NVIDIA GB300 NVL72 Supercluster

Overview

A complete NVIDIA GB300 NVL72 Supercluster in a single, turnkey containerised node. Eight liquid-cooled GB300 NVL72 racks arrive pre-integrated as one production-ready AI factory node — 576 NVIDIA Blackwell Ultra B300 GPUs and 288 Grace CPUs operating as a single coherent accelerator, delivered to your site and commissioned as a plug-and-play unit rather than assembled on the floor over months.

Key highlights

• 8× GB300 NVL72 racks in one containerised node — 576 Blackwell Ultra B300 GPUs + 288 Grace CPUs as one system.
• ~165.6 TB HBM3e plus ~320 TB fast system memory as node-level unified memory.
• 1,040 TB/s aggregate NVLink bandwidth — near-zero-bottleneck all-to-all GPU communication.
• ~11.5 EFLOPS peak (FP4 sparse) and ~8.8 EFLOPS dense per node for trillion-parameter reasoning.
• Warm-water direct liquid cooling with in-row CDU rated up to 1.8 MW thermal dissipation.
• Redundant power: 64× 33 kW shelves with integrated busbars and comprehensive BMS/safety.
• High-speed data spine: ConnectX-8 (800 Gb/s) + BlueField-3 DPUs; Quantum-X800 InfiniBand or Spectrum-X Ethernet options.
• Complete stack: NVOS, full NVIDIA AI Enterprise (576 GPU subscriptions), Mission Control & DOCA — containerised and plug-and-play.

AI workload fit

• Large-scale / foundation-model pretraining at trillion-parameter scale.
• Post-training alignment and fine-tuning (SFT / RLHF) on frontier models.
• Real-time, test-time-scaling inference and serving.
• Agentic AI and multi-step reasoning workloads.
• Generative AI (text, image, video), NLP & speech model families.
• HPC + AI convergence and sovereign / national-scale AI.

AI workload positioning

This is a rack-scale-to-node building block for an AI factory. The balance of unified HBM3e capacity, 1,040 TB/s NVLink, and a high-speed ConnectX-8 / BlueField-3 data spine lets 576 GPUs train and serve models that will not fit on a single rack, while warm-water DLC and redundant 33 kW power shelves sustain the density in continuous production. It sits at the top of the DRACO tier — above a single GB300 NVL72 rack — and scales out to 16-, 32- and 64-rack superclusters.

Industry use cases

• Sovereign & public sector: data-resident national AI, on-soil foundation-model programs.
• Neocloud / AI cloud: multi-tenant training-and-inference capacity as a deployable node.
• Research & higher education: frontier-scale training for labs and institutions.
• Enterprise AI factories & GCCs: in-house model development and agentic platforms.
• Defence, telecom & government: secure, air-gap-capable large-scale AI.

Performance & how to be sure

Published node figures: ~11.5 EFLOPS peak (FP4 sparse) / ~8.8 EFLOPS dense, with a generational step over the Hopper baseline of ~50×+ AI-factory output, ~10× user responsiveness and ~5× throughput-per-watt. Rather than quote a single tokens/sec number that will not match your models, RDP offers a “benchmark your model” POC: bring your training run or inference workload and we will size and validate it against this configuration before you commit.

Series & upgrade path

DRACO is the flagship tier. The GB300 family scales: a single GB300 NVL72 rack-scale AI factory → this 8-rack containerised Supercluster (576 GPUs) → 16 / 32 / 64-rack superclusters. Start at the node that matches your model scale and add nodes as demand grows.

On-prem vs cloud (TCO)

For sustained frontier training and inference, an owned node removes per-GPU-hour cloud billing, data-egress fees and multi-tenant contention, and keeps data on-soil for DPDP / sovereignty requirements. RDP provides predictable INR capital pricing, GST input-credit invoicing (HSN 8471), and financing / lease options to compare against a 3-year cloud run.

Software & day-one readiness

Ships with NVOS managing the hardware layer, full NVIDIA AI Enterprise licensing across all 576 GPUs, and Mission Control + DOCA fleet orchestration. CUDA, cuDNN, drivers and container runtime are pre-integrated so the node is workload-ready on power-up.

Serving stack, ready on day one. The node ships with the two engines the market actually runs: vLLM — the broad production standard with the widest hardware support — and SGLang, whose RadixAttention is strongest on RAG and multi-turn workloads. SGLang publishes day-0 support for new open models and reports up to ~25× inference throughput on GB300 NVL72 (SGLang’s own published figure, on their reference configuration — we will benchmark your model on this hardware rather than ask you to take it on trust). Also validated for FP8 serving, now the precision most frontier models ship in at launch, and sized for trillion-parameter MoE models such as Qwen3.8-2.4T-A95B.

Power, thermal & acoustics

Housed in a high-cube container: 64× 33 kW power shelves with redundant busbars, warm-water direct liquid cooling via an in-row CDU rated to ~1.8 MW. Site needs grid + network hookup and a warm-water loop; the container isolates acoustics and thermal from occupied space.

Deployment, warranty & support

Arrives pre-integrated and containerised, bypassing years of traditional data-center construction; plug-and-play for immediate grid and network hookup, with comprehensive training and detailed O&M documentation. Backed by full NVIDIA Enterprise support and RDP pan-India onsite service, GST invoice, and GeM availability. Lead time: 12–16 weeks. This is premium, constrained inventory — allocation is prioritised for strategic partners.

Why RDP

RDP Technologies is a Make-in-India OEM with 14+ years and 300,000+ units shipped, delivering sovereign / DPDP-ready AI infrastructure with predictable INR pricing, GST invoicing and pan-India onsite SLAs.

Full node configuration (bill of materials)

One node = 8× GB300 NVL72 racks. Per-node totals:

• Compute: 144× 1U liquid-cooled compute trays · 288× GB300 Grace-Blackwell Ultra superchips · 576× Blackwell Ultra B300 GPUs (288 GB HBM3e each; FP4/FP6/FP8 Tensor Cores) · 288× Grace CPUs (72-core Arm Neoverse V2, up to 480 GB LPDDR5X each).
• NVLink scale-up: 72× 5th-gen NVLink switch trays · 18 NVLink-5 links per GPU · 1.8 TB/s per GPU · 1,040 TB/s aggregate per node.
• In-rack networking: 576× ConnectX-8 SuperNICs (dual-port 800 Gb/s) · 144× BlueField-3 B3240 DPUs · 16× SN2201 OOB management switches.
• Scale-out fabric: NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet leaf/spine at 800 Gb/s per GPU (selected at design stage), with pre-terminated optical trunks.
• Storage: 1,152× E1.S 3.84 TB NVMe (PCIe Gen5) + 144× M.2 1.92 TB boot in-tray; external NVIDIA Enterprise RA-certified storage sized to workload.
• Power: 64× 33 kW power shelves · 48V DC busbars · IT load ~1.06–1.08 MW nominal, ~1.24 MW peak (EDPp), busway provisioned to ~1.54 MW.
• Cooling: warm-water direct liquid cooling · in-row CDU up to 1.8 MW with N+1 pumps · blind-mate manifolds and leak detection.
• Enclosure & safety: high-cube containerised module housing all 8 racks + CDU · clean-agent fire detection/suppression · environmental monitoring/BMS · physical access control.
• Software: NVIDIA AI Enterprise (576 GPU subscriptions) · Mission Control · NVOS · DOCA services on BlueField-3.
• Programme services: logistics, staging & commissioning · site integration and liquid-loop commissioning · burn-in & acceptance testing · operator training with as-built O&M documentation.

Sustained performance: ~8.6 EFLOPS FP4 sustained per node (peak ~11.5 EFLOPS sparse / ~8.8 EFLOPS dense).

Buy with confidence

Use Request a Quote to reach a named RDP solution architect for site assessment, benchmark-your-model POC, financing options and a delivery plan. No obligation.

Specifications