GPU selector
Pick what you'll run, compare the cards side by side, and see which servers can take each one.
Running a trained model so people and apps can use it.
6 GPUs suggested for Inference.
A low-power, low-profile card for inference, video and graphics.
Suits Inference, Rendering
A general-purpose card for LLM inference and training, 3D graphics and video.
Suits Inference, Training, Rendering
The most memory per card here, for LLM inference and HPC in air-cooled racks.
A power-efficient single-slot card for small-model inference, video and graphics.
Suits Inference, Rendering
A 96 GB data centre card for generative AI inference, rendering and scientific computing.
A 300 W, 96 GB workstation card for fine-tuning LLMs, AI development, rendering and simulation.
Suits Inference, Training, Fine-tuning, Rendering, HPC
The GPUs that suit Inference are highlighted. Change the tab to see others.
| Specification | NVIDIA L4Suits Inference | NVIDIA L40SSuits Inference | NVIDIA H200 NVLSuits Inference | NVIDIA RTX PRO 4500 Blackwell Server EditionSuits Inference | NVIDIA RTX PRO 6000 Blackwell Server EditionSuits Inference | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionSuits Inference |
|---|---|---|---|---|---|---|
| GPU memory | 24 GB | 48 GB | 141 GB | 32 GB | 96 GB | 96 GB |
| Max power | 72 W | 350 W | Up to 600 W, configurable | 165 W | Up to 600 W, configurable | 300 W |
| Architecture | NVIDIA Ada Lovelace | NVIDIA Ada Lovelace | NVIDIA Hopper | NVIDIA Blackwell | NVIDIA Blackwell | NVIDIA Blackwell |
| Memory type | GDDR6 | GDDR6 with ECC | HBM3e | GDDR7 | GDDR7 with ECC | GDDR7 with ECC |
| Form factor | 1-slot low-profile, PCIe | 4.4" (H) x 10.5" (L), dual slot | PCIe, dual-slot, air-cooled | Single-slot, FHFL (4.4" H x 10.5" L) | 4.4" (H) x 10.5" (L), dual slot | 4.4" x 10.5" L, dual slot, full height |
| Interconnect | PCIe Gen4 x16, 64 GB/s | PCIe Gen4 x16, 64 GB/s bidirectional | 2- or 4-way NVLink bridge, 900 GB/s per GPU; PCIe Gen5, 128 GB/s | PCIe 5.0 x16 | PCIe 5.0 x16 | PCIe 5.0 x16 |
| MIG | Not listed | Not supported | Up to 7 at 16.5 GB each | Up to 2 at 16 GB each | Up to 4 at 24 GB each | Up to 4 at 24 GB each |
| Servers that can take it |
|
|
|
|
|
|
| Source | NVIDIA L4 datasheet and product brief | NVIDIA L40S product page | NVIDIA H200 datasheet and H200 NVL product brief | NVIDIA RTX PRO 4500 Blackwell Server Edition datasheet, April 2026 | NVIDIA RTX PRO 6000 Blackwell Server Edition datasheet, December 2025 | NVIDIA RTX PRO 6000 Blackwell Max-Q datasheet, June 2026, and product page |
Source: NVIDIA L40S product page
Source: NVIDIA RTX PRO 4500 Blackwell Server Edition datasheet, April 2026
Source: NVIDIA RTX PRO 6000 Blackwell Server Edition datasheet, December 2025
Source: NVIDIA RTX PRO 6000 Blackwell Max-Q datasheet, June 2026, and product page
Published figures for one card, with the source beside each.
Per card, largest first. More memory fits larger models on one card.
Per card, largest first. Plan your power and cooling around this.
Configure a server around it, or tell us about your workload. We help you source and deploy the system that suits it.
Every figure on this page comes from the source listed for that card, and photos come from the manufacturer's pages. “Compatibility confirmed on your quote” means the ASUS compatibility list doesn't name that card for that server.