GPU Servers — NVIDIA / AMD Accelerated Compute

GPU servers are no longer just for AI — they power VDI virtual desktops, real-time ray-traced rendering, scientific computing (genomics, fluid dynamics, financial Monte Carlo), and video encoding/transcoding pipelines.

The hardware question is about how many GPUs you need, what generation (NVIDIA Ampere / Hopper / Blackwell, AMD CDNA), and whether you need NVLink or PCIe-only topology. PCIe Gen4 x16 gives 32 GB/s GPU-to-GPU; NVLink gives 600 GB/s on Hopper.

Pro Disk Network stocks the major GPU-server platforms: Dell PowerEdge XE9680 (8× H100 with NVLink mesh), R750xa (4× double-wide cards, PCIe), HP ProLiant DL385 Gen10 Plus v2 (3× A100), Supermicro SuperServer 4U/8U with 8× Tesla / Instinct, and Lenovo ThinkSystem SR670 V2 (4× A100 with NVLink).

We also stock the GPU cards themselves (new and refurbished), riser cards, GPU-specific PCIe cooling shrouds, and the upgraded high-wattage power supplies (1600W+, Platinum efficiency) that GPU configurations require. Single-card prices: Tesla V100 32GB ~$1.5-3K refurb, A100 80GB ~$10-20K, H100 80GB ~$25-35K.

Featured GPU Servers — NVIDIA / AMD Accelerated Compute

Frequently Asked Questions

How many GPUs can I fit in a 2U server?

Up to 4 PCIe-form-factor cards (2× double-wide × 2 slots each), or 2 SXM-form-factor cards. For 4× SXM (NVLink-connected) or 8 PCIe cards, you need a 4U chassis like the Dell XE8545, HPE Apollo 6500, or Supermicro 4029GP.

What's the difference between PCIe and SXM GPUs?

PCIe cards use standard slots (250-300W, x16 lanes). SXM is NVIDIA's higher-power form factor (400-700W) with NVLink between cards, only available in specific server chassis. For training workloads SXM is significantly faster; for inference, PCIe is fine and much cheaper.

Do I need MIG-enabled GPUs?

MIG (Multi-Instance GPU) lets you partition an A100 / H100 into up to 7 independent GPU instances, each with isolated memory and SM allocation. Useful for cloud / multi-tenant inference. Not needed for single-job training or rendering. All A100 SKUs we stock support MIG.

Can I retrofit an existing R740 with GPUs?

Limited yes — R740 supports up to 3× NVIDIA T4 (low-profile, 70W) or 1× double-wide card with the GPU-enabled riser kit. For real GPU work you want R740xa, R750xa, or one of the dedicated GPU SKUs above. We can quote the riser kit + cooling + power supply upgrade if R740 retrofit is the right path.

Other Use-Case Hardware

Part of

AI / ML Infrastructure Hub

View all 36 pages →

Server CPUs and GPUs for AI/ML and HPC — Intel Xeon, AMD EPYC, NVIDIA A100/H100, AMD MI300, RTX 6000.

Related Pages