DRACO 16× H200 SXM Rack-Scale AI System

Overview

The DRACO 16× H200 SXM Rack-Scale AI System is a turnkey, liquid-cooled quarter-rack AI training cluster — 16 NVIDIA H200 SXM GPUs across 2 HGX nodes, wired into one system with NVLink inside each node and a 400G InfiniBand spine between them. It delivers 2,256 GB HBM3e of aggregate HBM3e and arrives racked, cabled, cooled and tested, ready to train and serve the largest models on-premises — in INR, on a GST invoice.

Engineered for organisations building serious in-house AI capacity, it removes the integration risk of assembling a cluster yourself: RDP sizes the nodes, fabric, storage, power and cooling as one validated system and delivers it as a single SKU with one warranty and one support contract.

Key highlights

• 16× H200 SXM · 2,256 GB HBM3e aggregate HBM3e — cluster-scale GPU memory for trillion-token training and large-model serving.
• NVLink + NVSwitch in each node, 400G InfiniBand spine — full intra-node bandwidth and low-latency inter-node collectives for near-linear scaling.
• 4× Intel Xeon 6 + 4 TB DDR5 ECC — host compute and memory matched to 16 data-centre GPUs.
• 120 TB NVMe NVMe + parallel-FS ready — high-throughput data and checkpoint storage across the cluster.
• Quarter-Rack, liquid-cooled, turnkey — delivered racked, cabled, cooled and burned-in; one SKU, one warranty.
• 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.
• Scale path — grow to larger rack-scale systems and multi-rack AI SuperClusters on the same fabric.

AI workload fit (what it actually runs — honestly)

• Distributed training: data-, tensor- and pipeline-parallel training of 405B+-class models across the 16 GPUs.
• Large-scale inference: serve many large models, or shard the largest models across nodes for high throughput.
• RAG, multimodal & agentic at scale: production AI platforms on the cluster’s storage and fabric.
• Engineering note: within each HGX node the eight GPUs share an NVLink+NVSwitch fabric; across nodes, a 400G InfiniBand spine carries the collective (all-reduce) traffic. Real-world scaling efficiency depends on model and parallelism strategy — we validate it for your workload rather than quoting a peak FLOPS number.

AI workload positioning

This sits at the cluster-scale train-and-serve stage: a complete, validated AI system rather than a single server. With 2,256 GB HBM3e of HBM3e and a balanced NVLink/InfiniBand fabric, it is sized to sustain real distributed training and large-scale serving — the owned alternative to renting an equivalent cloud cluster, where the meter never stops.

Industry use cases

• Government & sovereign AI — a national or departmental AI cluster on GeM-procurable infrastructure.
• BFSI — a private training cluster for large risk, fraud and language models.
• Healthcare & life sciences — large-scale model and imaging training, data in-house.
• Neocloud / AI providers — a validated rack to build or expand a GPU cloud.
• Research & higher-ed — an institutional AI training cluster.
• Large enterprise & telecom — in-house foundation-model and platform development.

Performance — and how to be sure

We don’t publish inflated peak numbers. The honest picture: 2,256 GB HBM3e of HBM3e across 16 NVSwitch-linked GPUs on a 400G InfiniBand spine is sized for 405B+-class distributed training and large-scale serving. Want certainty? Request a free benchmark of your model and dataset — including a scaling test across the nodes — on this exact configuration before you buy; we’ll send back real tokens/sec, scaling efficiency and timings.

Series & upgrade path

• DRACO (flagship rack-scale tier) — this.
• Scale ladder: 16-GPU quarter-rack → 32-GPU half-rack → 64-GPU full-rack; step up to Blackwell B200/GB200/GB300 rack-scale systems for more memory per GPU.
• When to step up: for multi-rack scale, move to RDP AI SuperClusters built from these systems — talk to an architect about the fabric and facility.

On-prem vs cloud — the TCO case

For a sustained training cluster, owning beats renting decisively: an always-on cloud cluster of this size is the dominant line in an AI budget, 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 cluster TCO comparison.

Software & day-one readiness

Ships pre-integrated and validated: NVIDIA driver, CUDA, cuDNN, NCCL, the InfiniBand stack, Docker and NVIDIA Container Toolkit, with Slurm or Kubernetes, PyTorch and vLLM / Triton / TensorRT-LLM on Ubuntu LTS. Optional managed cluster operations, scheduler and observability setup.

Power, cooling & rack integration

A quarter-rack liquid-cooled system — plan facility power, CDU/manifold and water, and the InfiniBand spine. RDP scopes power, cooling and floor/rack requirements as part of the design. (Exact PSU/PDU ratings, BTU, flow and facility figures confirmed on the build sheet.) Full out-of-band management across the cluster.

Deployment, warranty & support

• Made to order, integrated, racked, cabled, cooled and burned-in in India; realistic lead time confirmed at quote.
• Delivered as one system: nodes, fabric switches, cabling, PDUs, cooling integration, and the pre-installed cluster software stack.
• Onsite warranty + AMC with pan-India coverage, cluster-level support 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 turnkey rack-scale AI system, made to order — talk to an RDP solution architect, size the cluster, fabric and facility, get a 3-year TCO, and benchmark your own model with a scaling test before you commit. Request a quote to begin.

Specifications