DRACO 128-Node InfiniBand XDR Fabric

Overview

The DRACO 128-Node InfiniBand XDR Fabric is a turnkey InfiniBand fabric that wires a GPU cluster end-to-end — switches, the subnet manager and the leaf-spine topology — providing 128 nodes · 512× 800G at 410 Tb/s aggregate over InfiniBand XDR (Quantum-X800), with Sub-600 ns. It is the interconnect that turns a hall of GPU servers into one cluster, carrying the collective (all-reduce) traffic that dominates multi-node AI training — on-premises, in INR, on a GST invoice.

Engineered as the fabric layer of an RDP AI cluster or supercluster, it is delivered designed, cabled and validated alongside the compute and storage it connects, as a single non-blocking system with one warranty and one support contract.

Key highlights

• 128 nodes · 512× 800G · 410 Tb/s aggregate — a complete non-blocking fabric to wire 128 GPU nodes at full bandwidth.
• InfiniBand XDR (Quantum-X800) — RDMA, lossless, in-network compute (SHARP) for fast collectives.
• Sub-600 ns — low latency so distributed training scales with near-linear efficiency.
• 800G per port — matched to your GPU NICs/DPUs.
• Turnkey, validated fabric — delivered cabled and tested as one system, not loose boxes.
• Validated with RDP compute & storage — part of a tested cluster.
• On-prem sovereignty — your fabric, your data path; DPDP-friendly, air-gappable.
• Make-in-India OEM — predictable INR pricing, GST tax invoice (HSN 8517), pan-India onsite support, GeM-procurable.

Where it fits

• GPU cluster fabric (primary): the interconnect that carries distributed-training collectives across the cluster.
• Storage fabric: high-bandwidth path between GPU nodes and the NVMe parallel-FS (GPUDirect).
• Scale-out spine: a non-blocking leaf-spine topology for large clusters and superclusters.
• HPC interconnect: tightly-coupled MPI and AI collectives.

How it works

The fabric forms a non-blocking leaf/spine topology with RDMA over InfiniBand; the subnet manager programs routes, in-network compute (SHARP) offloads reductions, and adaptive routing avoids hotspots — so all-reduce traffic, which dominates multi-node training, completes fast. RDP sizes the topology (oversubscription, spine count) to your cluster. Honest note: real scaling efficiency depends on topology, collective library and model — we validate it on your cluster, not just a peak number.

Industry use cases

• AI/ML & foundation-model teams — the fabric that lets training scale across a hall of GPUs.
• Government & national labs — sovereign HPC/AI interconnect at scale.
• Neocloud / AI providers — the fabric layer of a GPU cloud region.
• BFSI & research — low-latency clusters for training and HPC.
• Telecom — high-bandwidth AI fabric at scale.
• Manufacturing & energy — HPC + AI cluster interconnect.

Performance — and how to be sure

We don’t publish inflated peak numbers. The honest picture: 410 Tb/s aggregate non-blocking at Sub-600 ns is sized to carry cluster collectives without becoming the bottleneck. Want certainty? Request a free scaling test of your training job across nodes on this fabric before you buy; we’ll send back real all-reduce bandwidth and scaling efficiency.

Series & scale path

• DRACO (flagship fabric tier) — this.
• Speed ladder: 200G Ethernet → 400G NDR → 800G XDR InfiniBand.
• Scale: 16 → 32 → 64 → 128 → 256 → 512-node fabrics for thousands of GPUs.

On-prem vs cloud — the TCO case

An owned fabric is a one-time asset that removes the per-node networking premium of cloud GPU clusters and keeps the data path in-house. RDP pricing is fixed in INR with a GST input-credit-eligible invoice — ask for a fabric design and TCO for your cluster.

Software & integration

Integrates with the AI stack: the InfiniBand subnet manager, NCCL with SHARP and standard MPI, with fabric monitoring and telemetry. Validated with RDP GPU servers, rack-scale systems and storage.

Power, cooling & rack integration

A Leaf-spine fabric (multi-rack) with redundant PSUs and front-to-back airflow — plan rack power, cooling and cable plant (DAC/AOC/transceivers) across the topology. (Exact power draw, BTU and cabling confirmed in the design package.) Full out-of-band management.

Deployment, warranty & support

• Made to order, designed, cabled and validated in India as part of the cluster; project timeline confirmed at quote.
• Delivered as one fabric: switches, rails, cabling and the configured fabric profile; transceivers/cables scoped at quote.
• Onsite warranty + AMC with pan-India coverage and an RMA/escalation path (exact terms confirmed at quote).

Why RDP

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

Buy with confidence

This is a turnkey AI cluster fabric, made to order — talk to an RDP solution architect, design the topology and oversubscription for your cluster, get a fabric TCO, and run a scaling test before you commit. Request a quote to begin.

Specifications