video_OCP2026_台達電_AdvancedEnergy_ORv3高功率電源架構_20261015

All right. We'll get started with the session. My name is Harry Soin from Advanced Energy, and here's my colleague >> Jihan from Delta Energy Systems. >> And together we will talk about ORV3 power architectures, you know, ORV3 evolves, you know, next generation architectures for higher power workloads. The topics we plan to cover together include you know, Jihan is going to get into the overview of the data center, the HPR, the different types of power shelves, the efficiency curves, GPU profile, CBUs, HVDC transition, or LVDC transition as we call it, and then I'll get into the detail of the unique features of these products, the pulse challenges, and and other utility related challenges we are facing, and the new features. All right. Over to you. >> Thanks, Harry. So, welcome everyone. So, just to start with a introduction over the plan. So, here you see our actual transition plan. Right now we are at the first stages, so means we have the single phase PSUs, and the shelves up to 72 kilowatts on a single phase. Next step is also what we are right now working on is the three phase 110 kilowatts, still 50 volts, and uh let's start with the 50 volt architecture first. So, here you see a small summarization about the HPR and the HPR2. So, the HPR V2 is increased from 5.5 to 12 kilowatts, and this the latest ORV3, which is in MP now. Here you see a comparison of the uh outer dimensions and how the shelf looks. So, what you can see is we have added two ports to the backside to gain more space for the PSU. And here's a little bit comparison about the features. I mean, it's just a different. So, here's the mechanical model and you see that we needed to increase the connector. So, we have changed now to 2 * 60 amps to cover the needed current capability. And also like mentioned, the PMM shape has different size. So, from four ports from the HPR to two ports to HPR V2. Here's just a efficiency curve of our actual 5.5 kilowatts PSU. And here I would like to go now to the actual topics and challenges we are seeing. So, with the increased data center capacity and the need for more power, what we are seeing is challenges on the grid side. And what we need as a PSU or power provider is to cover different scenarios what might happen from the grid. So, here you can see different um sex. So, here the AC outage is shown and how we are reacting. So, this is a example of a low voltage right through and the PSU needs to recover very fast. And here you can see it, you have a outage with 100% load, it drops to zero and we are com- coming back below 500 milliseconds, which is quite challenging, but we need to fulfill it and we made it. Here's also example. I think this is also interesting for most of you. Some examples for dynamics and EDPP. So, this is how the pulse load pattern looks like and how the PSUs reacting. So, here you have also different frequencies. So, uh Yeah. Okay. So, uh one of the main things which we are introducing and where we see the need is the capacitor bank unit. So, we see the increasing power. We see the increasing demand for pulse load is and what we definitely need and what needs to be avoided is to reflect the dynamics to the grid. And due to this one, we are developing the capacitor bank unit. So, means um we're absorbing the pulse uh pulse load is and you do not see it on the grid side. So, maybe one example of this one. On the left-hand side, you see the theory and on the right side, uh this is the real current how we have measured it. So, you see we have times where the pulse load is occurring and we are living from the caps and during the downtime, we are charging back. So, for the PSU, it does not change anything, but if you look at the AC input current on the right side with the capacitor unit, you see it's just flat up. So, you do not have the fluctuations, and this is uh one of the biggest benefits and needed to stabilize the grid. So, here's the overview of our uh portfolio right now. So, like mentioned, we have the capacitor shelf for the 50 V. So, this everything is the 50 V architecture. You have the 12 kW PSU, which is an MP now. And uh Yeah. So, now we are going one step ahead and moving to the HVDC uh step. So here you can see that we are coming from the uh 50 V one phase and three phase and we are moving to the HVDC due to the need of above 250 kW. And as you can see, we have still the ACDC and the DCDC, so the sidecar. And uh as you can see also the next steps, so what is in plan, but right now the highlighted area is where we are. And as you can see from the sidecar rack, so this is a typical AI rack and what we are seeing is uh that the rack is occupied by the AI servers and uh no room for additional power shelves and no room for BBU backups or for cap shelves. That's why we moved these to the sidecar, which is enabling uh up to megawatts of rack and just putting the DCDC on the IT rack. Here you see the uh complete HVDC overview. And as you can see, so we are going in with the ACPDU and then we have the AC to DC PSU with 30 kW and uh 30 piece in it. The lower part is a configuration which can be uh changed up to your needs. You can have a BBU, but you can also just have cap shelves. It depends on your system and on your needs. So, the shelf is the same, but it needs to be completely uh populated with the same units. And uh what we also see is we have the DCPDU on top. So means so for safety reasons, everything needs to be fused and protected, even the rear side door. So once the door is open, the complete rack will shut down. So, safety is a quite big topic right now due to >> direct current. >> This is our actual portfolio for the HVDC. So, we have the 30 kW PSU model, the 18 kW DC-DC module, and the shelves for this one. So, as you can see also, it's configurable. You have the CBU and also the BBUs. >> [clears throat] >> So, I'm handing over to Harry. >> Thank you. Thank you, Johan. Okay. All right, as we as we move forward, I So, I'm planning not to My purpose is show you the complimentary side to what Johan presented and and the evolution as to, you know, we talked about ORV3, HPR, HPR V1, and how these these platforms are evolving to address the to address the market space. So, as you're seeing on your screen, right, the HPR V1 by itself, you know, had a the connector you know, the bar clip connector in the middle. So, for the to address the need for the various markets, the enterprise customers, you know, we have developed the variations of that. You could have a right hand connector or the left hand connector. So, essentially giving you flexibility feature sets of all HPR V1, but at the same time having the flexibility to deploy, you know, as your needs might be. So, that that helps address those needs, but the specifications of the power shelf and the PSU remain the same. Same for HPR V2. So, this is really interesting, right? If you look at the HPR V2 specification, that talks about the bolted interface. So, here, you know, we have created the bar clip interface for market just to create a differentiation, give people and our customers more flexibility. So, you know, they can adapt to to these needs. You know, so you can have you know, either uh the bolted connection or you can have the bar clip connection. And in addition, the the the power supply specification had uh only LBRT. We are enhancing that with the IT type specification as well. So, you know, it's a great contribution from the community where we take an existing specification and take it to the next step. So, efficiency as you know, the whole reason we are going up is you know, we need better and better efficiency. And you know, as uh as we have seen in the past, you know, we are meeting the spec, in fact, exceeding efficiency at some points. And that becomes a good number. I mean, 97.97.5 and approaching 98%. I mean, that becomes, you know, a challenge. But I think, you know, working hard towards, you know, as high efficiency as possible as we move forward. So, now I've covered the products. Now, I'm going to get into deep dive into into the challenges and the feature set of next generation. Before I do that, let me let me talk about little bit about, you know, you saw Jihad's presentation on the sidecar. So, so essentially, when you take as the need for the power grows in the AI infrastructure, what has happened now is, you know, the power is segregated in the Mount Diablo. So, this is down conversion. To make it more efficient, you know, there is a down conversion that that takes place within the IT rack, right? So, to address that, HPR V4 came into play. It was introduced last year. So, at Advanced Energy, we have uh 100 kW power shelf that takes uh plus and minus 400 or 800, converts that into 50 V. Uh we heard about OR uh so we heard about uh that went forward too fast. So we heard about OR the the wide rack this morning, right? So in order to have even higher power, uh you are now looking at a wider rack which uses 11 same PSUs as you saw in the previous slide in the 100 kW. 18 kW, 11 PSUs giving you a total of 160 kW of power. So, you know, these becomes building blocks for your 1 MW racks as you move forward in higher and higher power. So now, I'm going to deep dive into the features uh just to show what the challenges we are facing and how we plan to address them. So 72 kW, the pulse load challenge. What you're seeing in your screen is the load changes from 30% to 160% of the load. Now that is a huge change in load. And on the right, you're seeing going up to 175% of the load for 2 ms. That That's imagine that's almost your load has doubled, right? And and if you look at the current rate at the bottom, you are seeing three phases of input current. And you know, you you don't see almost anything there. It's not noticeable. We are well within the specifications and we are able to address using, you know, I I dynamic performance of the PSUs uh all these challenges. Now this is interesting. Uh Gihan touched on that and he showed you the waveform. So I'm just going to a little bit deep dive into it. You know, as the data centers are rapidly expanding, so what's happening? The power is increasing. Not only increasing, it's increasing significantly, right? So what's So utilities are getting extremely concerned. So they are putting forth a lot of regulations in place. You know, what you don't want to do is like a typical legacy power supply will trip. It will cause a breaker to trip. It will cause cause all kinds of issues downstream. So, those utilities do not want that. And so, the challenge we face now is how are we going to implement it that you know, when there is a voltage sag event or a voltage surge event, the as you see at the bottom that the data center should maintain its function perfectly and there should be no change. So, that's a challenge we started to address. You know, we call it low voltage right through. So, I'm just going to talk about the challenge in little bit of a detail as to exactly what it means, right? So, what happens, you know, you have a PSU, you have a BBU. What you do not control is the distance between the two, right? And you know, PSU and BBU do not share current. I mean, there is there is no direct current sharing path. The only way they share current is through droop, right? That's the only mechanism they have. So, what are we looking for, you know, as you know, as the voltage decreases, we have a voltage event. So, what's going to happen, you know, your power supply is going to come to a power limit. What do we want to happen? We want the BBU to help fill that gap. So, the numbers you're seeing the 22 whatever is just an example to show when the PSU power goes, you know, from 70 to the PSU load power shelf load goes from 72 kilowatts to 50 kilowatt, you want the BBU to come and fill in that gap. That's that's exactly what we're looking for and that's the challenge we are faced with. All right, so I'm going to get into the next slide to show how we address that. Oh, went too fast. No, that's right. So, what we need to do is in terms of a solution, we need to detect it. We need to ride through it. So, ideally what we are expecting is, you know, that the PSU power should go down, the BBU should gradually take over. I mean, that is the ideal thing and that's exactly what you're seeing and that's with the intelligent PSU current, you know, current limiting. So, essentially, you know, you want to limit the power supply current so the PSU is safe, it doesn't cause an event, and you want BBU to gradually ride through and and take over that extra load. And that's exactly what we achieved by intelligent power current limiting combined so that we are able to address it. And so, in the next slide, I'm going to show you the real example of that. So, let's start with the base case, like there's nothing In this case, you know, you have your power shelf and the example I'm sharing here is you have a power shelf and you have a BBU shelf very close to each other, right? And and in this case, you know, you're seeing a successful ride through. You're seeing the voltage from 230 V AC 2.45 or 103 V, the voltage drops for 350 ms. So, the PSU power goes down. Right in the middle screen, you're seeing the green. So, the BBU is taking load. Nothing happened. It's a successful ride through and the PSU, and as Jian pointed out, the fast recovery of the PSU should come back within 500 ms when the event disappears. All right, here. Again, the base case, all right? So, here we have what we have done is the BBC droop regulation was adjusted lower. We said imagine they actually got far away from each other. So, what happens in this case? The PSU in this case, if there was no intelligent control, will shut down, right? That's not the behavior we want. So, that is what happens if you don't have those intelligent mechanisms built in to address this challenge. So, here is a successful example when the LVRT upgrade function when the when the power shelf of the PSU was upgraded with this functionality and you are seeing a very smooth transition. In fact, BBU is taking a more power, PSU reaches its current limit and is happy and it's it does its job, stays online and when the event disappears, uh 500 microsecond right through and you know, everything is okay and we come back to normal operation. So, you know, sounds trivial, but I think all of us have worked really hard to, you know, address this very important challenge, uh you know, LVRT or low voltage right through, which was really an important step in the right direction and we're going to see as the power goes to, you know, one you know, 1 MW racks, 2 MW racks, this becomes equally more important and more stressful and we need creative ways how we address it. Uh just coming towards the end, slew rate challenge. You know, most of you are familiar with slew rates. So, as we have gone from 33 kW shelf to 100 kW shelf, our power shelf slew rates expectations have increased from 6 amp per microsecond to 30 amp microsecond. What does that do to us, right? So, having those uh uh slew rate challenge if that happens, if your load is varying that much, you're going to have a voltage spikes uh and that is something nobody wants. That's an abnormal behavior. So, how do we address it? I mean, this is where we need a very creative uh low inductance bus bar, uh proper bar clip, uh back panel design to make sure and the dynamic performance of your power supply. So, interactively together, we are able to address this challenge and uh you know, very successfully address uh the slew rate challenges that we all face. So, you know, so essentially I wanted to give you an update on the peak load challenges, LVRT, and the slew rate which we all face as, you know, critical things as the GPU loads are undefined, their profiles keep on changing, the right sort of expectations keep on changing. So, we have to be on our toes so to address these challenges. So, I think that brings to most of the parts and at the end I'm just going to talk you through the offerings we have at Advanced Energy. You know, we've been in this journey for a long time. We started with the 18 kW shelf and on the right you see 33 kW shelf, moving on to 72 kW shelf. We have the complimentary 3 kW PSU, 5.5, 12 kW, and not only the ORV3 versions, but also the versions for the market for with a different bar clip options. Right and left position connectors. So, you know, so we are able to address because everybody's needs are different. So, we are able to address the needs of the market and help address those challenges. And as we move, you know, we heard a lot about the Mount Diablo as we move towards Mount Diablo or or power desegregation, we have the LLVDC PSUs and power shelves. We have designed the 18 kW PSU which works with the 100 kW shelf and 160 kW shelf. So, that brings me to the end of my section and call to action as I think it's equally important. We need everybody to be involved and the only reason we have these variations is we hear from the community and our monthly calls, people tell us that we want to see this, we want to have these changes in the PSU, we want to have IT compliance and that helps us evolve and get better and you know, help improve the specifications. So, I think that concludes it and uh Gihan and I'm here to help answer your questions. >> Hi, I just have one question. My name is from Vern. We You showed some architecture earlier about there is no more space in the IT rack for the PSUs, which is why it's been moved to the sidecar at moment. >> Mhm. >> So, uh I know that the BBUs that are incorporated in the sidecar is ultimately help with the sort of the LV right through and the backup. But, what about the AIS moving because as we know the the rapid fluctuations will affect the grid compliance and things in Europe first, for example. So, I don't see anywhere in these both in the sidecar and in the computer rack where there is CBUs involved. So, what is what is the the plan for that? >> That's a great question. I think Gihan, you had your presentation. You talked about CBUs. And again, those are individual work cases. There are multiple architectures evolving, you know, and we had an earlier presentation. Those capacitors are best close to the power supply. So, there are various architectures that have been designed either you have a CBU shelf >> Okay. >> or if your workload is different, you have the PSUs as a as you have those capacitors as a part of your PSU to help address it. >> Okay. >> So, we haven't forgotten. And plus, for your sidecar, there are multiple energy shelf solutions evolving that combine BBU and CBU functionality together. >> Okay. Thank Thank you for that. >> Sure. >> Hello. I have another question. BBU and CBU, they are taking care of LVRT. I mean, it's good. But, when the AC voltage is back, the load must be restarted within 500 milliseconds. >> Mhm. Correct. >> How do you achieve it? I mean, I feel that's a wasted part compared to LVRD. May I know your opinion? >> You want to take that? >> Yeah, sure. So, actually, what [clears throat] we have implemented is we are not turning off the PSUs as Harry also shown. We are sharing the load. So, means the PSU is still operating. But nevertheless, we need to minimize the inrush or the effects on the grid. That's why we keep the PSU alive. So, it's easier for it to ramp up. But nevertheless, we also have the random times. So, we need to make it a little bit variable. >> Does it answer your question? >> Uh I mean, maybe I didn't get it completely. >> No, but what he's saying is if the PSU shuts down, then coming up in 500 milliseconds will be challenging. And I think what he explained is the PSU stays keeps on goes to current constant current limit, right? >> Exactly. >> Doesn't shut down. It can come back within 500 milliseconds. >> Okay. So, in the meantime, the PSU will be powered by S2 AC power supply? Let us say S1 S2 AC sources. So, do you keep the PSU on using S2 source? Or do you keep the PSU on using inbuilt using a bulk capacitor energy? >> No, your PSU stays on using a combination of the bulk's capacitance as well as the AC source because remember AC source has not disappeared. >> No, in a in a >> shuts down then comes here, right? >> Yes. >> So, it uses the hold up there at that time, correct? >> Correct. But you do not have two two feeds in this system. We have S1 S2. So, we have only one feed. There is no ATS functionality. >> Okay. >> So, means uh the grid is gone, but we still need to survive. >> Okay. Got it. Thank you. >> Sure. Good question. Thank you. >> Oh, hi Harry. Quick question for that PSU BBU droop regulation. What is the limit? You mentioned like 100 mV, right? >> We See, that's a that's a very good question. There is no limit. Right? Because in a data center, you do not define that your BBU could be far away, but you should be able to intelligently control your PSU current and sharing with BBU. So, there is no defined limit. It's just droop sharing. >> But there is um oring to block a reverse current between that line. That would be like a what? 0.7, 0.6 V? >> Yeah. >> So, that that's probably the limit, right? You cannot exceed the uh the the diode voltage, right? >> Yeah. We're looking at the Anything you want to add, Jian? I mean, you know >> I actually agree. So, you you want to have a low impedance busway because otherwise, like you mentioned, you have the voltage drop. >> Right. >> But at the end of the day, so it's compensated. So, the sharing is still there. Yeah. >> Right. >> the balance. >> Yeah, related question is uh that oring right now. Uh Is this like discrete Schottky or you using some ideal diode? For the >> have active oring. >> Active oring? >> It's active oring. >> That's part of the design, active oring. >> a very low drop, right? >> Oh, yeah. Absolutely. Absolutely. >> All right. Thank you. >> Sure. Thank you. Good question. >> To be clear, when you say constant current, you're talking about input current. >> That is the constant input current. That is absolutely correct. Not the output. Exactly. That's what you want to maintain, right? Exactly. So, you don't trip. Yep. >> And then um one thing that I haven't seen yet is any discussion on the 12 kW battery backup. Is I know it's still being developed, but do we have a timeline on that? When something might be published, Harry? >> Uh yes, good It is ready and it's being worked on. I I know I'll put on top of my list to have it published and and shared. >> Okay. >> Point taken. >> Thanks, Harry. >> Sure. >> Hi, I'm Chen from I nines. Very interesting presentation. Just so you know, I worked in Delta in the past as well. Quick one. When the PSU and the BBU are sharing the powers, is there any communication between the two units? Or this is just a still like the droop control? They are just the detector. >> As far as there's absolutely no communication. They are That's the best part. They are separate. There's only droop. That's it. >> So it's only the control algorithm in the PSU which is looking at the current and load. And the BBU will detect when it should step in to take over. >> That is correct. >> Okay. Thank you. >> All right. Well, thank you. Thank you. >> [applause]