Description
800-to-6-V proof-of-concept design
At Electronica 2026, Renesas’ Pietro Scalia, Head of Power System Marketing and Architecture, covers Renesas’ recent proof-of-concept 800-to-6 V intermediate bus converter (IBC) reference design, which, as shown in Figure 1, fits the standard form factor (400 × 60 × 6.5 mm, excluding the busbar) that OEMs and hyperscalers are targeting for 48-V, 12-V, and 6-V output. The reference design reaches 2.1 kW/in³ and nearly 96% efficiency, which Scalia calls “a stretch at this voltage.”

“I think we came maybe a couple of months later than others, but on the other side, we are differentiating a lot by introducing more density and a more compact form factor,” says Scalia. The design switches at >500 kHz but below the ~1 MHz of 48 V designs. “We also have to be realistic that at a 6-V output, we had to scale down the frequency,” says Scalia.
The 20-kW, fixed 128:1 ratio IBC debuted at OCP 2026 and is built from four identical 5-kW full-bridge LLC blocks interconnected in input-series, output-parallel (ISOP) and input-parallel, output-parallel (IPOP) configurations (Figure 2).

Magnetics
On magnetics, Scalia was candid: “The reason the losses are concentrated in the transformer is that it’s not our native territory.” According to the loss model in Renesas’ white paper, the transformer accounts for ~74% of the converter’s total losses, with transformer copper and shielding alone making up ~64%. Renesas co-developed the four-limb planar matrix transformer with Payton Planar Group, but Scalia stressed that it is “not only Payton,” pointing to work with Sumida and Würth as well.
Scalia notes that the next step is encapsulation: embedding components (e.g., FETs and capacitors) inside the PCB or underneath the transformer instead of mounting them as discrete parts. This effectively shortens current paths and further optimizes both efficiency and solution size, as Scalia points out, “ the more things we can embed inside the PCB…the better.”
HV GaN and LV Si
Scalia also discusses the power devices used in the design. The primary side uses 650-V GaN in dual-side cooled PQFN packaging, which is key to meeting the converter’s 6-mm height (versus TOLT packages). And while the primary GaN accounts for only 6.7% of modeled losses, Scalia argued that thermal headroom, not just RDS(on), is what enables 20 kW of rated power.
With liquid cooling “typically swinging between 55°C and 65°C,” the package dissipates heat through its bottom exposed pad and, critically, through its top directly into the cold plate. “It’s not only the RDS(on) of the GaN, but the way you dissipate,” he said. Scalia notes that this headroom is critical because the electrical design current (EDC) can be far higher than the thermal design current (TDC).
On the secondary side, silicon handles the low-voltage socket, with 32 of Renesas’ 25-V QFN3×3 silicon MOSFETs per 5-kW cell (accounting for 12.6% of total losses). The industry has largely converged on LV Si FETs for the secondary side of the 800-to-6 V designs. Scalia does add the potential for LV GaN to be deployed on the secondary side; however, “40 V is probably the boundary today, Scalia says. Below that, GaN’s performance “is not yet where we wish,” though Scalia expects the crossover to come in “quarters, not years.” For now, he summarizes: “At 50 V, we see the value [for LV GaN]. At 12 V, we might see it. At 6 V, not yet.”
Renesas has already signaled that dual-side-cooled low-voltage GaN is coming for the secondary side. In an earlier interview, Adil Salman, principal product marketing manager at Renesas, said that “is happening very soon,” pointing to the company’s partnership with EPC to enable “a full-fledged IBC solution eventually.”
When asked about the timing, Scalia first clarified the nature of that partnership. “It’s been a licensing agreement,” he said, explaining that Renesas is incubating the technology at its Naka factory in Japan, a step taken after TSMC announced it was exiting the GaN foundry business it shared with Renesas. “So, in short, Renesas owns the process,” he said, adding that the two companies continue to collaborate on co-packaging.
Transcript
AS: Pietro, thank you so much for meeting with me. How’s the show been so far?
PS: Fantastic. First of all, thanks so much, Aalyia, for having me today. The show is going magnificently. I think there’s really good momentum for Europe. And for Renesas specifically, we are having a lot of visitors and a lot of interesting threads these days.
AS: We’re going to talk specifically about the high-voltage IBC, the recent 800-V-to-6-V release. Could you talk about that in more detail?
PS: Sure. By the way, we are showcasing the solution at the booth, so a lot of customers are coming to see the density. This is 2.1 kW/in³, so a very impressive density. In terms of efficiency, we get very close to 96%, which is a stretch at this voltage. I think we came maybe a couple of months later than others, but on the other hand, we are differentiating a lot by introducing more density and a more compact form factor.
AS: Thank you. Let’s talk about topology. Why are most vendors going for the input-series, output-parallel (ISOP) topology with 650-V GaN on the primary side? Why not a higher-voltage technology, and why does GaN seem to win in these designs across the board?
PS: Very good questions. Let’s start with the switching frequency. In this isolated DC-DC, from 800 V down, we have been using a very exotic switching frequency, up to 1 MHz. Now, we also have to be realistic that at a 6-V output, we had to scale down the frequency, so maybe silicon carbide might become an option.
On the other side, think about a hyperscaler or an OEM that drives the spec across different suppliers. They try to have a uniform solution that can work at 48-V, 12-V, and 6-V output. Because at the end of the day, there is not really one solution that fits all.
It depends on the specific SoC or xPU demand. In that logic, I think this architecture is really on the spot. It can fit different approaches, and with the technology in 2026, GaN is definitely the sweet spot. That’s why the OEMs and hyperscalers have standardized the topological approach. If you visit different booths here, you’ll find similar solutions.
AS: Let’s talk about the three main high-voltage IBCs: 48 V or 50 V, 12 V, and 6 V. Where does each one fit, and what are you seeing?
PS: So you would like to see where the technology fits?
AS: Yes.
PS: First of all, 50 V, 12 V, and 6 V require different switches on the secondary side. We see a big transformation: silicon is leaving more space to GaN.
Clearly, on the secondary side at 100 V, GaN is already dominant, and I would say 40 V [GaN] is probably the boundary today. So at 12 V, you might have different vendors using silicon or GaN. We try to play GaN in a way that goes toward the customer’s spec. Typically, with GaN, you can reduce size, or you can increase efficiency. Sometimes we play both levers.
If I go back to the topology, the topology on the primary side is pretty much standard. Now, we have introduced this PQFN, which addresses the thickness. The new spec for 6 V is 6 mm, so a very, very low thickness.
The first one everybody did was 12 mm, with a 50-V output. TOLT was 2.5 mm, and thanks to the QFN, we scaled down. So this 20-mΩ QFN, 1 mm tall, is very important for this kind of compact packaging.
AS: Thank you so much, Pietro. I know that Renesas recently released a white paper on this.
As you were saying, the efficiency stands at 96% at the full 20-kW operating point, which is a total of 764 W of power loss. The 800–to-48-V DCX runs above 98%. Could you talk about why the change in efficiency is still a net win at the system level, and where those gains come from?
PS: Yes. First of all, there is a combined effect of this DC-DC and the low-voltage multiphase, which has become vertical in the meantime. Clearly, the market is transitioning toward vertical power due to the excess of load.
And going down to 6 V, we can leverage technologies that are no longer 30 V or 25 V, so we can accelerate efficiency on the core side. So it’s a combination of the DC-DC efficiency at 96% and another efficiency on the core, which depends a bit more on the power you get at the end, because there are xPUs that consume 600 W and others that consume 6 kW. You need to support them all.
In that logic, the multiphase increases the number of phases. We have systems with up to 144 phases. You need to support this amount of current, and as we debated this morning, it’s about 40 A TDC that we can do with one phase.
I wish to talk about the way you achieve 20 kW, for example. We have a 5-kW building block, and now it’s more complex than in the past. Now the trick is—let me go a little more into the topology—now, you have to convert 128 to 1. How do you do it?
You play with different elements. First of all, it’s IPOP first, so input-parallel, output-parallel, and then two 10-kW stages in ISOP. In this way, you can divide the transfer ratio across all the stages. The planar transformer is also a key element here, because it has four ports. You can inspect our way [design], where the turns ratio becomes extremely low; theoretically, it can be 2:1 or 1:1. Of course, this is proprietary, and there is a lot of IP there.
The planar transformer is essential for this kind of matrix transformer, and partnerships with the important players on the manufacturing side are very, very key for our final customers.
AS: Thank you so much.
PS: Sorry for the long answer.
AS: No, I think that was a very good answer.
Let’s talk about the matrix transformer in a little more detail. There are four 1.25-kW transformers, four limbs each, 25 V per limb. Why that partitioning, and what did working with Payton Planar add? I believe they worked on the magnetics.
PS: Yes. Let’s start first with the topology. At the end of the story, you need to bring the conversion where it is convenient to do from an efficiency perspective, and that’s the game we have played in the way of dividing. At the end of the story, the real turns ratio is 25 to 6, which is consistent with what we had at 16:1 and 64:1.
Like the rest of the market, we are following the same approach of scaling by leveraging a building block. This is also why GaN wins again. Otherwise, you need to change the topology, and that leads to an unreliable implementation time.
Going to the partnerships, it’s not only Payton. There are other players. There is clearly Sumida, and we partner with Würth. We have a big magnetics game.
Payton brings a lot of experience in planar. Planar is essential for these kinds of thicknesses; you cannot even think about not doing planar. A planar transformer here is essential, and Payton has been driving this for decades. They have a fantastic way to simulate losses, and at the end of the story, your efficiency depends a lot on these losses.
Maybe beyond the white paper, we are optimizing a little bit there, but there are a lot of losses in the magnetics. They’re copper losses; of course, there are fewer AC losses because the matrix transformer reduces those.
Where there is a frontier, and we are playing there already, is the secondary side. We have low-voltage 25-V MOSFETs, and then there are a lot of MLCCs. Getting very close to the termination is essential. One of the things that is not visible in the prototype is all the efficiency gain you get in that part. We are, of course, stronger on the electrical part. The reason the losses are concentrated in the transformer is that it’s not our native territory. We consider that we have optimized all the rest, and we left the weakness on the other side. The truth is, it’s a system design, and clearly, we are going to do even more on that side.
AS: Let’s talk about that a little more. In the loss model that the recent white paper described, the transformer accounts for 74% of the total losses, and the copper plus shielding alone is roughly two-thirds of the losses. What is optimized to bring that down: copper thickness, layer count, switching frequency, shield design, etc.?
PS: It’s the compromise of all these parameters. On switching frequency, as I said, we could not go so extreme. It depends. If customer X comes with a problem of space, you can sacrifice efficiency. So you always need to compromise these variables. At the moment, it’s something higher than 500 kHz, and it might go higher because, in the meantime, GaN is improving as well. Things are moving as we talk.
In terms of AC losses, it’s what I told you before. Embedding things underneath the transformer, for example, can be another step. Why not bring the FETs into the assembly? Today, we have more than 10 layers—many more than 10, actually—and it’s a very thick one. Most of the thickness comes from the PCB. But the idea is, again, encapsulation. We talked a lot about it at the panel today.
Encapsulation is key here. It’s one of the three main dimensions: technology, packaging, and architecture, let’s put it this way. If we go into encapsulation, the more things we can embed inside the PCB, including capacitors, the better. Silicon capacitors are now used everywhere. We have new ways of doing magnetics as well, but at the moment, this is discrete for this power level. I think it’s really the key to getting the best compromise between size and efficiency.
AS: Got it. Thank you. Let’s shift to the dual-side-cooled package, which is implemented in the high-voltage IBC on the primary GaN side. We’re talking 650 V, and that accounts for less than 7% of the losses. What does the dual-side-cooled package bring to this design?
PS: Very good. First of all, let me say I was a bit frustrated when we could not show the QFN on the Nvidia wall, because in the meantime we launched the part. You know there are rules, so we could not show it. We have two prototypes, one with TOLT, and we showed that on the Nvidia wall at OCP in October. Then we launched the part, and now we are showing the QFN here.
Of course, it adds a lot of value. All these boards in systems are now liquid-cooled; this is now the standard in the data center. And customers do very interesting, I have to say, phenomenal, encapsulation. The cooling is the tallest element in the entire encapsulation. It’s unbelievable.
In that logic [within those size constraints], we have an exposed pad, so we dissipate through the bottom, and we have improved even versus the TOLT, but we dissipate a lot of power to the top. This way, it goes directly into the cooler. With liquid cooling typically swinging between 55°C and 65°C, we can dissipate a lot of power. This gives GaN extra headroom.
This is why we get 20 kW. It’s not only the RDS(on) of the GaN, but also the way you dissipate. The customer is very comfortable that there is headroom, because we always talk about TDC [thermal design current], but there is also peak power, and peak power is enormous. We cannot disclose all the specs, but it can be much higher than the TDC. And of course, your partner needs to be happy, without any deterioration of performance at the end.
AS: Let’s talk about the secondary side, the low-voltage 25-V silicon MOSFETs. They account for 12.6% of the losses. I know that Renesas is working with EPC on dual-side-cooled low-voltage GaN for that particular socket. When could we expect this?
PS: First of all, let me clarify our partnership with EPC. It’s been a licensing agreement. On the other side, we are incubating our technology in our Naka factory in Japan, because this was the measure we took after TSMC announced it was going out of the foundry business we had with them.
So, in short, Renesas owns the process. That was a short-term solution to the TSMC problem, but we are in a great partnership with Alex [Lidow] and the team. We do co-packaging and a lot of other things together.
If I go into the innovation on low voltage, we are releasing these 25-V MOSFETs that are very similar to the devices we have done in our power stages. Now, where is GaN going there? GaN is still above 40 V. The performance of GaN below 40 V is not yet where we wish, because with silicon we can still go into the tens of megahertz without any issue at that voltage.
But the fight is very hard, and it’s going to happen. Now, the question is when. The time axis is not exactly given yet, but I think it’s a question of quarters, not years. The fight is there. Everybody’s testing GaN; we are testing GaN. At 50 V, we see the value. At 12 V, we might see it. At 6 V, not yet. It’s a moving target right now.
AS: Thank you so much for your time, Pietro. One more question. There are a lot of high-voltage IBCs on the market, specifically 800 V to 6 V. A lot of them were introduced at GTC earlier this year, and there’s news of more to come. What sets Renesas’ design apart in this regard?
PS: First of all, it’s the total grid-to-core approach. We are not the only one, but what the customer wants is for you, as a player, to play the entire game.
Let’s go a little into how to play. Clearly, we are an incumbent, a very heavy player in low voltage and core supply. Scaling up with GaN now, let me say in a bullish way, we are dominating the compute rack. Going to the sidecar, of course, we struggle; we don’t have all the technologies. So the strategy in this game is also about setting up partnerships, as we did with EPC for something else. These partnerships, and mergers and acquisitions, as we talked about this morning as well, are part of the game.
The pace of this market is very fast. The cycle time can be four or six quarters, which is exactly the time it takes us to make one product. So I think we will see more partnerships and more mergers and acquisitions, because you need to have more of a game all across the board.
There will be facility equipment, like ESS or SST, that you need to touch as well. In that logic, this game is getting more complex, and big players have to play partnerships, acquisitions, mergers, and so on.
This is a game we need to play because the customers are demanding, and these customers drive the market with trillions of dollars. We are still power suppliers, with $60 billion WSTS this year, so we are orders of magnitude lower. We need to follow where the market is going and try to be as responsive as we can.
AS: Well, thank you so much for your time, Pietro. I really appreciate it.
PS: Aalyia, thanks so much for having me here. I appreciate it as well.
Source: Power Electronics News







