Nvidia provides the first public view of its fastest AI supercomputer — Eos is powered by 4,608 H100 GPUs, tuned for generative AI

Nvidia has unveiled Eos, its cutting-edge enterprise-focused supercomputer tailored for advanced AI development on a massive scale within data centers. Eos, currently utilized by Nvidia itself, ranks as the 9th most powerful supercomputer globally in FP64 performance, potentially outstripping all others in AI tasks. Comprising 576 DGX H100 systems, each equipped with eight Nvidia H100 GPUs for AI and HPC workloads, Eos boasts a formidable 1,152 Intel Xeon Platinum 8480C processors and 4,608 H100 GPUs. This setup enables Eos to deliver exceptional performance, achieving 121.4 FP64 PetaFLOPS and 18.4 FP8 ExaFLOPS for HPC and AI, respectively.

With a design rooted in the DGX SuperPOD architecture, Eos prioritizes AI workloads and scalability. Leveraging Nvidia’s Mellanox Quantum-2 InfiniBand technology featuring data transfer speeds up to 400 Gb/s, Eos facilitates proficient training of large AI models and scalability. Complementing its robust hardware, Eos is accompanied by purpose-built software optimized for AI development and deployment. This comprehensive software stack includes AI development tools, orchestration and cluster management capabilities, accelerated compute storage, and network libraries, along with an AI workload-optimized operating system.

Nvidia’s Eos holds the promise of revolutionizing diverse applications, from generative AI akin to ChatGPT to comprehensive AI factory operations. The supercomputer’s integrated software stack has been refined based on Nvidia’s extensive experience with prior DGX supercomputers like Saturn 5 and Selene, underscoring Nvidia’s expertise in AI advancement. Eos facilitates enterprises in tackling their most challenging projects and realizing their AI goals both now and in the future.

While the pricing of Eos remains undisclosed, the confidential nature of Nvidia’s DGX H100 system costs, which can be influenced by various factors like volume, complicates cost estimation. Given that each Nvidia H100 GPU can range from $30,000 to $40,000 or potentially beyond, the financial implications of Eos are substantial. This advancement underscores the evolving landscape of AI supercomputing and the significant investments in cutting-edge technologies.

For enthusiasts seeking the latest insights on PC tech news, Tom’s Hardware provides expert analysis and updates on CPUs, GPUs, AI technology, and maker hardware over its 25-year history. Stay informed with breaking news and in-depth reviews delivered directly to your inbox by subscribing to Tom’s Hardware.