SuperMicro Unveils Super Clusters Powered by NVIDIA Chips

San Jose and Taipei, Taiwan – Supermicro is introducing at COMPUTEX in Taiwan a ready-to-deploy liquid-cooled AI data center, designed for cloud-native solutions that accelerate generative AI adoption for enterprises across industries with its SuperClusters. The servers are optimized for the NVIDIA AI Enterprise software platform for the development and deployment of generative AI. With Supermicro’s 4U liquid-cooled, NVIDIA recently introduced Blackwell GPUs can fully unleash 20 PetaFLOPS on a single GPU of AI performance and demonstrate 4X better AI training and 30X better inference performance than the previous GPUs with additional cost savings.

Supermicro recently announced a complete line of NVIDIA Blackwell architecture-based products for the new NVIDIA HGXTM B100, B200, and GB200 Grace Blackwell Superchip.

“Supermicro continues to lead the industry in creating and deploying AI solutions with rack-scale liquid-cooling,” said Charles Liang, president and CEO of Supermicro. “Data centers with liquid-cooling can be virtually free and provide a bonus value for customers, with the ongoing reduction in electricity usage. Our solutions are optimized with NVIDIA AI Enterprise software for customers across industries, and we deliver global manufacturing capacity with world-class efficiency. The result is that we can reduce the time to delivery of our liquid-cooled or air-cooled turnkey clusters with NVIDIA HGX H100 and H200, as well as the upcoming B100, B200, and GB200 solutions. From cold plates to CDUs to cooling towers, our rack-scale total liquid cooling solutions can reduce ongoing data center power usage by up to 40%.”

CEO Liang is a Taiwan native.

At COMPUTEX 2024, Supermicro is revealing its upcoming systems optimized for the NVIDIA Blackwell GPU, including a 10U air-cooled and a 4U liquid-cooled NVIDIA HGX B200-based system. In addition, Supermicro will be offering an 8U air-cooled NVIDIA HGX B100 system and Supermicro’s NVIDIA GB200 NVL72 rack containing 72 interconnected GPUs with NVIDIA NVLink Switches, as well as the new NVIDIA MGX™ systems supporting NVIDIA H200 NVL PCIe GPUs and the newly announced NVIDIA GB200 NVL2 architecture.

“Generative AI is driving a reset of the entire computing stack — new data centers will be GPU-accelerated and optimized for AI,” said Jensen Huang, founder and CEO of NVIDIA. “Supermicro has designed cutting-edge NVIDIA accelerated computing and networking solutions, enabling the trillion-dollar global data centers to be optimized for the era of AI.”

The rapid development of large language models and the continuous new introductions of open-source models such as Meta’s Llama-3 and Mistral’s Mixtral 8x22B make today’s state-of-the-art AI models more accessible for enterprises. The need to simplify the AI infrastructure and provide accessibility in the most cost-efficient way is paramount to supporting the current breakneck speed of the AI revolution. The Supermicro cloud-native AI SuperCluster bridges the gap between cloud convenience of instant access and portability, leveraging the NVIDIA AI Enterprise, allowing moving AI projects from pilot to production seamlessly at any scale. This provides the flexibility to run anywhere with securely managed data, including self-hosted systems or on-premises large data centers.

With enterprises across industries rapidly experimenting with generative AI use cases, Supermicro collaborates closely with NVIDIA to ensure a seamless and flexible transition from experimentation and piloting AI applications to production deployment and large-scale data center AI. This result is achieved through rack and cluster-level optimization with the NVIDIA AI Enterprise software platform, enabling a smooth journey from initial exploration to scalable AI implementation.

Managed services compromise infrastructure choices, data sharing, and generative AI strategy control. NVIDIA NIM microservices, part of NVIDIA AI Enterprise, offer managed generative AI and open-source deployment benefits without drawbacks. Its versatile inference runtime with microservices accelerates generative AI deployment across a wide range of models, from open-source to NVIDIA’s foundation models. In addition, NVIDIA NeMoTM enables custom model development with data curation, advanced customization, and retrieval-augmented generation (RAG) for enterprise-ready solutions. Combined with Supermicro’s NVIDIA AI Enterprise ready SuperClusters, NVIDIA NIM provides the fastest path to scalable, accelerated Generative AI production deployments.

Supermicro’s current generative AI SuperCluster offerings include:

  • Liquid-cooled Supermicro NVIDIA HGX H100/H200 SuperCluster with 256 H100/H200 GPUs as a scalable unit of compute in 5 racks (including 1 dedicated networking rack)
  • Air-cooled Supermicro NVIDIA HGX H100/H200 SuperCluster with 256 HGX H100/H200 GPUs as a scalable unit of compute in 9 racks (including 1 dedicated networking rack)
  • Supermicro NVIDIA MGX GH200 SuperCluster with 256 GH200 GraceTM Hopper Superchips as a scalable unit of compute in 9 racks (including 1 dedicated networking rack)

Supermicro SuperClusters are NVIDIA AI Enterprise ready with NVIDIA NIM microservices and NVIDIA NeMo platform for end-to-end generative AI customization and optimized for NVIDIA Quantum-2 InfiniBand as well as the new NVIDIA Spectrum-X Ethernet platform with 400Gb/s of networking speed per GPU for scaling out to a large cluster with tens of thousands of GPUs.

Supermicro’s upcoming SuperCluster offerings include:

  • Supermicro NVIDIA HGX B200 SuperCluster, liquid-cooled
  • Supermicro NVIDIA HGX B100/B200 SuperCluster, air-cooled
  • Supermicro NVIDIA GB200 NVL72 or NVL36 SuperCluster, liquid-cooled

Supermicro’s SuperCluster solutions are optimized for LLM training, deep learning, and high volume and batch size inference. Supermicro’s L11 and L12 validation testing and on-site deployment service provide customers with a seamless experience. Customers receive plug-and-play scalable units for easy deployment in a data center and faster time to results.