xAI's Colossus supercomputer: 100,000 Hopper GPUs in Memphis, built in 122 days

2024-10-28 · Publisher : xAI · Domain : Cloud & DevOps

NVIDIA details Colossus, the 100,000-GPU Hopper cluster xAI built in Memphis in 122 days, linked with Spectrum-X Ethernet, with a doubling to 200,000 GPUs under way.

### Background On 28 October 2024, NVIDIA announces that xAI's **Colossus** cluster, made of 100,000 Hopper GPUs in Memphis, Tennessee, reached that scale using its Spectrum-X Ethernet networking platform. The facility and the supercomputer were built by xAI and NVIDIA in 122 days, versus many months or years for a system of this size. Colossus is used to train the Grok family of language models. xAI is doubling its size to reach 200,000 Hopper GPUs. ### Key facts - Timeline: 122 days for the site and system, including 19 days between the first rack arriving and training starting. - Size: 100,000 Hopper GPUs, with a doubling to 200,000 under way. - Networking: Spectrum-X Ethernet with SN5600 switches (800 Gb/s) and BlueField-3 SuperNICs, instead of InfiniBand. - Claimed performance: 95% data throughput sustained, versus 60% for standard Ethernet, without packet loss from flow collisions. - Use: training the Grok family of models. ### Why it matters Colossus sets a new benchmark for deployment speed: an AI campus of this size in four months, where classic projects take years. It also shows that Ethernet networking can replace InfiniBand at very large scale, which opens the market to other equipment makers. Throughput and timeline figures come from NVIDIA and xAI and are unaudited. ### Takeaway Build speed becomes a competitive advantage of AI cloud, with Ethernet as an alternative to InfiniBand.

Tags : Datacenter, NVIDIA, États-Unis, Puces