specific PCIe editions; H100 SXM, H100 NVL and complete HGX or DGX systems are separate configurations.
On smaller screens, scroll the table horizontally. Source numbers link to the manufacturer references below.
| Specification | NVIDIA H100 PCIe 80GB | NVIDIA H200 NVL |
|---|---|---|
| Architecture | Hopper [1] | Hopper [2] |
| Memory per GPU | 80 GB HBM2e with ECC [1] | 141 GB HBM3e [2] |
| Memory bandwidth per GPU | 2,039 GB/s [3] | 4.8 TB/s [2] |
| Power limit per GPU | 350 W; minimum configurable limit 200 W [1] | Up to 600 W, configurable [2] |
| Host connection | PCIe Gen5 x16; Gen5 x8 and Gen4 x16 supported [1] | PCIe Gen5; 128 GB/s bidirectional [2] |
| NVLink reference bandwidth | Up to 600 GB/s between two compatible PCIe H100 GPUs [3] | 900 GB/s [2] |
| MIG partitioning | Up to 7 isolated instances [1] | Up to 7 isolated GPU instances [2] |
| Reference form factor | 10.5 in, full-height/full-length, dual-slot [1] | Dual-slot air-cooled PCIe card [2] |
Specifications are read from the same verified reference records used on our product pages. A reference design is not a guarantee of the exact OEM board, accessories, firmware or system configuration in a quotation.
No SXM benchmark is presented as a PCIe result. Ask for a workload test using the intended card edition, software, batch size and numerical precision before estimating throughput or cost per token.
PLAN THE COMPLETE CONFIGURATION
Compare the model weights, numerical format, KV cache, batch size and application overhead with the per-GPU memory shown below. Additional memory can change which workload fits; it does not by itself establish application throughput.
Review the exact server qualification, PCIe topology, auxiliary power, airflow or liquid-cooling arrangement, rack power and supported firmware. A PCIe connector does not make these cards interchangeable in every server.
A bridge and supported card grouping are separate configuration decisions. Capacities and prices here are per listed GPU, not the combined memory or price of an NVLink pair, four-card assembly or complete server.
PRICE & ORDERING
Prices below come from the current product catalogue and apply to each stated priced variant. Specifications above describe NVIDIA reference hardware. Confirm the complete board/OEM part number, condition, quantity, compatibility, warranty and dispatch before ordering.
CURRENT CATALOGUE OFFER
Call or WhatsApp for our best available price.We’ll prepare a competitive quotation for your exact requirements.
CURRENT CATALOGUE OFFER
Call or WhatsApp for our best available price.We’ll prepare a competitive quotation for your exact requirements.
DEFINE THE WORKLOAD
Share your application, software stack, GPU count, memory needs, host configuration and delivery destination. Compare a complete server or workstation quotation against the component-only options.