Artificial intelligence is growing faster than ever, and with it comes the need for infrastructure capable of supporting massive training clusters, real-time reasoning, and multimodal AI applications. That’s where Supermicro’s NVIDIA HGX™ B300 Systems, powered by the NVIDIA Blackwell Ultra architecture, step in.

These systems are designed to deliver ultra-performance computing for organizations pushing the boundaries of AI. With support for both air-cooled and liquid-cooled configurations, they provide flexibility, scalability, and unmatched performance.

Why the B300 Systems Matter

• Up to 7.5x performance gains over the previous NVIDIA Hopper generation.

• 288GB of HBM3e memory per GPU, ensuring enough bandwidth and memory capacity to handle the largest models.

• Support for scaling from single systems to 72-node clusters with thousands of GPUs.

The NVIDIA HGX B300 platform is a building block for the world’s largest AI training clusters. It is optimized for delivering the immense computational output required for today’s transformative AI applications.

Some key advantages include:

This combination means businesses and research institutions can train larger models faster, deploy more responsive AI, and handle workloads that were previously unthinkable.

The System Configurations

Supermicro offers two primary system designs for the B300 platform—an air-cooled 8U and a liquid-cooled 4U version (coming soon). Each is optimized for different deployment needs.

• Air-Cooled 8U System
• Processors: Dual Intel® Xeon® CPUs (5th Gen Scalable processors)
• GPUs: 8x NVIDIA Blackwell B300 GPUs with NVSwitch connectivity
• Memory: Up to 8TB DDR5 across 24 DIMM slots
• Storage: Up to 32 NVMe drives for high-speed data access
• Networking: Dual port 400GbE/IB + OCP slots
• Power: 6x 6000W redundant (N+1) Titanium level power supplies

This setup is perfect for organizations that prefer traditional air-cooled infrastructure while still delivering top-tier GPU density and performance.

Liquid-Cooled 4U System (Coming Soon)

• Processors: Dual Intel® Xeon® CPUs
• GPUs: 8x NVIDIA Blackwell B300 GPUs
• Memory: Up to 4TB DDR5 across 16 DIMM slots
• Storage: 16 NVMe drives for fast local storage
• Networking: Dual 400GbE/IB + OCP slots
• Cooling: Supermicro 250kW capacity CDU (Cooling Distribution Unit) with hot-swappable pumps
• Power: Redundant PSU design

The liquid-cooled option is designed for maximum efficiency and density, ideal for data centers seeking reduced operational costs and improved cooling at scale.

Scaling Beyond a Single System

Supermicro doesn’t stop at standalone servers. The B300 systems are available in rack-level and cluster-level solutions, giving enterprises the ability to scale to thousands of GPUs.

Air-Cooled Rack

• Up to 32x NVIDIA B300 GPUs per rack
• 9.2TB of HBM3e memory per rack
• NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet networking
• Out-of-band 1G/10G IPMI switch for management

This option provides a non-blocking, air-cooled network fabric, suitable for organizations with existing air-cooled infrastructure.

Liquid-Cooled Rack

• Up to 64x NVIDIA B300 GPUs per rack
• 18.4TB of HBM3e memory per rack
• Flexible storage fabric with full NVIDIA GPUDirect RDMA support
• Vertical Cooling Distribution Manifold (CDM) for efficient cooling

This is the next step in efficiency and density, making it ideal for high-performance AI clusters where space and power optimization are critical.

Scaling to Clusters: 72-Node Deployments

For organizations training the largest AI models, Supermicro offers fully integrated 72-node clusters.

• Air-Cooled 72-Node Cluster: Up to 576 NVIDIA B300 GPUs
• Liquid-Cooled 72-Node Cluster: Same GPU density, but with liquid cooling for even higher performance efficiency

Each cluster is pre-integrated with NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet fabric, delivering up to 800Gb/s per link. These are ready-to-deploy solutions built for enterprises that need to train trillion-parameter AI models.

Why Enterprises Should Care

AI models are rapidly expanding in both size and complexity. To remain competitive, enterprises need infrastructure that:

• Scales seamlessly as workloads grow
• Handles trillions of parameters without bottlenecks
• Offers flexibility between air-cooled and liquid-cooled designs
• Maximizes efficiency per watt and per square foot

Supermicro’s NVIDIA B300 systems deliver all of this, empowering organizations to stay at the forefront of AI innovation.

Final Thoughts

The Supermicro NVIDIA HGX B300 systems are more than just servers—they’re the foundation for next-generation AI. With industry-leading performance, scalability, and efficiency, these solutions are built for the future of AI training, inference, and deployment at massive scale.

Whether you’re starting with a single 8-GPU system or scaling up to a 72-node cluster, the B300 platform ensures you have the infrastructure to handle what’s coming next in AI.