DGX B300 Core Computing Architecture

FAQs

• What is the NVIDIA DGX B300, and what scale of performance does it offer for enterprise AI?

  The NVIDIA DGX B300 is a high-performance AI infrastructure platform designed to serve as a universal foundation for enterprises building large-scale AI capabilities. Its power is best demonstrated by its deployment in the world’s first NVIDIA DGX SuperPOD, which integrates 1,016 Blackwell Ultra GPUs to perform over 9 quintillion calculations per second. This architecture is engineered to accelerate complex computations and provide a breakthrough in operational efficiency for modern “AI factories”.

• What is the "Blackwell Ultra" architecture that powers the DGX B300’s compute engine?

  At the core of each DGX B300 system are eight NVIDIA Blackwell Ultra (B300) GPUs that function as a tightly coupled compute complex. Each B300 GPU features a dual-reticle design with 208 billion transistors across two silicon dies, connected via a High-Bandwidth Interface (NV-HBI) delivering 10 TB/s of on-package bandwidth. Crucially, this design is transparent to software, appearing as a single logical GPU within the CUDA programming model to simplify scheduling and memory management.

• How does the DGX B300 optimise performance specifically for AI inference and reasoning tasks?

  The system introduces the NVFP4 precision format, a low-precision format optimised for transformer-based models that delivers up to 15 petaFLOPS of dense compute per GPU. This format reduces memory usage by approximately 1.8× compared to FP8, allowing more model parameters to remain in-memory for higher throughput. Additionally, the architecture doubles the throughput of Special Function Units (SFUs), resulting in up to 2× faster performance for attention layers, which is critical for reasoning-centric models with long context windows.

• What memory architecture is used to support these massive AI models without performance stalls?

  To ensure compute engines remain fully utilised, the DGX B300 features a unified HBM3e memory subsystem. Each system provides 2.3 TB of total GPU memory, with 288 GB of HBM3e per GPU—a 3.6× increase over the previous H100 generation. This high-capacity memory is paired with a bandwidth of up to 8 TB/s per GPU, allowing models with over 300 billion parameters to remain resident in GPU memory and preventing data transfer bottlenecks during real-time inference or training.

• How do the individual GPUs within the system communicate to act as a single unit?

  The eight Blackwell Ultra GPUs are linked using fifth-generation NVIDIA NVLink, which provides 1.8 TB/s of bidirectional bandwidth per GPU. This high-speed intra-system interconnect enables seamless memory sharing and distributed computation across the entire set of GPUs. This allows complex workloads, such as generative AI and long-context reasoning, to treat the system’s total resources as a single logical unit, eliminating internal bottlenecks during data movement.

• How does the DGX B300 scale beyond a single system to form larger clusters?

  For “scale-out” performance, the DGX B300 is equipped with robust networking hardware, including eight OSFP ports supporting up to 800 Gb/s InfiniBand or Ethernet via NVIDIA ConnectX-8 SuperNICs. It also incorporates two dual-port NVIDIA BlueField-3 DPUs to handle storage acceleration, infrastructure management, and security isolation. This combination ensures low-latency, high-throughput communication when the systems are deployed as part of a multi-node SuperPOD.

• What are the physical and mechanical specifications of the DGX B300 chassis?

  The system is housed in a 10 Rack Unit (RU) chassis designed for high compute density and ease of serviceability. It features front-accessible I/O to simplify cabling and maintenance, while cooling fans are located at the rear to allow for thermal management without system downtime. In terms of power, the system consumes approximately 14.5 kW under full load and is compatible with both AC/PDU and DC/busbar configurations to fit various data centre architectures.

• What supporting hardware is included to manage the system’s orchestration and data processing?

  Beyond the GPUs, the DGX B300 includes two Intel Xeon Platinum 6776P CPUs to handle orchestration, preprocessing, and auxiliary workloads. The system is also equipped with 2 TB of DDR5 system memory (expandable to 4 TB), ensuring that end-to-end AI pipelines—from initial data ingestion to final inference—can operate smoothly without being limited by the supporting hardware layers.

• How can an organisation prepare for the deployment and optimisation of a DGX B300 system?

  Successful deployment requires meticulous planning regarding power, cooling, and networking alignment to avoid underutilisation. The Uvation Marketplace serves as a central platform for organisations to assess data centre readiness, plan integration with existing infrastructure, and access advisory support. Experts are available through this marketplace to help align DGX B300 strategy with long-term AI goals and specific workload requirements.

  To understand the DGX B300, imagine a high-performance racing team: the Blackwell Ultra GPUs are the powerful engines, the HBM3e memory is the high-octane fuel delivered at massive speeds, and the NVLink is the precision transmission that ensures every part of the vehicle works in perfect synchronisation to achieve maximum velocity.