• If I learned one thing at NANOG 96, it’s how to build an up-to-date state of the art AI data centre! • As I now understand it, the process is quite simple. • Take everything that is at the raw bleeding edge of today’s technology: A collection of 60,000 advanced GPUs using 3nm silicon fabrication processes, 3D-stacked high bandwidth memory. • You’ll need massive amounts of power, capable of delivering some 150kW to 200kW per equipment rack, and a local power system capable of sustaining up to 100MW of total power for the entire centre. • Equip every rack with a liquid cooling system, and provision each rack with power bus bars. • Then, mesh-connect the GPUs to both each other and high speed storage in a lossless connectivity mesh using 800G optics and high density switches.

Article Summaries:

  • If I learned one thing at NANOG 96, it’s how to build an up-to-date state of the art AI data centre! As I now understand it, the process is quite simple. Take everything that is at the raw bleeding edge of today’s technology: A collection of 60,000 advanced GPUs using 3nm silicon fabrication processes, 3D-stacked high bandwidth memory. You’ll need massive amounts of power, capable of delivering some 150kW to 200kW per equipment rack, and a local power system capable of sustaining up to 100MW of total power for the entire centre. Equip every rack with a liquid cooling system, and provision each

Sources: