For over a decade, Power Usage Effectiveness (PUE) has reigned supreme as the gold standard for data center efficiency metrics. End users and operators have relied on this metric to gauge the energy efficiency of their facilities' electrical and mechanical infrastructure design.
However, as its popularity has grown, so have concerns about its limitations. Critics argue that PUE fails to capture the full efficiency picture, particularly the performance of IT equipment and the concept of "useful work." As PUE improvements have stalled in recent years, industry leaders are calling for a more comprehensive metric.
This raises critical questions: Is PUE still relevant in 2024? Are data centers calculating and reporting the metric correctly? Can it be tweaked to better reflect the actual work performed? Are there superior alternatives to PUE available for measuring data center efficiency?
Power Usage Effectiveness (PUE) measures how efficiently a data center uses power or energy to deliver its design IT equipment load. Introduced by The Green Grid in 2006, it has also been recognized as a global standard under ISO/IEC 30134-2:2016.
Mathematically, PUE is the ratio of the total energy or power used in a facility to the energy or power used to run the IT equipment. The total energy includes the energy used by cooling equipment, lighting, security, power delivery components, distribution losses, etc.
We can express PUE mathematically as:
The lowest PUE value a data center can ever obtain is One. That means all the energy the data center consumes is used by its IT equipment. However, since this isn’t possible, data centers aim to get as close to 1 as possible.
The closer to one the PUE is, the more efficient the facility is. Here’s a simple example:
Partial PUE (pPUE) is a variant of the PUE metric that considers only certain parts of the data center’s system. It shows the data center’s PUE within a specific boundary.
For example, we can decide to disregard all other energy-consuming equipment in the data center and take a look at only the cooling components. Here’s an example:
As you can see, if we disregard electrical losses and set a boundary around the cooling and IT equipment, we have a lower value than the PUE.
Partial PUE is not a replacement for PUE; instead, it helps calculate the efficiencies of different critical components and designs.
Green Grid Initiative, the creators of the PUE metric, authored a comprehensive white paper with guidelines for calculating and reporting PUE. The guide specifies three different levels for measuring energy usage for PUE: Level 1(Basic), Level 2(Intermediate), and Level 3(Advanced).
Here are some of the differences between the different levels.
For Level 2, they are taken hourly/daily, while for Level 3, the measurements are taken continuously in 15-minute or less intervals.
At level 3, these readings are taken at rack level or right at the input of the IT equipment.
The PUE metric has been a massive game changer in data center operations. In the early days of the metric, data centers were operating at efficiencies of 2.5 and higher. However, today, industry averages have shrunk to about 1.58, which is largely due to the PUE-driven efficiency and sustainability narrative.
Here are some of the benefits of using the PUE metric:
PUE provides a framework for data center operators to quantify efficiency improvements in their data centers. It gives operators a methodology for measuring efficiency before and after applying changes to gauge the improvements brought about by the changes.
The data gathered during the energy measurements process enables facilities to capture real-time facility and ITE load more effectively. It also helps provide a snapshot of the energy used in the facilities on any given day.
PUE provides a means for marketers to communicate the efficiency of data center equipment to customers. It is a simple and trackable metric that can be used to certify the quality of services being provided when measured instead of modeled.
It has also been codified into an ISO standard, which makes it a more standardized and trustworthy metric when calculated correctly.
Although it’s a good metric for measuring efficiency, PUE also has limitations. In these sections, we’ll discuss some of these drawbacks and how they affect the metric.
The Green Grid’s paper emphasizes using measured data instead of modeled data when calculating PUE. However, most providers and their marketers do not use this method.
Instead, they opt for modeled data, which, in most cases, does not reflect what is gotten in real life. This leads to inaccurate PUE calculations and makes it impossible to obtain snapshots of energy usage on certain days throughout the year.
PUE is not a great tool for benchmarking and comparing different data center designs. This is because it doesn’t fully take into account various location-specific factors like altitude, climate, etc.
All the mentioned factors affect the cooling and power usage of the data center. For example, two data centers with the same design but running in different climates will have different efficiencies but may have the same modeled PUE.
If you want to compare data centers, you’ll have to use a metric that considers all those factors. Also, additional equipment like economizers aren’t accounted for in PUE calculations unless they are annualized and based on local bin weather data.
PUE should not be used to compare different cooling strategies, especially liquid cooling. This is because PUE underreports and underemphasizes the gains from liquid cooling do to the reduced IT equipment energy usage.
For example, assuming a server is switched from air-cooling to direct-to-chip cooling, the server fans aren’t going to use as much energy. So the energy drawn by the IT equipment decreases thus increasing PUE even though the system is more efficient.
This also decreases the Total facility energy, but the fan motors comprise a larger overall percentage of the IT equipment energy as compared to the total facility. So, the denominator decreases at a greater rate than the numerator, increasing the PUE.
In data centers, IT equipment does not always run at 100% capacity. The loads on the servers vary throughout the day, which in turn influences the power consumed by the IT equipment. PUE doesn’t take this into account, and it assumes it is a fixed load.
In highly virtualized server environments, the server loads vary a lot, which means that the PUE number will be incorrect most of the time.
Assuming all other factors stay the same, PUE increases with increasing IT equipment efficiency, which is quite counter-intuitive if we consider it.
For example, let’s say we replace aging IT equipment with a newer ones that are about 25% more efficient, and the power input stays the same. In this scenario, the old GPU IT Equipment draws 1.2MWh, while the Total Facility energy is 1.5MWh.
From our PUE formula, the PUE is equal to:
Now, assuming we replace the old GPU with one that is 25% more efficient, our new PUE will be:
As we can see, this will result in a higher PUE while your actual electricity bill reduces.
Many debates are springing up on what new efficiency metric is going to be the successor to PUE. Several bodies like ASHRAE are working on new efficiency metrics that can make up for PUE's shortcomings.
Some other data center metrics can also evaluated along with PUE to provide a clearer picture of data center efficiency and sustainability. Let’s look at some of the more interesting ones.
The Mechanical Load Component (MLC) is a relatively new metric introduced in the ASHRAE 90.4 Data Center Efficiency Standard. It is calculated by dividing the energy consumed by the mechanical load components by the power consumed by the IT equipment. We can express this as:
The cooling energy, pump energy, heat rejection energy, etc., form the mechanical energy.
The Annualized Mechanical Load Component (AMLC) is a slight modification of this metric. It takes the mechanical energy measurements for IT equipment running at 25%, 50%, 75%, and 100% load over the various months of the year, combines them, and divides them by the ITE energy.
Here’s how we can express it mathematically:
The Total Utilization Efficiency (TUE) is obtained by combining two metrics: IT Equipment Utilization Efficiency and PUE. We can get the ITUE by Dividing the Total energy supplied to the IT equipment by the Total energy into the compute components.
Here’s how:
The compute components are the parts of the IT equipment that do the actual work, e.g., chips, memory, hard drives, Raid controllers, etc.
This metric can be a bit tough to implement because all the IT equipment in data centers isn’t the same. Obtaining data for calculations can be difficult if the manufacturer does not provide it.
However, despite all these, it may be the better metric for evaluating the efficiency of liquid cooling systems.
Power Usage Effectiveness (PUE) still remains a key metric for measuring data center efficiency in 2024. However, it's essential to recognize its limitations and use additional metrics like TUE, AMLC and WUE alongside PUE for a more comprehensive picture of data center health.
The quest for greater data center efficiencies continues, and Stulz is leading the charge. By employing our custom cooling solutions, built for specific applications, customers are able to increase efficiency while maintaining optimal uptime.