Guide
Power and cooling for GPU racks
What a data center must provide for GPU servers: rack power, redundancy and air or liquid cooling, and the figures to bring.
By Tera Compute · GPU rental and colocation
GPU servers draw much more power than general-purpose servers, and they release it all as heat. A data center that suits ordinary racks does not always suit GPU racks. Before you look for colocation, know what your servers need.
How much power does a rack need?
Start with the maximum power of one server from the manufacturer's specification, in kilowatts (kW). Multiply by the number of servers in a rack, and add the switches and other equipment. The result is the rack density. A rack of 8-GPU servers commonly needs tens of kilowatts, and many older halls were designed for much less.
What does redundancy mean?
Redundancy is the spare capacity in the power and cooling systems. N+1 means one spare unit more than the load needs. 2N means a complete second system. A server usually has several power supplies on two separate power feeds, so one feed can fail and the server stays on. Ask which parts of the path are redundant.
Air cooling or liquid cooling?
Air cooling moves cold air through the server and removes hot air at the back of the rack. It works up to a rack density that depends on the design of the hall. Direct liquid cooling brings a coolant to cold plates on the GPUs and CPUs, and the liquid removes most of the heat. Some of the newest GPU servers are made for liquid cooling only. The server model sets the cooling type, so check it before you select a facility.
Which figures should you bring?
- The server model and the number of servers.
- The maximum power of each server, and the total power in kW.
- The cooling type: air, direct liquid or either.
- For liquid cooling: the supply temperature and the flow that the manufacturer specifies.
- The power connection of the server: voltage, phase and connector type.
- The number of racks that you expect, and the growth that you plan.
- The network needs: the connection between racks, and carriers or bandwidth.
What should the offer confirm?
The offer should confirm that the facility supports your rack density and cooling type at the positions offered, and the date that power and cooling are available. It should also say who installs the servers. What goes into GPU colocation costs? lists the cost items to compare.
Tera finds colocation for customer-owned GPUs. The colocation form asks for the total power and the cooling type. A request does not reserve space or authorize payment.
About Tera Compute. Tera Compute provides GPU rentals to startups: dedicated bare-metal nodes for AI teams. Tera also finds colocation for the GPUs that a startup owns. The company is based in the United States and is backed by Alliance. It is built by an operating team with deep experience in industrial-scale compute and data centers. Learn more at tera.gw.