Supermicro has launched a new set of solutions to advance the deployment of generative AI, marking a significant step in the evolution of infrastructure for large language models (LLMs). These SuperCluster solutions are designed as core components to support current and future AI demands.
Supermicro has launched a new set of solutions to advance the deployment of generative AI, marking a significant step in the evolution of infrastructure for large language models (LLMs). These SuperCluster solutions are designed as core components to support current and future AI demands.
This new release includes three distinct SuperCluster configurations tailored for generative AI tasks. Options include a 4U liquid-cooled system and an 8U air-cooled setup engineered for intensive LLM training and high-capacity LLM inference. Additionally, a 1U air-cooled variant featuring Supermicro NVIDIA MGX systems is geared toward cloud-scale inference applications. These systems are built to deliver unparalleled performance in LLM training, boasting features such as large batch sizes and substantial volume handling capabilities for LLM inference.
With the ability to produce up to 5,000 racks per month, Supermicro is positioned to rapidly supply complete generative AI clusters, promising faster delivery speeds to its clients. A 64-node cluster, as an example, can incorporate 512 NVIDIA HGX H200 GPUs, utilizing high-speed NVIDIA Quantum-2 InfiniBand and Spectrum-X Ethernet networking to achieve a robust AI training environment. In conjunction with NVIDIA AI Enterprise software, this config is an ideal solution for enterprise and cloud infrastructures aiming to train sophisticated LLMs with trillions of parameters.
The new Supermicro 4U NVIDIA HGX H100/H200 8-GPU systems leverage liquid cooling to double the density compared to the 8U air-cooled alternatives, resulting in lower energy consumption and a decrease in the total cost of ownership for data centers. These systems support next-generation NVIDIA Blackwell architecture-based GPUs, featuring efficient cooling technologies that maintain optimal temperatures for maximum performance.
The Supermicro SuperClusters are scalable solutions for training massive foundation models and creating cloud-scale LLM inference infrastructures. With a highly scalable network architecture, these systems can expand from 32 nodes to thousands, ensuring seamless scalability. Integrating advanced liquid cooling and comprehensive testing processes guarantees operational efficiency and effectiveness.
Supermicro details two primary configurations: the SuperCluster with 4U Liquid-cooled System, capable of supporting up to 512 GPUs in a compact footprint, and the SuperCluster with 1U Air-cooled NVIDIA MGX System, designed for high-volume, low-latency inference tasks. Both configurations are highlighted for their high network performance, which is essential for LLM training and inference.
Here is a quick rundown of their specifications:
SuperCluster with 4U Liquid-cooled System in 5 Racks or 8U Air-cooled System in 9 Racks
SuperCluster with 1U Air-cooled NVIDIA MGX System in 9 Racks
Supermicro is also announcing the expansion of its AI system offerings, including the latest in NVIDIA’s data center innovations aimed at large-scale generative AI. Among these new technologies are the NVIDIA GB200 Grace Blackwell Superchip and the B200 and B100 Tensor Core GPUs.
To accommodate these advancements, Supermicro is seamlessly upgrading its existing NVIDIA HGX H100/H200 8-GPU systems to integrate the NVIDIA HGX B100 8-GPU and B200. Furthermore, the NVIDIA HGX lineup will be bolstered with the new models featuring the NVIDIA GB200, including a comprehensive rack-level solution equipped with 72 NVIDIA Blackwell GPUs. In addition to these advancements, Supermicro is introducing a new 4U NVIDIA HGX B200 8-GPU liquid-cooled system, leveraging direct-to-chip liquid cooling technology to handle the increased thermal demands of the latest GPUs and unlock the full performance capabilities of NVIDIA’s Blackwell technology.
The new Supermicro’s GPU-optimized systems will soon be available, fully compatible with the NVIDIA Blackwell B200 and B100 Tensor Core GPUs and certified for the latest NVIDIA AI Enterprise software. The Supermicro lineup includes diverse configurations, from NVIDIA HGX B100 and B200 8-GPU systems to SuperBlades capable of housing up to 20 B100 GPUs, ensuring versatility and high performance across a wide range of AI applications. These systems include first-to-market NVIDIA HGX B200 and B100 8-GPU models featuring advanced NVIDIA NVLink interconnect technology. Supermicro indicates they are poised to deliver training outcomes for LLMs (3x faster) and support scalable clustering for demanding AI workloads, marking a significant leap forward in AI computational efficiency and performance.
Supermicro Liquid Cooling Technology
Engage with StorageReview
Newsletter | YouTube | Podcast iTunes/Spotify | Instagram | Twitter | TikTok | RSS Feed
JetCool has launched an innovative liquid cooling module tailored for NVIDIA's H100 SXM and PCIe GPUs, claiming a significant advancement…
iXsystems has launched the TrueNAS Enterprise H-Series platforms, designed to give organizations ultimate performance. The H10 model is now available,…
Hannover Messe 2024 represents a significant event in the global industrial sector, serving as the world's largest industrial trade fair.…
The IBM Storage Assurance program offers access to the latest FlashSystem hardware and software, supporting investment protection from day one.…
Proxmox Backup Server 3.2 has been released - open-source solution designed for backup of VMs, containers, and physical hosts. (more…)
IBM has unveiled the FlashSystem 5300, setting a new standard for entry-level all-flash storage systems by providing impressive performance, high…