Description

In this week’s Power Corner, James Hitchcock, Vice President, General Manager, and Managing Director of Tektronix’s EA Elektro-Automatik, discusses the recently released EA-ELR 21000 Dynamic Test System. The Dynamic Test System is Tektronix’s solution for coordinating multiple ELR 21000 regenerative electronic load units so they act as one synchronized, rack-scale system rather than a collection of independent loads. These are paralleled together via an arbitrary function generator and dedicated synchronization lines that ensure every unit transitions in lockstep. 

The system is optimized for 240 kW at 12 A/μs  slew rates (at 8 x 30 kW) and scales to 1.9 MW, for headroom with next-generation sidecars. Slew rate scales with the number of units paralleled. The same ELR 21000 unit works as a standalone benchtop load at low power (30 kW) and scales into the full dynamic rack system. 

Linear vs regenerative loads

Linear electronic loads, or resistance load banks, dissipate energy as heat, while regenerative electronic loads take energy from the device under test (DUT) and return it as electricity. This is particularly relevant for burn-in tests, as the load is static (versus dynamic) to ensure product quality. Hitchcock explains: “For a megawatt-class burn-in test where you’d set up a piece of equipment and continuously operate it for several tens of hours, if not hundreds of hours—our customers can expect savings upward of half a million dollars per year in energy savings doing long-term burn-in testing at that megawatt scale.” 

The technology, supported by SiC WBG technology, uses both a resonant converter and power factor correction (PFC) circuit for two-way energy transfer (bidirectionality) and grid-side synchronization, respectively. 

Power shelf and sidecar validation

Hitchcock separates regenerative load testing into two categories with different needs: the static, steady-state burn-in stress test and the high-duty-cycle, fast-transient dynamic test. Data center sidecar testing involves dynamic testing in which the ELR 21000 targets the Open Compute Project (OCP)- defined envelope of what the power bus experiences. 

Hitchcock notes that early attempts to meet the Mount Diablo Open Compute Project (OCP) specifications used linear electronic loads for dynamic testing because they optimized for very fast switching at low power. However, scaling to 200 kW and beyond proved difficult because it required extensive engineering to extract heat from the test system.

Hitchcock says the ELR 21000 platform avoids the synchronization and cooling problems that come with stacking benchtop loads. Hitchcock continues: “In dynamic testing, we provide arbitrary functions to the load, to the power shelf, to observe its performance under different conditions, and the purpose of that is to characterize how it will perform and predict its suitability for use inside the data center.” 

Transcript

AS: Welcome everyone to Power Corner. I’m your host, Aalyia Shaukat, editor-in-chief of Power Electronics News. We’re sitting here with James Hitchcock, Vice President, General Manager, and Managing Director at Elektro-Automatik. Hi, James. Good to meet you. Thank you for sitting with me.

JH: Thank you for having me, Aalyia. It’s great to be here.

AS: No problem. And we’re going to be talking about regenerative electronic loads and the new EA ELR 21000 Dynamic Test System. But before we even start on that, do you mind talking a little bit about your background at Keithley, Elektro-Automatik, and then Tektronix?

JH: Absolutely. My background has been in scientific instrumentation and test and measurement for my whole career. I moved from life sciences to the test and measurement field in 2020 when I joined Tektronix as the general manager of Keithley Instruments. 

Tektronix owns several brands—one of those is Keithley Instruments, and another is EA Elektro-Automatik. We acquired EA Elektro-Automatik in 2024. We became interested in the solutions that EA provides through our work with automotive EV customers on the electronic powertrain. 

My focus has been on the power storage and energy management side through precision DC on the Keithley brand, and now through higher-efficiency regenerative solutions as we scale up in power with the EA Elektro-Automatik brand.

AS: Thank you for that—that helps give a little bit of background. Maybe we could start with the basics: what is a regenerative load, and how does it differ from resistive load banks?

JH: A regenerative load takes the energy from your device under test and returns it back as electricity. That electricity can be used within a facility to power other equipment, or returned back to the grid if needed elsewhere. A resistive load bank takes that energy and burns it as heat, so you end up paying for it twice. First, you pay for the electricity in your test, and second, you pay to remove that heat through HVAC systems.

AS: Yeah—so more power put into just cooling it. Can you talk about the underlying technologies that enable regenerative power? What power topologies, what semiconductors are we using, that enable bidirectionality in these test systems?

JH: I’ll start with the semiconductors we’re using. EA was a pioneer in the adoption of silicon carbide technologies inside our equipment. We learned alongside the automotive EV customers on the advantages of high-speed switching inside silicon carbide devices, where efficiency gains moving from silicon can be quite dramatic. 

Anyone who’s familiar with the transition over the years in electric vehicles will be familiar with the early electric vehicles that were out as early as the 1990s, which had severe range problems and batteries being damaged over time. 

A lot of that was because of the power topologies, the electronics, and silicon that were being used at that time. We simply didn’t have the technology ready yet. The silicon carbide capabilities in the semiconductor industry really unlocked this new wave of cars that we’re seeing out on the road. And EA learned right alongside, partnering with our customers, like we always do, to develop the high-efficiency test solutions as our customers developed the power electronics.

Those silicon carbide chips are used inside a core technology we call a resonant converter. It’s what allows for the bidirectional and high-efficiency regenerative capability. The resonant converter is a circuit that matches the load, or the device under test, to the power electronics inside the electronic load, so there’s no mismatch in moving the energy from one to another. If there’s a disconnect—just think about trying to drive across a curb, or take a step—you want that to be as flat as possible, so everything moves very smoothly from one side to the other. And that’s what we call a resonant converter. It’s the big difference between a resistive load and a regenerative load. 

The second piece is on the AC grid side, which is the power factor correction circuit—that circuit monitors the electric grid, the local electric grid, and ensures that we match both the frequency and the phase of the electricity on this grid, so that as we move that energy back onto the facility grid, we don’t impact or affect any other equipment that is nearby. It’s a very important part of the regenerative technology that unlocks what we’re able to do.

AS: Thank you so much. Do you mind talking about the power savings that are associated with a regenerative power supply? Do you have specific examples and numbers?

JH: As we already said, in a resistive load, every kilowatt you burn is heat, and you pay for that over time. With efficiencies in the 95 or 96% range—depending on whether you choose our industrial loads or our high-dynamic loads—that savings can really add up. 

The actual savings completely depend on the duration of your test, how often you’re scaling to high power and low power (we call it the duty cycle), and the cost per watt, so it’s very unique to every user. We have a return-on-investment calculator on our website that our customers quite often use. But as a general example, for a megawatt-class burn-in test—this is where you’d set up a piece of equipment and continuously operate it for several tens of hours, if not hundreds of hours—our customers can expect savings of over half a million dollars per year in energy savings doing long-term burn-in testing at that megawatt scale.

AS: Well, that’s a pretty significant savings; thanks for sharing that. I understand that regenerative loads are used in industrial testing facilities a lot. So I wanted to know about the data center applications. How is the ELR 21000 built to validate 800-V DC data center power? Can you name specific examples?

JH: Our ELR 21000s are built on our industrial platform, and with that, it brings all of the reliability and scalability that our customers have known. What’s changed—or what we’ve designed and innovated—is the synchronization of the different units for speed. 

The Dynamic Test System, which is the solution that brings the ELR 21000 together and scales up in power, uses an arbitrary function generator and synchronization lines to ensure every unit is transitioning together. Similar to how a boat crew rows in unison, and the speed of that boat will determine how synchronized each rower is, the Dynamic Test System synchronizes our ELR 21000s to ensure the high-speed testing requirements for the 800-V DC architecture.

AS: And I imagine that has to do with the fast load transitions of dynamic AI load—their load profile. Do you mind talking about that in a little more detail, how that synchronization helps any particular test setup?

JH: Absolutely. In an AI compute rack, when you’re doing training, or when you’re prompting or running an agent, those AI chips are switching very, very quickly, and hundreds or thousands of them will activate all at once. When that happens, the 800-V power architecture that is powering all of this will experience fluctuations in the loading on its power source—we call this load transients. 

Now, the load transients aren’t as fast as the AI chips that are switching, so we’re not trying to emulate the AI switching itself. But we’re working to a standard that has been defined by the Open Compute Project for what to expect on an 800-V DC bus for the newest generation of AI chips. This standard allows the power shelf and power sidecar suppliers to test and qualify their designs and their suitability to go in there. So what we’ve provided on the Dynamic Test System is a set of applications and functions that can be performed to certify the performance of a sidecar or a power shelf.

AS: And just to give an example there—I saw that the ELR 21000 offers current slew rates of 12 A/µs at 240 kW. Why is that slew rate dependent on the number of parallel-connected electronic loads? Are they close to the edge of what you’re seeing in the field? And you were saying there were some OCP standards that you were in alignment with.

JH: Yes. The Dynamic Test System is optimized for 240 kW at 12 A/µs, as you said, and the reason is that’s the most popular power shelf on the market today. But that is not where our customers are stopping—many customers have announced megawatt-scale power sidecars that are coming to market. We see that quite often in the news, and several engineers that we work with are pushing power levels even farther. 

So the 1.9-MW capability of the Dynamic Test System provides headroom for engineers to continue to optimize and increase their power sidecar rack design. The Dynamic Test System gets faster the more units you run in parallel, similar to how your boat would speed up the more rowers that are being placed in it and working in unison together.

So for a single high-speed ELR 21000 unit, we operate at 1.9 A/µs on the bench, and power engineers testing individual power supplies within the shelf use those single units to ensure that the components going into the larger designs are each operating successfully. Then, as our customers’ needs evolve and change, for a small power shelf at 120 kilowatts, we can operate with several units—four or five units—and for 240 kW, we operate with further units, continuing to scale and adding units as needed to meet the power demands in the future.

AS: You did mention testing the components that go into the larger design. What components do you mind elaborating on?

JH: Oh, absolutely. The power sidecar, or shelf, is made up of multiple power shelves—a power sidecar is a number of power shelves, and a power shelf is a number of power supplies, and a power supply breaks down into individual semiconductor designs and components. At the semiconductor power electronics level, we often engage with customers with our Tektronix scopes, our IsoVu probes, and our Keithley SMUs, source-measure units. Once the equipment is fully built into a power supply—which is the smallest component that’s sold on the open market—that will range in power from about 10 to 20 kW, and this is where our 30-kW  high-speed ELR is very useful for our customers.

Each power supply is checked for its speed and stability before being placed in a power shelf, which will contain somewhere between 3-6 power supplies, depending on the manufacturer, and that power shelf may operate in the 100-kW  or higher range, depending on how many power supplies go into it. Then, by stacking power shelves along with battery backup units (BBUs) and switching capabilities, we arrive at the power sidecar, with full control over that 800-V megawatt-class power that gets transferred to the compute rack when needed.

AS: And when you talk about specific tests, are we primarily talking about burn-in tests, or are there more tests involved?

JH: There’s a class of test considered burn-in and quality testing, which is a very stable test. Here, customers often use our industrial power supplies or electronic loads to get that extra bit of efficiency. We don’t need to change or swing the amount of load on it for burn-in; it’s just sitting, operating, typically at an elevated temperature, to ensure the quality of the product. 

And then there’s dynamic testing, another class of testing that is standardized through the Open Compute Project. In dynamic testing, we provide arbitrary functions to the load, to the power shelf, to observe its performance under different conditions, and the purpose of that is to characterize how it will perform and predict its suitability for use inside the data center.

AS: Thank you, that’s very helpful. I noticed on the data sheet there’s a 95% efficiency at 100% power and 100% output. I know that AI workloads will switch between low load and full utilization—what efficiency metrics are we looking at for the lower loads?

JH: The 95% efficiency is an average between the power conversion efficiency—that path we talked about earlier, of moving energy from the device under test to the electrical grid—and the power overhead of running the cooling systems and the control systems. So, there’s a combination. As the power draw for, say, the compute rack not doing AI computation reduces, the power conversion efficiency remains essentially constant. But the average of the energy being used for the control and the cooling does slightly increase, meaning our efficiency does come down a few percentage points at low power. 

However, at low power, the heating-effect and the power-loss-effect are actually quite small. So for our customers, the real concern is when the power scales and increases. The losses generate heat, which damages equipment and requires large spacings to avoid interactions, and requires extra cooling at high power—is why the systems are designed to be so efficient at full power.

AS: So that 95% efficiency—that’s the average of the input to output—and what was the other metric?

JH: The power overhead: the power required to maintain cooling and control of the system.

AS: Do you have any metrics regarding the efficiency of the regenerative capabilities specifically?

JH: I’m not open to sharing the internal efficiencies—only the overall averages of the system.

AS: Gotcha. So then the next question would be: what can a rack-scale regenerative load system do that a collection of smaller benchtop electronic loads can’t, for a team that’s testing at lower power today?

JH: Well, the beauty of EA’s electronic loads is they’re modular, and the majority of our customers use the ELR 21000 High Speed as a benchtop load when working at low powers, and they can use that same exact equipment scaled up inside their Dynamic Test System. 

But for customers who are using more traditional benchtop test and measurement equipment, that equipment is traditionally what we call a linear-style electronic load, which is resistive and optimized for very fast switching at low power. These systems can be stacked inside of a larger solution and controlled through automation software. However, those systems are not optimized for synchronization, and it can be very challenging to ensure that every single one of those systems is coordinated when switching. 

In addition, the heating generated by those solutions requires extensive engineering to extract, often through water cooling and additional solutions. So customers prefer a dedicated solution optimized for that testing at higher power, rather than designing benchtop equipment into broader racks.

AS: Is there a customer today asking for that full 1.92-MW configuration, or is that more future-proofing? And what’s the constraint that stops it from going further than that 1.92-MW rating?

JH: I mentioned earlier several announcements for 1-MW class, and the majority of customers announcing new power sidecars are right around 1.2 MW. I won’t talk about any research or engineering that’s going on, but it’s safe to say that power levels are continuously increasing, and engineers are hard at work to find ways to increase the amount of power available to our data centers inside the rack. 

EA Elektro-Automatik is constantly engaged with these innovators to scale up in power. For our customers who need more than two megawatts, we often supply our 60-kW industrial bidirectional solutions, which are scalable further to 4 MW today. I don’t know what the future holds—we are under constant innovation, but we’re focused on staying in lockstep with our customers and providing the power levels required as they increase.

AS: So this is directly, as you’re saying, in lockstep with the industry—you’re getting industry feedback and working to those specs.

JH: Yes, exactly.

AS: So you mentioned the OCP guidelines—do you mind talking about that? Is there a sort of de facto test methodology? What does it look like, and do any of these test methodologies include regenerative loads in the standards practice?

JH: Yes. The industry recognized this challenge a few years back, when power supplies and power shelves coming into the data centers were having different levels of performance depending on where they were sourced from. It was really important for the hyperscalers to be able to get consistent power delivery across multiple vendors. So together, under the Open Compute Project, they published a specification referred to as Mount Diablo, which is a mountain in Washington. This sets the guidelines for how to test the power shelves and power sidecars. 

It does not specify any type of testing equipment or what to use for that. In fact, the first testing attempts for Mount Diablo specifications were primarily linear electronic loads. Linear technology does have a significant advantage in speed, so as customers looked at these new speed requirements, they immediately turned to that technology. Shortly after scaling past 200 kW, they recognized a significant problem in the heating of this linear technology, and were requiring almost as much investment in the cooling systems as the testing systems to continue to be operational. A new solution was needed. 

That’s when EA started really working in the industry; we’re a specialist in very high power, and experts in hundreds-of-kilowatt systems and megawatt systems. So this is when we started designing the Dynamic Test System to solve this heating challenge when the power shelves and sidecars exceed 100 or 200 kW. That’s what allows the customers to test to the Mount Diablo guidelines, even though regenerative loading is not specified in those guidelines.

AS: That makes sense, thank you so much for all of this information, James. I really appreciate it, and thanks for answering all my questions. I was wondering: is there potential to use electronic loads or regenerative loads for testing other data center systems as well? Do you mind describing some of that?

JH: Absolutely. Connected to the 800-V DC infrastructure are battery backup units, energy storage systems, solid-state transformers, and several other electronic equipment designs. Each of these designs needs to be able to handle the load transients on that 800-V bus—depending on the location of those systems, the transients can be lower than the specifications set by the Open Compute Project. We work with each customer and each product to understand where they are in that power architecture and recommend the appropriate equipment.

AS: And would you say that the power shelves and the sidecar racks experience the most intense transients—potentially more than the others?

JH: Absolutely. The power shelf is directly connected to the AI compute rack, and it observes the largest transient. One of the specifications out of the Mount Diablo paper is that the power shelf must withstand the turn-on, turn-off transient without activating the battery backup unit. 

The battery backup unit is the next piece of equipment connected into the system, and it’s there to ensure that if the power supply doesn’t fully control itself with that transient, it will activate and protect the compute process. It’s also there in case there’s any power disruption to the sidecar itself—it can run the sidecar for several minutes while a workaround is identified. 

It doesn’t take the place of the energy storage system, which is much farther back in the infrastructure, and an energy storage system will not see the large, high-speed transients that the battery backup would. But as you move to something like an energy storage system, or even a battery backup unit, bidirectional capability—versus just loading—becomes really important. Because we need to be able to charge the batteries as well, and test energy moving in and out of that system. Our industrial bidirectional solutions become well suited for that.

AS: That’s very helpful to know. I suppose, same with solid-state transformers—the bidirectionality is built into a lot of these systems.

JH: Solid-state transformers are also a bidirectional technology. They connect directly to the medium-voltage AC grid and bridge to the 800-V DC power infrastructure. They’re a critical piece of technology that’s being developed right now to further increase the data center, but likely a whole other talk would be needed if we wanted to get into that technology.

AS: Gotcha, no problem. All done here—please let me know if there’s anything I’ve missed, anything you’d like to say, any last comments.

JH: I want to say thank you for having me on the podcast—it was a joy speaking with you. I’m always thrilled to talk about our customers and our innovations. It’s a very exciting time in the AI power infrastructure space, as we’re solving multiple problems that allow us to scale our technologies in ever-increasing power.

AS: Thank you so much, James. Thank you.

The post Power Corner: Testing Data Center Sidecars with Regenerative Loads appeared first on Power Electronics News.