#USA #Supermicro #San Jose #NVIDIA Blackwell #HGX B300

Supermicro Unveils NVIDIA HGX B300 Systems

In a significant stride towards enhancing AI infrastructure, Super Micro Computer, Inc. has announced the expansion of its NVIDIA Blackwell portfolio with the introduction of new 4U and 2-OU liquid-cooled NVIDIA HGX B300 systems. These innovative solutions are tailored for high-density hyperscale and AI factory implementations, offering unprecedented performance and energy efficiency.

Meeting Rising AI Demands

As the global demand for AI infrastructure surges, Supermicro’s new liquid-cooled NVIDIA HGX B300 systems promise to deliver the computational density and energy efficiency essential for hyperscale data centers. Charles Liang, the company’s president and CEO, stated, “With our latest offerings, we provide the most compact NVIDIA HGX B300 solutions in the market, able to house up to 144 GPUs in a single rack. Our Direct Liquid Cooling (DLC) technology not only reduces power consumption but also enhances cooling efficiency.”

Technical Specifications

The 4U liquid-cooled NVIDIA HGX B300 systems are designed to fit standard 19-inch EIA racks and can support up to 64 GPUs per rack. Utilizing DLC-2 technology, these systems can capture up to 98% of the heat generated, resulting in remarkable energy efficiency and lower operational costs.

On the other hand, the 2-OU (OCP) NVIDIA HGX B300 system adheres to the 21-inch OCP Open Rack V3 (ORV3) specifications. It is designed to cater to cloud providers requiring space-efficient racks, with a staggering capacity of 144 GPUs in a single rack. This model incorporates blind-mate distribution connectors, a modular architecture for GPUs and CPUs, and state-of-the-art liquid cooling solutions. Specifically equipped with eight NVIDIA Blackwell Ultra GPUs, each with a TDP of 1100 W, it manages to keep both the rack footprint and power usage significantly lower.

Scalability and Flexibility

One ORV3 rack can support up to 18 nodes, totaling 144 GPUs, allowing for seamless scaling in large AI operations. The Supermicro solution is also compatible with NVIDIA Quantum-X800 InfiniBand switches and includes cooling distribution units (CDUs) capable of handling up to 1.8MW. When combined, eight NVIDIA HGX B300 compute racks, along with three NVIDIA Quantum-X800 InfiniBand network racks and two Supermicro in-row CDUs, form a scalable SuperCluster unit accommodating 1,152 GPUs.

In addition to the 2-OU model, Supermicro’s 4U Front I/O HGX B300 liquid-cooled system maintains similar computing power but utilizes a traditional 19-inch rack form factor, making it suitable for large-scale AI factory deployments. It leverages Supermicro’s advanced DLC-2 technology to maximize heat capture, promoting superior energy efficiency, reduced noise, and enhanced maintainability of dense training and inference clusters.

Advancements in Performance

The newly introduced Supermicro NVIDIA HGX B300 systems provide critical performance advancements. Each system is equipped with 2.1 TB of HBM3e GPU memory, enabling the processing of larger models at the system level. Both the 2-OU (OCP) and 4U platforms also significantly enhance cluster-level performance, doubling the compute fabric network throughput to 800 Gb/s when used in conjunction with NVIDIA's high-capacity networking equipment.

These achievements are poised to accelerate demanding AI workloads such as agentic AI applications, foundational model training, and large-scale multimodal inference in AI factories. Supermicro has tailored these offerings according to core customer demands for total cost of ownership (TCO), maintainability, and efficiency. With the DLC-2 technology stack, data centers can achieve energy savings of up to 40% while minimizing water usage by utilizing warm water at 45°C, therefore reducing the need for traditional coolant systems.

The Future of AI Infrastructure

Supermicro continues to deliver fully validated and tested racks as part of its Data Center Building Block Solutions (DCBBS), streamlining the deployment process for hyperscale, enterprise, and governmental customers. The expansion of the NVIDIA Blackwell portfolio now includes several robust solutions such as the NVIDIA GB300 NVL72, NVIDIA HGX B200, and NVIDIA RTX PRO 6000 Blackwell Server Edition. All Supermicro NVIDIA-certified systems have undergone rigorous testing to ensure optimal performance for various AI applications, ensuring compatibility with NVIDIA networking and software, including NVIDIA AI Enterprise and NVIDIA Runai.

As AI demands grow, Supermicro is committed to providing scalable AI infrastructures that can range from individual nodes to comprehensive AI factory ecosystems.