NVIDIA has announced that xAI's Colossus supercomputer cluster, featuring 100, 000 NVIDIA Hopper GPUs located in Memphis, Tennessee, achieved this scale using the NVIDIA Spectrum-X™ Ethernet networking platform. This platform is engineered for optimal performance in large AI factories, utilizing standards-based Ethernet for its Remote Direct Memory Access (RDMA) network. Colossus, now the world's largest AI supercomputer, is employed in training xAI’s Grok family of large language models, which includes chatbots for X Premium subscribers. xAI plans to expand Colossus to a total of 200, 000 Hopper GPUs. Built in a remarkable 122 days—a significant reduction from the typical timeline for such systems—Colossus began training just 19 days after the first rack was set up. During the training of the extensive Grok model, Colossus experiences exceptional network performance, maintaining 95% throughput and zero application latency degradation or packet loss, thanks to Spectrum-X congestion control. In contrast, standard Ethernet struggles with flow collisions, achieving only 60% data throughput. Gilad Shainer, NVIDIA's senior vice president of networking, emphasized the critical nature of AI, requiring enhanced performance, security, scalability, and cost efficiency, which the Spectrum-X platform provides for innovators like xAI.

Elon Musk praised Colossus as the "most powerful training system in the world, " acknowledging the efforts of xAI and NVIDIA. An xAI spokesperson highlighted that the combination of NVIDIA's Hopper GPUs and Spectrum-X technology allows for unprecedented training of AI models on a massive scale. Central to the Spectrum-X platform is the Spectrum SN5600 Ethernet switch, supporting speeds up to 800Gb/s, paired with NVIDIA’s BlueField-3® SuperNICs for enhanced performance. The Spectrum-X Ethernet networking brings advanced features previously associated only with InfiniBand, including adaptive routing, congestion control, and improved visibility and performance for multi-tenant AI environments. NVIDIA leads in accelerated computing, but forward-looking statements about the benefits and performance of its technologies come with risks that may impact actual results. Factors such as global economic conditions, reliance on third parties, and technological competition can affect outcomes. NVIDIA provides regular updates through SEC filings. © 2024 NVIDIA Corporation. All rights reserved.