AI & GPU infrastructure
Compare H100, H200 and RTX PRO 6000 Server Edition by the exact variant, then plan the server, power and cooling together.
GPU acceleration
NVIDIA · Data center
Nominal form factor: 10.5 in long × 4.4 in high; dual-slot, air-cooled
Server Edition only. Nominal form-factor dimensions are not the complete bracket-and-cable envelope.
Dimensions & installation clearance →Per-GPU specifications. SXM, PCIe and NVL variants are different products. We confirm the exact OEM option, server configuration, power and cooling requirements with your quotation.
Memory capacity and bandwidth below are per GPU. A bridge group does not turn several cards into a single card, and quoted module power is not the complete server power budget.
| GPU variant | Memory | Bandwidth | Maximum power | Connection |
|---|---|---|---|---|
| H100 SXM 80GB | 80 GBHBM3 | 3.35 TB/s | Up to 700 W | NVLink: 900 GB/s |
| H100 PCIe 80GB | 80 GBHBM2e | 2,000 GB/s | Up to 350 W | NVLink: up to 600 GB/s |
| H200 SXM 141GB | 141 GB per GPUHBM3e | 4.8 TB/s | Up to 700 W | NVLink: 900 GB/s |
| H200 NVL 141GB | 141 GB per GPUHBM3e | 4.8 TB/s | Up to 600 W | 2- or 4-way NVLink bridge: 900 GB/s per GPU |
| RTX PRO 6000 Server | 96 GBGDDR7 with ECC | 1,597 GB/s | Up to 600 W (configurable) | NVLink not supported |
H200 specifications are preliminary and subject to change. H100 PCIe here is the 80 GB HBM2e version, not H100 NVL 94 GB. RTX PRO 6000 is the air-cooled Server Edition, not a Workstation or Max-Q card.
H100 SXM and H200 SXM are modules for an integrated GPU subsystem. The carrier, interconnect, heatsink or liquid-cooling assembly, firmware and system power design form part of that platform. An available PCIe slot is not an installation path for an SXM module. For replacement enquiries, provide the original server model, service identifier and full GPU assembly part number. Confirm whether the supplied item is a module, a service assembly or a complete system.
H100 PCIe, H200 NVL and RTX PRO 6000 Server Edition require a supported server configuration. Check the riser and slot position, front GPU bay, supplemental power cable, power supplies, fans and airflow. Passive cards depend on the server cooling system. For an upgrade, the installed CPU generation and cooling package can restrict the GPU options even when the card appears to fit.
A nominal form factor does not describe every projection. Allow for the I/O bracket, retention hardware, extension pieces, power connector and cable route. Measurements for a bare board must not be substituted for a complete service assembly.
Maximum GPU power is only one part of the system budget. The server configuration, CPU count, redundancy policy and input voltage determine the power supply arrangement. Confirm inlet temperature, altitude limits and the correct high-performance fans.
NVLink bridge requirements differ by GPU variant. Use the appropriate bridge and supported card spacing. A server with eight GPU positions does not imply a single eight-card bridge group, and the maximum GPU count does not approve every possible intermediate population.
A NVIDIA GPU model describes the accelerator family and variant. Dell, HPE and Lenovo may supply that accelerator under their own option or service part numbers, with different brackets, cables, firmware requirements and installation kits. Matching the GPU name alone is not enough to substitute one OEM option for another. Send both the server identity and the full existing part number when requesting a replacement; our quotation will identify the proposed item and included accessories.
Share the exact GPU variant, required quantity, new-build or upgrade requirement, target server and current hardware configuration. Include your rack power, cooling conditions and delivery destination. If the purchase replaces an installed component, a clear label photo and the original OEM part number help avoid a mismatched option. Product images are for presentation; condition, supplied accessories, availability, lead time and delivery terms are confirmed in the formal quotation.
Distinguish PCIe cards from integrated accelerator baseboards.
This guide adds the decisions, documents and next steps that support this page.
GPU platforms
Start with model size, precision, parallelism and dataset movement.
GPU platformsDefine request concurrency, latency target and model-memory footprint.
AI systemsCheck slot width, risers, auxiliary power and supported cooling.
GPU checklistDistinguish PCIe cards from integrated accelerator baseboards.
Plan host networking, GPU interconnect and storage access separately.
Request platform power requirements and facility cooling review.
A planning sequence, not a record of completed work.
Counts show records in the site catalog. They do not represent stock, available quantities or verified compatibility.
Snapshot: 2026-09-15
| Decision | Record in your brief | Evidence to retain |
|---|---|---|
| Accelerator architecture | Distinguish PCIe cards from integrated accelerator baseboards. | Complete identifier and source revision |
| Fabric & data path | Plan host networking, GPU interconnect and storage access separately. | Platform document or written clarification |
| Rack power & cooling | Request platform power requirements and facility cooling review. | Offered-unit details and agreed scope |
Tick items as you prepare them. This is your personal checklist, not a hardware validation.
Saved in this browser when local storage is available.
Reference-based generated scenes; actual-unit photos are available only through an explicit unit-specific request. Related models are separate catalog records, not substitutes for the item on this page.
Continue the relevant directory or service flow.
Open resourceUse the connected tool to develop the requirement.
Open resourceRetain the supporting reference with your project brief.
Open resourceWorkload and precision defined · Accelerator model and quantity named · Platform support documents requested · Power and cooling review assigned.
