DRACO 8 PB Parallel-FS NVMe AI Storage

Overview

The DRACO 8 PB Parallel-FS NVMe AI Storage is a large scale-out parallel-filesystem NVMe storage system built to feed entire GPU superclusters during the largest AI training runs. It delivers 8 PB usable (NVMe) at 4.8 TB/s read with GPUDirect Storage across a clustered namespace, so datasets, checkpoints and weights stream to hundreds of GPU nodes in parallel without the storage becoming the bottleneck — on-premises, in INR, on a GST invoice.

Engineered for supercluster-scale AI/HPC data pipelines, it presents a single parallel namespace over high-speed networking and scales capacity and bandwidth together by adding nodes — delivered racked, configured and validated as one system with one warranty and one support contract.

Key highlights

• 8 PB usable (NVMe) · 4.8 TB/s read — supercluster-scale bandwidth sized to feed hundreds of GPU nodes.
• Parallel filesystem + GPUDirect Storage — a single namespace; data streams from NVMe straight to GPU memory across the cluster.
• Parallel FS (Lustre/GPFS-class) + NFS/S3, GPUDirect Storage — parallel and standard protocols so existing pipelines and schedulers just work.
• 768× 15.36 TB NVMe (32 nodes) — dense enterprise NVMe with end-to-end data integrity, scaling across nodes.
• 128× 400G InfiniBand/Ethernet — very high aggregate bandwidth to the GPU fabric; bandwidth grows with capacity.
• 320M random read — high aggregate random-read IOPS for metadata- and small-file-heavy datasets.
• On-prem data sovereignty — datasets 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.

Where it fits

• Supercluster training data lake (primary): streams datasets to a large GPU cluster in parallel without starving any node.
• Checkpoints & weights: fast parallel write/read of very large checkpoints during long runs.
• RAG & vector stores: low-latency storage for very large embeddings and indexes.
• HPC scratch: high-throughput scratch for large simulation alongside AI.

How it works

A clustered parallel filesystem stripes data across many NVMe nodes and presents one namespace over 128× 400G InfiniBand/Ethernet. With GPUDirect Storage, reads bypass the CPU and land directly in GPU memory across the cluster. Capacity and bandwidth scale together as nodes are added. Honest note: real throughput depends on dataset shape, file sizes and the client fabric — we validate it on your data, not just a peak number.

Industry use cases

• Government & national labs — sovereign data lakes for national AI and HPC.
• AI/ML & foundation-model teams — feed GPU clusters at full speed.
• BFSI & healthcare — data-resident storage under compliance.
• Media & design — fast scratch and asset storage.
• Energy & manufacturing — large datasets for AI/HPC.
• Neocloud / AI providers — the high-bandwidth storage tier for a GPU cloud.

Performance — and how to be sure

We don’t publish inflated peak numbers. The honest picture: 8 PB usable (NVMe) at 4.8 TB/s read. Want certainty? Request a free benchmark with your datasets and training loop on this exact configuration before you buy; we’ll send back real sustained throughput, IOPS and GPU-utilisation.

Series & scale path

• DRACO (flagship storage tier) — this.
• Tiering: NVMe parallel-FS hot tier (GPU feeding) + object capacity tier (durable datasets/archives).
• When to step up: add nodes for more capacity and bandwidth — talk to an architect about the tiering.

On-prem vs cloud — the TCO case

For AI data at scale, owning beats renting: cloud egress and per-GB costs add up fast against large datasets, and on-prem keeps data resident. RDP pricing is fixed in INR with a GST input-credit-eligible invoice — ask for a 3-year TCO comparison including egress savings.

Software & integration

Integrates with your stack: Parallel FS (Lustre/GPFS-class) + NFS/S3, GPUDirect Storage, NVIDIA GPUDirect Storage, and standard NFS/S3 clients, with monitoring and quota/multi-tenant management. Works with PyTorch/TensorFlow data loaders, Slurm/Kubernetes and common MLOps tooling.

Power, cooling & rack integration

A Multi-rack (32-node) liquid-cooled system with redundant PSUs — specify rack power and cooling. (Exact power draw, BTU, airflow and footprint figures confirmed on the build sheet.) Full out-of-band management and drive hot-swap.

Deployment, warranty & support

• Made to order, built, racked and burned-in in India; lead time confirmed at quote.
• In the box: storage nodes, drives, rails, cabling, quick-start, and the configured parallel filesystem.
• Onsite warranty + AMC with pan-India coverage, drive-replacement SLA 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 (HSN 8471), pan-India onsite engineers, GeM availability, and DPDP / sovereign-AI-ready deployment.

Buy with confidence

This is an AI storage system, made to order — talk to an RDP solution architect, size capacity, bandwidth and fabric for your AI estate, get a 3-year TCO, and benchmark your own datasets before you commit. Request a quote to begin.

Specifications