DRACO 4× B200 SXM GPU Server

Overview

The DRACO 4× B200 SXM GPU Server is a Rack 4U rack server that brings NVIDIA’s Blackwell generation into your own data centre. Four NVIDIA B200 SXM (HGX B200 baseboard) GPUs deliver 720 GB HBM3e of HBM3e, linked by NVLink and NVSwitch into one tightly-coupled accelerator with Blackwell’s FP4/FP8 throughput — sized to train and serve up to 405B-class models, behind your firewall, in INR, on a GST invoice.

Engineered for AI platform teams adopting Blackwell on-prem, it pairs the four GPUs with a 2× Intel Xeon 6 host, 2 TB DDR5 ECC and 30 TB NVMe, with 400G InfiniBand for scale-out, redundant power and full BMC/IPMI — a liquid-cooled training node that racks and runs.

Key highlights

• 720 GB HBM3e of HBM3e across 4× B200 SXM — Blackwell-generation memory and bandwidth to train and serve up to 405B-class models.
• NVLink + NVSwitch fabric — full all-to-all GPU bandwidth for efficient tensor-parallel training.
• Blackwell FP4/FP8 — next-generation throughput for both training and high-efficiency inference.
• 2× Intel Xeon 6 + 2 TB DDR5 ECC — high core count and memory bandwidth to feed four Blackwell GPUs.
• Rack 4U, liquid-cooled, redundant PSU, BMC/IPMI — sustained clocks, hot-swap drives, lights-out management.
• 30 TB NVMe + InfiniBand NDR 400G — fast local storage with a high-bandwidth scale-out fabric; no egress fees.
• On-prem data sovereignty — training data and weights stay in-house; DPDP-friendly, air-gappable.
• Make-in-India OEM — predictable INR pricing, GST tax invoice (HSN 8471), pan-India onsite support, GeM-procurable.

AI workload fit (what it actually runs — honestly)

• Training & fine-tuning: full and parameter-efficient training of 405B-class models, tensor- and pipeline-parallel across the four NVSwitch-linked Blackwell GPUs.
• Inference: high-efficiency FP4/FP8 serving of 405B-class models, or several large models concurrently.
• RAG, vision, multimodal & agentic: production pipelines on the 30 TB NVMe and multi-agent back-ends.
• Engineering note: the four SXM GPUs sit on an HGX B200 baseboard with NVLink + NVSwitch — full all-to-all bandwidth for tensor-parallel models; Blackwell’s FP4 path is most valuable for inference efficiency and large-model serving.

AI workload positioning

This sits at the train-and-serve stage with Blackwell-generation efficiency. With 720 GB HBM3e of HBM3e on an NVLink+NVSwitch fabric and 400G InfiniBand, it is sized to sustain real large-model training and high-efficiency inference — a current-generation alternative to renting Blackwell cloud capacity around the clock.

Industry use cases

• BFSI — train and serve private large models under data-residency rules.
• Healthcare & life sciences — on-prem training of medical and imaging models, PHI in-house.
• Government & PSU — sovereign Blackwell-class AI on GeM-procurable infrastructure.
• Neocloud / AI providers — a Blackwell building block for a GPU cloud.
• Telecom & manufacturing — large-scale model development near operations.
• Research & higher-ed — a current-generation institutional training node.

Performance — and how to be sure

We don’t publish inflated peak numbers. The honest picture: 720 GB HBM3e of HBM3e across four NVSwitch-linked Blackwell GPUs is sized to train/fine-tune and serve up to 405B-class models on-prem. Want certainty? Request a free benchmark of your model and dataset on this exact configuration before you buy — we’ll send back real tokens/sec and training/fine-tune timings for your workload.

Series & upgrade path

• DRACO (flagship training tier) — this.
• GPU-count ladder: 4-GPU → 8-GPU Blackwell within the line; step up to 8× B200 or 8× B300 HGX for the largest models.
• When to step up: for multi-rack scale, move to RDP Rack-Scale AI Systems and AI SuperClusters — talk to an architect about the fabric.

On-prem vs cloud — the TCO case

For sustained training and inference, owning beats renting: four continuously-running Blackwell cloud GPUs add up fast, and on-prem removes egress fees and keeps data and weights in-house. RDP pricing is fixed in INR with a GST input-credit-eligible invoice — ask for a 3-year TCO comparison against your current cloud spend.

Software & day-one readiness

Ships pre-configured to train and serve: NVIDIA driver, CUDA, cuDNN, NCCL, the InfiniBand stack, Docker and NVIDIA Container Toolkit, with PyTorch, vLLM / Triton / TensorRT-LLM on Ubuntu LTS. Optional Slurm/Kubernetes, managed AI-stack and observability setup available.

Power, cooling & rack integration

A Rack 4U liquid-cooled node with redundant PSUs and substantial power draw — this node is liquid-cooled; plan CDU/manifold and facility water, plus InfiniBand cabling. (Exact PSU rating, BTU, flow, fabric cabling and rack-depth figures confirmed on the build sheet.) BMC/IPMI provides remote power, console and health monitoring.

Deployment, warranty & support

• Made to order, built, racked-and-stacked, cabled and burned-in in India; realistic lead time confirmed at quote.
• In the box: server, rails, power cables, quick-start, and the pre-installed AI software stack; fabric switches and cabling scoped at quote.
• Onsite warranty + AMC with pan-India coverage and an RMA/escalation path (exact term & response window confirmed at quote).

Why RDP

14 years of Make-in-India infrastructure and 300,000+ devices shipped. Indian OEM, INR pricing, GST tax invoice (HSN 8471), pan-India onsite engineers, GeM availability, and DPDP / sovereign-AI-ready deployment.

Buy with confidence

This is a Blackwell-generation training-and-serving server, made to order — talk to an RDP solution architect, get a configuration and 3-year TCO tailored to your workload, and benchmark your own model on it before you commit. Request a quote to begin.

Specifications