Super Micro Computer, Inc. (NASDAQ: SMCI), a global leader in enterprise computing, today announced a portfolio of superclusters to accelerate the adoption of generative AI.

The three supercluster solutions can be utilized for generative AI workloads, and can be used for powerful LLM learning performance as well as large-scale batch and large-capacity LLM inference by choosing a 4U water-cooled cooling system or an 8U air-cooled cooling system. Superclusters featuring 1U air-cooled Supermicro NVIDIA MGXTM systems are optimized for cloud-scale inference.

“In the era of AI, compute is measured not simply in number of servers, but in clusters,” said Charles Liang, President and CEO of Supermicro. “With our global manufacturing capacity scaling to 5,000 racks per month, Supermicro can deliver full-fledged, generative AI clusters faster than ever before.”

“The scalable cluster building block is a 64-node cluster with 400Gb/s NVIDIA Quantum-2 InfiniBand and Spectrum-X Ethernet networking, supporting 72TB of HBM3e and 512 NVIDIA HGX H200 GPUs,” added CEO Charles Liang.

“Supermicro’s Supercluster solution, combined with NVIDIA AI Enterprise software, is ideal for enterprise and cloud infrastructures training LLMs with up to trillions of parameters today,” he continued. “When interconnected GPUs, CPUs, memory, storage and networking are deployed across multiple nodes within a rack, they can enable modern AI.”

“NVIDIA’s latest technologies enable system makers to accelerate a diverse range of next-generation AI workloads for global markets,” said Kaustubh Sanghan, vice president of GPU Products at NVIDIA. “Supermicro is leveraging the NVIDIA accelerated computing platform with products based on the Blackwell architecture, delivering cutting-edge, easily deployable server systems for data centers.”

Supermicro 4U NVIDIA HGX H100/H200 8-GPU system uses liquid cooling to double the density of 8U air-cooled systems, reducing energy consumption and data center TCO.

Also designed to support next-generation NVIDIA Blackwell architecture-based GPUs. Supermicro Cooling Distribution Units (CDUs) and Manifolds (CDMs) are key components for distributing cooled liquid to Supermicro’s custom Direct-to-Chip (D2C) cold plates, keeping GPUs and CPUs at optimal temperatures and maximizing performance. This can reduce total data center electricity costs by up to 40% and also save physical space.

Systems powered by NVIDIA HGX H100/H200 8-GPU are ideal for generative AI learning, and the high-speed GPUs interconnected by NVIDIA NVLink, along with the large GPU memory bandwidth and capacity, play a key role in cost-effectively running LLM models. Supermicro’s supercluster creates a massive pool of GPU resources that acts as a single AI supercomputer.

Supermicro’s NVIDIA MGX system design addresses key bottlenecks in generative AI with the NVIDIA GH200 Grace Hopper superchip and provides a blueprint for future AI clusters.