Inside NVIDIA DGX B300: Architecture, Hardware Specs, and Data Center Engineering

The rapid expansion of enterprise artificial intelligence adoption requires computing infrastructure capable of managing massive scaling demands while keeping operational complexity and energy consumption under control. Introduced at GTC 2025 in San Jose, California, the NVIDIA DGX B300 system is a re-engineered enterprise AI platform designed for high-performance training and inference workloads.

Enclosed in an air-cooled 10 Rack Unit (RU) chassis, the DGX B300 integrates eight NVIDIA Blackwell Ultra GPUs, 5th-generation NVLink switches, PCIe Gen6 network connectivity, and modular power infrastructure. This technical brief details the core physical, electrical, and compute innovations introduced in the DGX B300 platform.

Technical Specifications Summary

Redesigned 10 RU Chassis and Front Accessibility

The DGX B300 introduces a new structural layout compared to previous generations, optimized for denser deployments and varied data center topologies.

While retaining the signature gold foam bezel of the NVIDIA DGX platform, the DGX B300 bezel does not cover the entire front face. This design change enables network cabling, storage, and maintenance interfaces to be routed and serviced directly from the front of the node.

The lower section of the front chassis provides direct access to:

• Cluster Networking: Direct front-facing OSFP network connections.
• Storage Drives: Front-accessible drives for local data caching and system booting.
• Management Ports: Dedicated I/O ports, including MiniDP video, USB, LAN, Baseboard Management Controller (BMC) port, serial port, power button, UID button, and BMC reset button for direct crash-cart access.
• Slide-Out Motherboard Tray: Mounted on sliding rails, the motherboard tray pulls out from the chassis to facilitate maintenance of internal components such as DIMM modules and security controllers.

GPU Computing and High-Throughput Interconnects

The computing core of the DGX B300 resides in a modular, sliding GPU tray housing eight Blackwell Ultra GPUs.

5th-Generation NVLink Architecture

The eight Blackwell Ultra GPUs are interconnected via two onboard NVSwitches running 5th-generation NVLink. This fabric delivers 14.4 TB/s of aggregate bandwidth, doubling the bandwidth available in previous Hopper-generation systems. This high-speed mesh ensures minimal latency during multi-GPU model parallelization.

Direct PCIe Gen6 Networking

A major architectural update in the DGX B300 is the relocation of the network interfaces directly into the GPU tray. The system incorporates eight 800 Gbps NVIDIA ConnectX-8 OSFP network ports.

In conventional server designs, network adapters may connect through the host CPU’s PCIe hierarchy, introducing additional I/O hops and potentially constraining the path between GPUs and network interfaces. DGX B300 instead places the ConnectX-8 interfaces in the GPU tray and connects them directly to the GPUs over PCIe Gen6. The DGX B300 connects the ConnectX-8 interfaces directly to the GPUs using PCIe Gen6, doubling the bandwidth to achieve the full 800 Gbps line rate. This maximizes throughput and accelerates RDMA transactions across DGX clusters. The ConnectX-8 interfaces can be configured in either NVIDIA InfiniBand or Ethernet mode to integrate with existing networking environments.

Motherboard, Memory, and Storage Architecture

The slide-out motherboard tray contains the host compute, local memory, storage caching, and management infrastructure:

• CPUs & Memory: Houses two Intel Xeon Platinum 6776P CPUs paired with 2 TB of system memory (32 × 64 GB DIMMs), which can be upgraded to 4 TB (32 × 128 GB DIMMs).
• DPUs: Features two NVIDIA BlueField-3 DPUs to manage system communication, isolation, and telemetry offloading.
• High-Speed Cache: Eight E1.S NVMe SSDs provide 30 TB of local cache storage to support intensive training and data retrieval pipelines.
• Boot Media: Two 1.92 TB M.2 NVMe boot drives sit inside an OCP carrier connected to PCIe, making them fully serviceable from the front panel.
• Security Subsystem: Includes a Data Center Secure Control Module (DC-SCM) containing the BMC, Trusted Platform Module (TPM), and management ports.

Power Infrastructure: AC PSUs vs. DC Busbars

The DGX B300 offers two distinct power delivery mechanisms to match differing data center electrical layouts.

1. Traditional AC Power Configuration

• Designed for facilities distributing 200–240V AC to the rack.
• Features twelve integrated power supplies that are replaceable from the front of the chassis.
• Utilizes twelve power cords connected at the rear to distribute the electrical load evenly across three phases.
• The system requires six active power supplies to operate at maximum performance, providing full N+N redundancy.

2. High-Density DC Busbar Configuration (NVIDIA MGX)

• Designed for modern data centers delivering 400–480V AC directly to the rack.
• Eliminates internal server power supplies; instead, a single DC power clip at the rear connects to a vertical rack busbar.
• Deployed in NVIDIA MGX-compliant racks equipped with stiffeners and dedicated power shelves.
• Power shelves (containing six 5.5 kW power supplies and a management module) convert AC to 54V DC at the rack level for all installed systems.
• Reduces intra-chassis heat and conversion losses while allowing operators to set redundancy levels up to full N+N based on local maintenance needs.

Operational Takeaways

The NVIDIA DGX B300 addresses enterprise AI scaling challenges within an air-cooled, 10 RU server footprint. By integrating eight Blackwell Ultra GPUs, 14.4 TB/s NVLink fabric, direct PCIe Gen6 800 Gbps networking, front-serviceable components, and flexible AC/DC power options, the platform provides scalable, high-throughput compute designed to integrate smoothly into traditional and next-generation data centers.