Uvation MarketplaceMarketplace

NVIDIA DGX SuperPOD with DGX B300 Systems — Spectrum-4 Ethernet, DC Busbar

NVIDIA DGX SuperPOD with DGX B300 Systems integrates Blackwell Ultra GPUs, high-speed networking, storage, DC busbar power, and centralized management for large-scale AI and HPC workloads.

32×
NVIDIA Blackwell Ultra GPUs per compute rack
8×
Intel Xeon Platinum 6776P CPUs per compute rack
8 TB
Standard system memory per compute rack (up to 16 TB)

Integrated Infrastructure for Large-Scale AI

Built around modular 64-system Scalable Units, NVIDIA DGX SuperPOD combines accelerated computing, high-speed Ethernet and InfiniBand fabrics, certified storage, and centralized management in a validated AI infrastructure architecture.

Accelerated Computing

Each DGX B300 system features eight NVIDIA Blackwell Ultra GPUs, fifth-generation NVIDIA NVLink, and fourth-generation NVIDIA NVSwitch technology.

High-Speed Ethernet Fabric

NVIDIA Spectrum-X uses Spectrum-4 Ethernet switches and ConnectX-8 SmartNICs to deliver high-performance, low-latency connectivity for distributed AI workloads.

Unified Infrastructure Management

NVIDIA Mission Control manages cluster operations and workload orchestration; NVIDIA NetQ provides Spectrum-X network telemetry; NVIDIA Unified Fabric Manager manages the InfiniBand fabric.

64 DGX B300 Nodes. 512 Blackwell Ultra GPUs. Rack-Scale DC Architecture.

A reference Scalable Unit (SU) integrating 64 DGX B300 nodes into MGX racks with DC busbar power infrastructure and Spectrum-4 Ethernet networking for massive AI scale-out.

NVIDIA DGX SuperPOD with DGX B300 Systems rack architecture
512 GPUs
Reference Scalable Unit across 64 DGX B300
compute nodes
133.12 TB GPU Memory
2.1 TB HBM3e GPU memory per compute node
(288 GB per GPU)
Spectrum-4 Ethernet
High-throughput scale-out fabric with ConnectX-8
NICs & BlueField-3 DPUs
DC Busbar Power
4× 33 kW power shelves per MGX compute rack

Workloads

Built for massive-scale enterprise AI infrastructure, frontier LLM training, real-time reasoning, and supercomputing workloads.

Frontier Large Language
Model Training

512 Blackwell Ultra GPUs linked across a Spectrum-4 high-bandwidth network deliver multi-petaflop compute density to train massive generative models.

Enterprise AI Factory and
Multi-Tenant Inference

Dual BlueField-3 DPUs per node provide hardware-accelerated infrastructure services, security, and multi-tenant isolation for high-throughput inference services.

High-Performance Computing and Analytics

Dual Intel Xeon Platinum 6776P CPUs with up to 4 TB of system RAM per node feed continuous data streams to GPU acceleration clusters.

Engineered to dominate

Rack-Scale Ethernet Supercomputing Power

This SuperPOD reference configuration combines 512 Blackwell Ultra GPUs, fifth-generation NVLink interconnect, and Spectrum-4 Ethernet networking with DC busbar power infrastructure to deliver peak compute efficiency for enterprise AI factories.

NVIDIA DGX SuperPOD with DGX B300 Systems racks
144 PFLOPS
FP4 sparse (108 dense) per node / ~9.2 EFLOPS FP4 aggregate (64 nodes)
72 PFLOPS
FP8/FP6 sparse compute per node / ~4.6 EFLOPS FP8 aggregate (64 nodes)
1.8 TB/s
NVLink 5 bandwidth per GPU (14.4 TB/s aggregate per DGX B300 node)
800 Gb/s
Spectrum-4 Ethernet networking via 8× ConnectX-8 ports & 2× BlueField-3 DPUs per node
NVIDIA DGX SuperPOD with DGX B300 Systems racks
144 PFLOPS
FP4 sparse (108 dense) per node / ~9.2 EFLOPS FP4 aggregate (64 nodes)
72 PFLOPS
FP8/FP6 sparse compute per node / ~4.6 EFLOPS FP8 aggregate (64 nodes)
1.8 TB/s
NVLink 5 bandwidth per GPU (14.4 TB/s aggregate per DGX B300 node)
800 Gb/s
Spectrum-4 Ethernet networking via 8× ConnectX-8 ports & 2× BlueField-3 DPUs per node

The rack, itemized.

SUBSYSTEM

SPECIFICATION / CONFIGURATION DETAILS

Processor

2 × Intel Xeon 6776P per node

Operating System

NVIDIA DGX OS

Chipset

NA

Accelerators / GPU

8× Blackwell Ultra GPUs per node; reference SU: 512 GPUs

Memory

Per compute node: 2 TB standard, up to 4 TB system memory

Primary Storage

Per node: 8 × 3.84 TB E1.S NVMe

Secondary Storage

Per DGX B300: 2x 1.92 TB M.2 NVMe boot; 8x 3.84 TB E1.S NVMe cache

Security

DGX B300 BMC; cluster security configuration dependent

Chassis

MGX racks; reference layout: 4 DGX B300 nodes per compute rack

Power Supply

DC busbar; reference layout: 4× 33 kW power shelves per compute rack

Ports

Per DGX B300 node: 8 × OSFP, 2 × dual-port QSFP112, 1GbE management; Spectrum-4 Ethernet switching

Slots / PCIe Architecture

Per node: 2 NVLink switches; network fabric determined by reference architecture

Bezel

DGX B300 front bezel

Form Factor

Reference SU: 64 DGX B300 nodes

Network Card

Per compute node: 8× ConnectX-8; 2× dual-port BlueField-3

Warranty

DGX nodes: 3-year business-standard hardware and software support; complete solution coverage per contract.

NVIDIA DGX SuperPOD with DGX B300 Systems power and networking

Power Fabric

4× 33 kW power shelves

DC busbar distribution to feed 4 DGX B300 compute nodes.

32× ConnectX-8 at 800 Gb/s

OSFP ports per Blackwell Ultra GPU for inter-rack cluster networking.

8× dual-port BlueField-3 DPUs

For storage and in-band fabric, alongside dedicated RJ45 OOB management.

Nuclear-Ready, Plug and Play AI Factories

Bypass grid constraints with self-contained, high-density inference ready infrastructure directly adjacent to clean, base load nuclear power.

Global Deployment Engine

Any Infra, Anywhere in the world.

Global Shipping and Deployment Capability treats advanced infrastructure as mobile, modular freight.

Packaged complex computing nodes, cooling manifolds, and power distribution units into certified, standard form factors.

360° Service Ecosystem

Unifying Legacy with Next-Gen Compute.

Vertically integrated service ecosystem that serves as an operational bridge between legacy environments and next-generation, AI-native infrastructure.

Finance and Leasing

Capital-as-a-Service for High-Density Compute

Convert upfront capital expenditures (CapEx) into predictable operating expenses (OpEx).

Uvation allows enterprises and sovereign entities to deploy multi-megawatt AI factories and advanced GPU clusters without liquidity constraint.