Catalina is Meta’s implementation of GB200 NVL72. A single “pod” consists of:1
- 4x “Air Assisted Liquid Cooling” (liquid-air) heat exchanger racks
- 2 “IT racks,” each with
- 18 compute trays, each with
- 9 NVLink switches
- 94 kW
Unlike the standard GB200 NVL72 with cross-rack cabling, Catalina employs 1:1 CPU to GPU.
It employs cross-rack cabling to create a single NVLink domain across the two “IT racks.” Here’s what a single pod looks like:
It is a stark visual reminder of what happens when you cheap out on datacenter infrastructure and try to wedge HPC infrastructure into it. By comparison, a properly designed datacenter with facility water would be able to fit these six racks into a single rack. See Fairwater > Racks, Microsoft’s purpose-built datacenter for this same GB200 NVL72 pod, as a point of comparison.