In a recent development, NVIDIA has publicly introduced its cutting-edge AI supercomputer, Eos, specifically crafted for sophisticated AI development within expansive data center environments. Eos stands as NVIDIA's fastest AI supercomputer to date, tailored to meet the demands of advanced AI development within enterprise settings.
Eos boasts a configuration of 576 NVIDIA DGX H100 systems, each housing 8 H100 GPUs, resulting in a cumulative total of 4,608 Nvidia H100 GPUs. Additionally, it incorporates 1,152 Intel Xeon Platinum 8480C processors, each equipped with 56 cores, showcasing impressive performance capabilities across both High-Performance Computing (HPC) and AI domains. Leveraging NVIDIA's Mellanox Quantum-2 InfiniBand technology, Eos supports data transfer speeds of up to 400 Gb/s, playing a pivotal role in training expansive AI models and facilitating system scalability.
According to NVIDIA's disclosed data, Eos has secured the ninth position globally in the latest Top500 supercomputers list, achieving a remarkable peak performance of 188.65 Peta FLOPS. Notably, its FP64 performance reaches 121.4 Peta FLOPS. While Eos is a key asset for NVIDIA's internal use, it also serves as a blueprint for other enterprises aiming to develop supercomputing solutions tailored for their unique needs.
NVIDIA underscores that, beyond its robust hardware, Eos features potent software solutions designed explicitly for AI development and deployment. This includes sophisticated coordination and cluster management tools, accelerated computing storage and network libraries, and meticulously optimized operating systems. Eos is designed to accommodate a diverse range of applications, from generative AI, such as ChatGPT, to applications within AI factories.
NVIDIA emphasizes that Eos represents the amalgamation of their extensive expertise and technologies in the field of AI. It stands as the culmination of insights gained from previous DGX supercomputers, positioned to aid enterprises in addressing their most challenging projects and attaining their AI objectives.
While the precise cost of Eos and the pricing structure of the NVIDIA DGX H100 systems remain undisclosed, contingent on various factors, the estimated cost of $30,000 to $40,000 per H100 suggests that the overall system cost is likely to be substantial for industry stakeholders.