The Secrets of xAI Colossus: 100,000 GPUs

What's inside a 100,000 GPU cluster? 🤔

Patrick from ServeTheHome takes us on a tour with his camera through different parts of the server, offering a bird's-eye view of its operations. While some details More specific details about the supercomputer, such as its power consumption and the size of the bombs, could not be revealed due to a confidentiality agreement, xAI was responsible for blurring and censoring parts of the video before its release. 🎥

Despite this, the most important thing, like the servers GPU from Supermicro, remained virtually untouched throughout the footage. These GPU servers are Nvidia HGX H100, a powerful server solution featuring eight H100 GPUs each. 🚀 The HGX H100 platform is integrated within the 4U Universal GPU Liquid Crystal Display System Cooled from Supermicro, providing easily hot-swappable liquid cooling for each GPU. ❄️

These servers are organized in racks containing eight servers each, totaling 64 GPU per rack. 1U manifolds are sandwiched between each HGX H100, providing the necessary liquid cooling for the servers. At the bottom of each rack, we find another 4U Supermicro unit, this time equipped with a redundant pump system and a rack monitoring system. 🔍

🖥️ These racks are organized in groups of eight, allowing for 512 GPU per array. Each server is equipped with four power supplies redundant. At the back of the racks GPU, there are three-phase power supplies, Ethernet switches, and a rack-sized manifold that provides all the liquid cooling. 💧

There are more than 1,500 racks in the Colossus cluster. GPU, distribuidos en cerca de 200 conjuntos de bastidores. Según Jensen Huang, director ejecutivo de Nvidia, las GPU de estas 200 matrices se instalaron completamente en solo tres semanas. 🚀

Since an AI supercluster constantly training models requires huge bandwidth, xAI went further in its interconnectivity of redEach graphics card has a dedicated 400GbE NIC (network interface controller), with an additional 400Gb NIC per server. 🔗 This means each HGX H100 server has 3.6 terabits per second of Ethernet. Impressive, isn't it? And yes, the entire cluster runs on Ethernet, rather than InfiniBand or other exotic connections that are standard in the supercomputing world. 🌐

Of course, a supercomputer like the Grok 3 chatbot, which trains AI models, needs more than just GPU to run at its best. 🔥 While details on storage and CPU servers in Colossus are somewhat limited, thanks to Patrick's video and the blog post, we know that these servers are usually in Supermicro chassis. 🚀

1U NVMe-forward servers with x86 platform CPUs inside are used, providing both storage and computing power, and are equipped with liquid cooling at the rear. 💧 Additionally, very compact Tesla Megapack battery banks can be seen outside. ⚡️

The start-stop feature of the array, with its millisecond latency between banks, was too much for the conventional power grid or Musk's diesel generators. So several Tesla Megapacks (each with a capacity of 3.9 MWh) are used as an intermediate power source between the red electrical and the supercomputer. 🖥️🔋 This ensures optimal and efficient operation, avoiding interruptions. 🚦✨

🌟 Using Colossus and Musk's stable supercomputer 🌟

The xAI supercomputer Colossus is currently, according to Nvidia, the largest AI supercomputer in the world. 🤯 While many of the world's leading supercomputers are used in research by contractors or academics to study weather patterns, diseases, or other complex tasks, Colossus has sole responsibility for training X's (formerly Twitter) various AI models. Most notably, Grok 3, Elon's "anti-woke" chatbot that's available only to X Premium subscribers. 🤖

Additionally, ServeTheHome was informed that Colossus is training AI models «del futuro»; modelos cuyos usos y capacidades supuestamente están más allá de las capacidades actuales de la IA. 🚀 La primera fase de construcción de Colossus está completa y el clúster está completamente operativo, pero aún no está todo terminado. La supercomputadora de Memphis pronto se actualizará para duplicar su capacidad de GPU, con 50.000 GPU H100 adicionales y 50.000 GPU H200 de próxima generación. 🔥

This update It will also more than double its energy consumption, which is already too much for the 14 diesel generators Musk added to the site in July to handle. ⚡ While it is short of Musk's promise of 300,000 H200s inside Colossus, that could be part of Phase 3 of updates. 🔋

On the other hand, the 50,000 GPU Cortex supercomputer at Tesla's "Giga Texas" plant also belongs to a Musk company. Cortex is dedicated to training Tesla's autonomous AI technology through camera streaming and image detection, as well as Tesla's autonomous robots and other AI projects. 🤖🚗

Additionally, Tesla will soon see the construction of the Dojo supercomputer in Buffalo, New York, a $500 million project coming soon. 💸 Meanwhile, industry speculators like Baidu CEO Robin Li predict that 99% of AI companies could collapse when the bubble bursts. Whether Musk's record spending on AI will backfire or pay off remains to be seen. ⏳