Supermicro has begun the global rollout of its NVIDIA Blackwell Ultra solutions, marking a significant milestone in the company’s long-standing collaboration with NVIDIA. The announcement centers on the broad availability of Plug-and-Play (PnP)-ready NVIDIA HGX B300 systems and GB300 NVL72 racks, pre-engineered and validated at system, rack, and even full data center scale.
These turnkey solutions are targeted at enterprises building the next generation of AI factories, where massive computational performance and efficiency are critical to training, inference, and deployment of advanced artificial intelligence models.
Charles Liang, president and CEO of Supermicro, highlighted the importance of these developments for data center operators struggling with the complexity of scaling AI infrastructure. “Supermicro has the best track record of fast and successful deployments of new NVIDIA technologies,” he said, adding that the company’s Data Center Building Block Solutions and on-site deployment expertise make it possible to deliver “the highest-performance AI platform” with reduced time-to-market. According to Liang, the challenges of cabling, power distribution, and cooling are growing as AI clusters expand, and pre-validated, plug-and-play systems represent a crucial advantage for organizations racing to deploy large-scale AI capacity.
The Blackwell Ultra platform brings significant generational advancements. At the system level, Supermicro integrates advanced air and direct liquid cooling designs optimized for GPUs capable of consuming up to 1,400 watts each. Compared to the previous Blackwell generation, the Ultra version delivers 50% greater inferencing performance using FP4 compute, alongside 50% more HBM3e memory capacity. These enhancements are aimed at running larger models with higher efficiency, a necessity as AI workloads push into trillions of parameters.
Supermicro’s NVIDIA Blackwell Ultra portfolio spans a wide range of configurations, including rack-scale GB300 NVL72 systems and 8U air-cooled or 4U liquid-cooled HGX B300 servers. The GB300 NVL72 rack-scale platform alone delivers 1.1 exaFLOPS of dense FP4 compute performance. Meanwhile, individual B300 systems provide up to 144 petaFLOPS of FP4 performance and 270 GB of HBM3e memory per GPU, representing up to 7.5x the performance of systems based on NVIDIA’s Hopper architecture.
Networking forms a key part of these solutions. Enterprises can choose between NVIDIA Quantum-X800 InfiniBand and Spectrum-X Ethernet fabrics, each offering bandwidth up to 800 Gb/s, ensuring Blackwell Ultra systems can be interconnected into large-scale clusters without bottlenecks. Integration of NVIDIA’s ConnectX-8 SuperNICs further boosts network throughput, doubling compute network bandwidth compared to earlier platforms.
NVIDIA AI Enterprise Software
Supermicro positions these systems not only as hardware but as part of a full-stack ecosystem. Each deployment is integrated with NVIDIA AI Enterprise software, along with blueprints and NIM microservices designed to accelerate AI workloads. Beyond hardware and software, the company also provides extensive deployment services through its Data Center Building Block Solutions (DCBBS). This includes on-site cabling, thermal management, and power integration, ensuring faster time-to-online for customers building AI factories.
Sustainability is also an integral part of the pitch. Supermicro claims that its DLC-2 liquid cooling technology can save up to 40% in power consumption, reduce data center footprint by 60%, and cut water usage by 40%. Collectively, these improvements can lower total cost of ownership by as much as 20%, a key factor for enterprises facing mounting energy costs and sustainability requirements.
The availability of Blackwell Ultra solutions underscores the rapid evolution of infrastructure designed for AI. With pre-validated, scalable systems capable of exascale performance, Supermicro aims to simplify the complexity of deploying AI at scale, while enabling enterprises to future-proof their operations for the demands of multimodal models, agentic AI, and real-time inference.
Discover more from WIREDGORILLA
Subscribe to get the latest posts sent to your email.
