One of the considerations when building any data center is the amount AC you are going to need (in tons) to keep your equipment properly cooled. This is particularly important as we consider the densities that we are able to accomplish now with systems optimised for Cloud, Hyperscale and HPC environments. Unless you are going to follow Facebook’s lead with the systems described this past week in the OpenCompute announcement, you’ll likely need some sort of additional cooling source. If you are hosting your environment at one of the many hosting facilities such as Internap, Savvis or others, they will want to know this information as well.
To get this information, we really need just a few pieces of information. First, we need to understand the wattage of a system. For the purposes of this example, lets say that you are planning on deploying a large HPC solution using a combination of GPGPU and large CPU systems, and you want to pack as much into the rack as possible. The total wattage per CPU and GPU combinations is 7192 watts. To calculate BTUs/hr you would take the total wattage, and multiply that by 3.41. This would give you a total of 24,524.72 BTUs/hr. Next, you divide that number by 12000, which gives you 2.04. That is the amount of tonnage in AC you would need to properly cool your system.
So, if we want to put eight of these systems in a rack for maximum density, you would simply multiply 2.04 by 8, and you come up with 16.5 tons of AC per rack. So…what we see is while we can technically fit the solution in the rack, the per-rack cooling requirement may be too high. To off-set that, we can simply cut back on the number of systems in a rack to reach the per-rack requirement.
There are, of course, other variables you’ll take into consideration when calculating wattage, but once you have that, you should use the PEAK load when performing your calculations. This will keep you from having any unwanted surprises after the system is deployed.