Heat density: The Challenges of Compact Compute Nodes
With the popularity of physically smaller compute nodes like Advanced Clustering’s own Pinnacle Flex system, comes the issue of heat density. While the size of nodes is shrinking, the power draw of these systems is staying about the same, or going up with the addition of accelerators like GPUs, and Xeon Phi co-processors.
Compared to previous generation 1U or 2U systems, we are doubling up on power draw and when power goes up, of course, heat does too. It’s quite easy to get to over 1,000 watts of power consumption per rack unit (U) of space. Even if you were to fill a standard 42U rack 3/4 full, at around 30U power consumption could be near to 30kW. To put that into perspective, an average US residential customer uses only 10.8kW. (http://www.eia.gov/tools/faqs/faq.cfm?id=97&t=3). Now you have the equivalent of 3 households use of power in the size of a refrigerator.
This shrinking of compute nodes combined with high power draw leads to a major heat density problem: lots of power and heat in a very small space. Most traditional raised floor data centers were not designed to handle this kind of heat load, as many are designed for only 2-5 kW per rack.
From talking to customers over the years, we’ve found the concept of heat density is often overlooked, or not taken into consideration. This is a conversation that needs to happen with facility or data center managers to make sure the proposed system can be cooled adequately in the space available. Many times clusters have to be split into multiple cabinets with lots of empty rack Us to compensate for the limitations of the room’s air conditioning units. There are other alternatives that can also be deployed including in row cooling, rear door heat exchangers, air conditioned cabinets, etc.
when designing your next cluster and we can help you navigate all of the often overlooked challenges of deploying larger systems. We look forward to working with you.