In a recent interview, Garrett Furlong, product manager at Server Technology, a brand of Legrand, spoke to Matt Vincent, Editor-in-Chief of Data Center Frontier, about how intelligent power infrastructure is helping data center operators navigate the AI transition.

Matt Vincent:

AI dominates many infrastructure conversations today. But the reality is that most data centers are still running mixed workloads. How should operators be thinking about power infrastructure that supports both today's requirements and tomorrow's higher density deployments?

Garrett Furlong :

AI is definitely getting all the headlines right now, but legacy enterprise workloads are still what pay the bills. Data centers aren't exclusively deploying 100 kW liquid cooled racks — the average rack is still in the 15 kW to 20 kW range.

If operators want to future-proof their environments now to support mixed and higher-powered workloads in the future, they should be looking at slightly upsizing PDUs — adding in more circuit breakers per outlet and making load balancing easier with alternating phase outlets.

Detachable cord PDUs offer the most flexibility. You could start with a smaller cord for what you actually need right now and then upgrade later to a larger input cord that supports your next higher density deployment.

Those are the kind of examples of decisions that will give you greater flexibility in the future, because who knows what you're going to need tomorrow?

Vincent:

Absolutely. We often hear about power capacity. But intelligent power management is becoming equally important. How is software changing the way operators manage rack power?

Furlong:

Your ability to manage power in the data center, especially at the rack level, is only as good as your telemetry readings from the PDU. When you can look at power consumption on an individual outlet level, you move from reactive power management focused on uptime or power consumption, to the predictive power management where operators can run a failover scenario accurately, ensuring that mission critical servers are actually protected.

But before any of that happens, Zero Touch Provisioning of PDUs simplifies deployment, and  gets the PDU into operation faster. Deploying 500 new PDUs used to mean physically going in and connecting to every single unit. Today, you connect the PDU to the network and the firmware automatically pulls the config file. We can use software and firmware to eliminate human error and cut deployment times.

Vincent:

Open integration is becoming increasingly important across the data center. How do open APIs help operators integrate power infrastructure into their wider management and automation platforms?

Furlong:

Power infrastructure is no longer operated in a silo, forcing teams to use different vendor-locked software just to see their basic power metrics. You want your PDU to be able to communicate with the DCIM and with an IT orchestration platform.

When your power infrastructure is all built on open, RESTful APIs such as Redfish, you give yourself a lot more flexibility. You open up the possibilities of what your equipment can do for you. Trying to automate anything like the migration of a virtual machine to a different part of the data center would be a lot more difficult if you weren't able to easily communicate between those applications.

Vincent:

Environmental monitoring is becoming much more granular. Beyond basic temperature monitoring, what kinds of intelligence can modern rack level sensors provide?

Furlong:

Nobody wants their gear overheating, so temperature is definitely the one that most people are most concerned with. But everyone's environments are going to have different factors they're going to want to monitor. You're likely also worried about humidity. If your humidity is too high, equipment corrodes and you're shortening the gear’s total useful life. If it's too low, you're risking ESD discharge.

Vibration is another common one. If your connections come loose in your cabinet and you’re at risk of losing power or connectivity, you want to know.

If your facility is undergoing construction, you're worried about dust collecting on heat sinks, which reduces the cooling capabilities of that piece of equipment.

Or you're worried about zinc whiskers if you've got a raised floor.

Differential air pressure and airflow sensors are another option. If you've got a hot aisle or a cold aisle you would place sensors at the top, middle and bottom of a rack to make sure you have the proper airflow front to back.

And of course, monitoring your liquid cooled system for leaks, whether it’s direct-to-chip or a rear door heat exchanger, is critical. You can put leak detection sensors on the floor, or you can use a rope style that tells you the position of the leak along the rope.

Those are just examples of sensors and reasons why people might be using them in the data center. If you’ve made a significant investment in all that equipment, the last thing you want is something preventable causing it to go down. Environmental sensors are proactive  measures you can take that ensure uptime and reliability.

Vincent:

As power architectures become more complex, we're seeing increasing interest in integrated cabinet designs. What advantages does bringing together cabinets, bus bars, rack PDUs and monitoring deliver, compared with treating each component separately?

Furlong:

With the shift to AI infrastructure, a lot of changes have been happening all at once — you’ve got OCP-based architectures, rear door heat exchangers, rear DC bus bars, and even beyond that, you're moving into low voltage DC architectures.

These changes force the facilities people and IT gear people to actually work together, where typically they're at each other's throats. You've got to make sure all these pieces are integrated and actually able to function properly from the start. For example, it used to be no problem if you had to make a last minute change when ordering the cabinet because there were too many whips in the back. You just increased the size of the brush cutouts and your problem was solved.

Today, solving those kinds of problems is no longer a last-minute decision. You have to be thinking about everything — the cabinets, the bus bars — ahead of time, when you're specifying everything out. Even if you're not doing Open Compute Project and you’re using rack PDUs, as AI-dense deployments scale and you're looking at a hundred 125-amp PDUs, the amount of room in the back of the cabinet becomes critical space.

Those things all need to be thought of well ahead of time and in conjunction with one another, rather than just piecemealing it as you go.

Vincent:

Great points. Efficiency is a major priority across today's data centers. Where are you seeing the biggest opportunities for operators to improve efficiency through intelligent rack power distribution?

Furlong:

From a PDU standpoint, being able to switch individual machines on and off remotely is going to decrease your total power usage. Making sure that all your phases are balanced assures that you get maximum effectiveness out of your PDU.

Looking into a bit of a newer architecture, the Open Compute Project architecture improves power distribution efficiency by disaggregating power supplies from the servers themselves. PSUs are consolidated onto dedicated shelves, freeing up additional space in the cabinet for compute. You can also run them at higher power and get better efficiency by running them at 50% load.

Those are the kind of opportunities available for operators to improve efficiency. You either learn to cycle your individual PDU outlets on and off, or you look at how your architecture can become more efficient as a whole.

Vincent:

Going back to the design point, I have a big picture question. Many organizations want infrastructure that can evolve rather than be replaced. What design principles help create a more scalable and standardized rack power platform?

Furlong:

What sticks out to me on this question is the standardized rack power platform. We've seen Open Compute Project move from early adopters to broad hyperscaler deployment. They've put out this standardized rack power platform that is intentionally designed to be upgraded as you go. It makes the installation process incredibly easy because there's no cabling between the PDU and the servers themselves.

We've seen it even in the most recent version of the open rack standards. They started with 18 kilowatt power shelves with three kilowatt power supplies. The next iteration was 33 kilowatt power shelves with 5.5 kW power supplies because they said 18 kW wasn't enough power. Then they said the 33 kW shelves weren't enough power, so they specified 72 kW shelves. Even beyond that, they're already looking at 110 kW shelves. And all those components are designed to be easily added or removed from of the shelves, so nothing about the cabinet, nothing about the IT gear itself really needs to change. They can simply upgrade the power shelves and the bus bar in the back of the cabinet to increase the capacity of it.

Those are the kind of design decisions that help make things more scalable and standardized.

Vincent:

We always like to end with the crystal ball questions. Looking ahead over the next 3 to 5 years, what developments do you expect will have the biggest impact on intelligent rack power infrastructure?

Furlong:

To bring it full circle, the standard data center architecture isn't going anywhere. Those 15 to 20 kW racks are and will continue to be the majority of infrastructure out there. That said, Open Compute Project is continuing to progress and see adoption from more than just hyperscalers.

And the hyperscalers — Meta, Google, Microsoft — are pushing to introduce low voltage DC architectures like a ±400 volt or 800 volt. Those are the changes that are going to be introduced in the next 3 to 5 years. There are a lot of technological advancements behind that, from the introduction of solid-state transformers upstream to making sure the overhead busway can actually support those low voltage DC currents.

Those architectures will look nothing like a traditional data center. So, when you think about intelligent rack power infrastructure right now, things are moving at an unprecedented pace. I would love to say that in 3 to 5 years, all that we're going to be focused on is low voltage DC architecture, but 3 to 5 years ago, we weren't even really talking about that or dreaming of it. Things can move so rapidly in this space right now that who knows where we'll be.

Learn how Server Technology high-density rack PDU solutions can support evolving AI infrastructure requirements here.