The NVIDIA DGX B300 is not just a server; it is a fully integrated AI factory designed to solve the world’s most complex computing challenges. Built on the revolutionary NVIDIA Blackwell Ultra architecture, this system sets a new standard for AI infrastructure, combining unprecedented compute density, massive memory capacity, and lightning-fast networking into a single, production-ready platform. Whether you are training trillion-parameter models or deploying real-time reasoning engines, the DGX B300 provides the horsepower needed to turn raw data into intelligence.

Key Features & Performance

• Unmatched Compute Power: Equipped with 8x NVIDIA Blackwell Ultra GPUs, the DGX B300 delivers up to 144 petaFLOPS of FP4 AI inference performance and 72 petaFLOPS of FP8 training performance. This represents a generational leap, offering 11x faster inference and 4x faster training compared to the Hopper generation, drastically reducing time-to-solution for the most demanding AI workloads.

• Massive Memory Landscape: With a total of 2.3TB of high-bandwidth GPU memory (288GB per GPU), the system is engineered to handle massive datasets and the largest Mixture-of-Experts (MoE) models without memory bottlenecks. This expanded capacity allows developers to fit larger models into memory, speeding up processing and enabling more complex reasoning tasks.

• Unified High-Speed Fabric: The system features 8x NVIDIA ConnectX-8 VPI network adapters, delivering up to 800Gb/s of throughput per port. This advanced networking architecture ensures seamless scaling across clusters, making it the ideal building block for NVIDIA DGX SuperPOD™ and BasePOD™ deployments.

• Turnkey Enterprise AI: The DGX B300 comes fully integrated with the NVIDIA AI Enterprise software suite, providing a pre-optimized stack for generative AI, data science, and computer vision. From the operating system to the drivers and frameworks, every component is tuned for maximum performance and reliability, allowing your team to focus on innovation rather than infrastructure management.
  
  

Power & Cooling Infrastructure

• High-Density Power Architecture: The system is engineered for maximum performance with a robust power subsystem capable of handling a peak consumption of approximately 14kW per node. It supports flexible power configurations, including N+N redundancy with up to 12 power supply units (PSUs) or direct connection to rack busbars for streamlined data center integration.

• Thermal Management Options: To accommodate diverse data center environments, the platform is available in both air-cooled (8U) and liquid-cooled (4U) configurations. The liquid-cooled variant leverages direct-to-chip technology to capture significant heat loads, enabling higher rack densities and improved energy efficiency for large-scale deployments.
  
  

Software & Ecosystem

• NVIDIA Base Command & Mission Control: Integrated management software simplifies the orchestration of AI workloads, providing a unified view of cluster health, job scheduling, and resource utilization. NVIDIA Mission Control enhances operational agility by automating essential tasks and ensuring infrastructure resiliency for AI factories at scale.

• Full-Stack Optimization: The system ships with the NVIDIA DGX OS, which includes a tuned Linux kernel, optimized drivers, and the NVIDIA Container Toolkit. This pre-installed stack ensures out-of-the-box readiness for rapid deployment of containerized AI applications and seamless integration with MLOps workflows.