Supermicro, Inc. (NASDAQ: SMCI), a leading provider of total IT solutions for artificial intelligence, cloud, storage, and 5G/Edge, announced at the GTC 2025 conference its new rack systems and solutions based on the NVIDIA Blackwell Ultra platform, including the NVIDIA HGX™ B300 NVL16 and NVIDIA GB300 NVL72 platforms. These innovations reinforce Supermicro and NVIDIA’s position at the forefront of artificial intelligence, delivering breakthrough performance for the most demanding workloads, such as AI reasoning, AI agents, and large-scale video inference.
Optimized Solutions with Air and Liquid Cooling
The company has designed its new systems with thermal efficiency and internal optimization in mind, adapting both air-cooled and liquid-cooled configurations. Supermicro’s innovative liquid cooling solution operates with warm water at 40 °C in 8-node rack configurations or at 35 °C in 16-node double-density configurations, using the company’s latest CDU systems. This technology can reduce energy consumption by up to 40% and contributes to water resource conservation, providing both environmental and operational benefits for enterprise data centers.
Blackwell Ultra Platform: Unprecedented Memory and Connectivity
The NVIDIA Blackwell Ultra platform offers unprecedented capacity of 288 GB of HBM3e memory per GPU, providing significant increases in FLOPs for training and inference of large-scale AI models. Additionally, the integration of network platforms with NVIDIA Quantum-X800 InfiniBand and Spectrum-X™ Ethernet doubles the bandwidth of the compute network to 800 Gb/s, eliminating traditional bottlenecks in cluster environments.
Supermicro NVIDIA HGX B300 NVL16: The New Standard for Data Centers
The Supermicro NVIDIA HGX B300 NVL16 sets a new standard for AI training clusters, integrating a domain of 16 interconnected GPUs with 1.8 TB/s via NVLink™ and a total of 2.3 TB of HBM3e memory per system. The platform includes eight NVIDIA ConnectX®-8 NICs directly on the motherboard, offering speeds of up to 800 Gb/s for node-to-node communications via NVIDIA Quantum-X800 InfiniBand or Spectrum-X™ Ethernet.
Supermicro NVIDIA GB300 NVL72: Exascale Supercomputing in a Single Rack
The NVIDIA GB300 NVL72 system incorporates 72 NVIDIA Blackwell Ultra GPUs and 36 NVIDIA Grace™ CPUs, achieving exascale capability in a single rack. It offers over 20 TB of HBM3e memory, interconnected in a domain of 72 GPUs with 1.8 TB/s bandwidth. The connectivity, backed by NVIDIA ConnectX®-8 SuperNIC, provides 800 Gb/s speeds for GPU-to-NIC and NIC-to-network communication, enhancing cluster-level performance of the AI compute infrastructure.
Comprehensive Focus on Liquid Cooling and Data Center Deployment
Supermicro also stands out for its rapid deployment capabilities, supported by a complete portfolio of liquid cooling solutions that include direct-to-chip cold plates, 250 kW rack CDU systems, and cooling towers. The company offers on-site deployment for enterprises, managing the entire process from planning, design, and validation to installation and configuration of racks, servers, switches, and network equipment, adapting to the specific needs of each organization.
Innovations Showcased at GTC 2025
Attendees at GTC 2025 in San Jose, California, from March 17 to 21, can visit Supermicro’s booth #1115, where the X14/H14 B200, B300, and GB300 systems will be on display along with scalable liquid cooling rack solutions.