2× RTX 6000 Ada Rackmount RAG Server — 96GB Aggregate VRAM

This planning baseline targets an on-prem RAG node that can separate ingestion, embedding, reranking, and generation work across two professional GPUs. Each RTX 6000 Ada supplies 48GB of ECC graphics memory; the two boards provide 96GB aggregate capacity, not one transparent 96GB memory pool.

Configuration at a glance

Memory boundary: 96GB is aggregate board memory. A normal process still sees 48GB on one GPU; spanning both cards requires explicit framework and model-parallel support.

Where this dual-GPU RAG plan fits

• Assign embedding or reranking services to one GPU and generation services to the other.
• Run multiple model workers or evaluation jobs with explicit per-GPU process placement.
• Use a model-parallel runtime when one workload must span both cards; support and scaling efficiency depend on the software stack.

Platform, storage, power, and cooling checks

• PCIe: ASUS documents four PCIe 5.0 x16 slots with a Xeon W-2400 processor on this W790 board. Verify the final two-card slot placement, riser topology, and chassis clearance against the board and chassis manuals.
• Storage: ASUS documents M.2_1 and M.2_2 as disabled with Xeon W-2400. The modeled second NVMe therefore needs a documented alternate path, such as a supported SlimSAS or add-in adapter, before procurement.
• Power: The two GPUs account for 600W of maximum board power before CPU, memory, drives, fans, and conversion losses. ComputeAtlas models a 1350W target and selects EVGA SuperNOVA 1600 P+ (1600W Platinum); validate connectors, rail distribution, and transient behavior with the final PSU vendor.
• Cooling: Each RTX 6000 Ada is an active dual-slot card, but a dense 4U deployment still requires an unobstructed front-to-back airflow path and a chassis-specific thermal review.

Related planning tools

Specification sources and evidence boundary

Manufacturer sources support the component specifications below. The complete build fit, PSU target, and workload guidance remain ComputeAtlas planning models—not benchmark evidence, live pricing, or procurement approval.

• NVIDIA RTX 6000 specifications48GB ECC GDDR6 per GPU, 300W maximum power, PCIe 4.0 x16, dual-slot active cooling.
• ASUS Pro WS W790E-SAGE SE specificationsCPU-dependent PCIe and M.2 behavior for the W790 platform used by this planning baseline.

Use case

RAG server

Example system budget

$17,500 modeled planning estimate.

Source: ComputeAtlas generated-build budget model (generated-build-budget-v1) • freshness: model-derived • uncertainty: high • not live. Heuristic planning budget derived from ComputeAtlas build-tier formulas. It is not a summed retailer cart, seller quote, or live market observation.

Hardware breakdown

• GPU: 2x RTX 6000 Ada 48GB
• CPU: Xeon w7-2495X
• RAM: Kingston Server Premier DDR5-5600 ECC RDIMM 128GB (4x32GB KSM56R46BD8-32HA)
• Motherboard: ASUS Pro WS W790E-SAGE SE
• Storage: 2x WD Black SN850X 2TB NVMe + 8TB enterprise SSD tier
• PSU: EVGA SuperNOVA 1600 P+ (1600W Platinum)

What This Build Includes

Includes:

• GPU(s)
• CPU
• RAM
• Storage
• Motherboard

Not Included:

• Case / chassis
• Cooling system
• Power cables / adapters
• Peripherals

Deployment Notes

• High-power multi-GPU systems require proper airflow
• Ensure PSU headroom for GPU transient spikes
• Verify motherboard PCIe lane and spacing compatibility
• Suitable for workstation or rack environments