According to a recent Market Watch article, “The Rise of Artificial Intelligence Comes with Rising Needs for Power,” datacenters became less efficient (PUE) for the first time since 2007. So unless you have been living in the Shire, pay close attention!
Today, there are more than 7,000 mega datacenters worldwide, IDC reports. The energy efficiency of these facilities is determined by a metric known as power usage effectiveness (PUE). IDC considers a PUE score of 1.2 to be “very efficient,” but the average datacenter still has lots of room for improvement, notching a PUE of 1.67 this year. According to an Uptime Institute survey, this mark is up a tenth from a year earlier – hence the first time PUE values have deteriorated since 2007.
Are AI Workloads Evil?
No, this revelation is not a plot by the Dark Lord Sauron to rule all of Middle Earth. As it turns out, neural networks now require custom-built hardware for training. Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs) are common accelerators that deliver improvements in the amount of time required to train a neural network to complete a desired task. Though newer processors can achieve similar results in a shorter time, they consume more power than legacy hardware. So, Frodo, listen up: Training a neural network can emit 17 times more carbon dioxide than an average American in a year, and five times the lifetime emissions of an average car.
VMware’s Project Magna – AI for Operating the Data Center- The Two Towers?
A couple weeks back, VPS wrote about Project Magna and VMware’s vision for the software defined datacenter of the future. It’s not just AI workloads that are going to challenge datacenters’ power requirements. AI also has a growing role in running a datacenter – sort of a “Datacenter AI Operating System.” Think of this as the second movie in the “Lord of the Rings” trilogy.
So, if PUE is actually getting worse and AI workloads are sucking every available kW and Datacenter AI is running the place (sucking even more power), what do we do? Each one of these “rings of power” is a force of its own and not going away.
One Ring (of Power) to Rule Them All
Thank goodness there is one single “ring of power,” which many people, articles, blogs, analysts and companies are missing – it is Software Defined Power (SDP), the fourth and final pillar of the SDDC.
Too many people in the industry continue to assume power will be acquired, provisioned and deployed in the same model that has been used for decades. Literally, decades. All the low-hanging PUE fruit have been picked. Renewable energy improvements can still be made, but they do not address the “stick in the eye” problem that is happening inside datacenters.
While many people brag about PUE, we have found many then admit 50% to 80% of their datacenters’ power capacity is stranded–and it shows up every single time at the rack level. Perhaps datacenter designers and operators should switch their thinking to Rack Power Efficiency (RPE) or Rack Power Utilization (RPU)?
SDP is a software-based power control platform that uses machine learning and big data analytics to provide awareness and recommendations, including how, when and where power should be utilized to better inform the capacity planning process. Furthermore, when SDP is combined with hardware, the union automates datacenter operations and power management. The result increases utilization of installed capacity by removing unnecessary buffers and automating SLAs while mitigating risks and improving uptime of critical applications.
There is one ring to rule them all: With SDP, greenfield datacenters that originally were designed at 20 MW now can be designed for 14 MW. With capex at $10 million per MW, that’s $70 million in savings just on the original design! For brownfield facilities with 50% to 80% of their power stranded and unusable, SDP enables reuse of 30% to 60% of the power. See our ROI for an Enterprise Client to understand how SDP can recapture 352 kW to reuse out of 1 MW provisioned.
While it’s easy to make excuses for rising PUE or staying with the status quo, embracing SDP is the best way for this story have a happy ending.