
Move the voltage regulator from beside a processor to underneath it, and the resistive loss in the final stretch of copper can fall sharply. In the illustrative 1,000 A model in this post, I²R loss drops from 70 W to 10 W, a cut of about 86%.
The total loss from the 48 V bus to the load falls by much less, from about 161 W to about 114 W, or about 29%, because a converter squeezed under the chip usually gives back some efficiency. That gap between the copper saving and the system saving is the subject of this post, with a calculation you can redo using your own numbers.
Key takeaways
- I²R loss depends on the square of current, so the place to convert from 48 V to under 1 V sets how far the high current has to travel.
- Vertical delivery moves loss more than it removes it. In our model, conversion accounts for about 91% of the vertical design's loss.
- Under our assumptions, the two approaches tie at about 300 A. By 2,000 A, vertical saves nearly half the loss.
- The figure to demand from a vertical module is its efficiency at your current and your voltage. Peak efficiency can sit far from the full-load number.
What moves, and what does not
Vertical power delivery and backside power are easy to confuse because both describe power arriving from below. They work at different scales. Backside power delivery rearranges metal inside the die, and the fight there is over on-die IR drop and routing tracks. Vertical power delivery moves the converter, which sits on a board or in a package, closer to the point of load.
In a lateral layout, the final regulator sits on the board next to the package, so high current at under 1 V crosses board copper, package layers and bumps. In a vertical layout, 48 V or an intermediate voltage is carried to a spot beneath the die, and the high current has only a short path upward. Published work places that converter at several heights:
| Converter location | Example in the literature | What it asks of the design |
|---|---|---|
| Underside of the motherboard | UC Berkeley's APEC 2024 converter is intended for this position, feeding the processor through board vias | Board via planning and cooling on the bottom side |
| Inside the package substrate | University of Illinois Chicago's 2026 framework embeds GaN switches and arrays of small passives | Embedded passives, plus a share of the area under the die |
| On or within an interposer | The same group's 2023 architectures | Compact high-ratio converters and integrated passives |
| Dedicated power die | The 2023 two-stage architectures, for the lower-voltage stage | A further die in the stack, plus heat at that level |
The closer the converter gets, the smaller and hotter it has to be. Our post on silicon interposers gives useful background on the structures involved at the interposer level.
Why the square law decides the layout
Two points explain it. The first is the voltage argument. Deliver 1 kW at 1 V and the current is 1,000 A. Deliver the same power at 48 V and it is about 20.8 A. Through an illustrative 1 mΩ path, the first case loses 1,000 W and the second about 0.43 W, a ratio of 48 squared, or 2,304. The University of Illinois Chicago (UIC) authors make the same point in their 2023 paper: PCB-level conversion means 48 times the current and 48 squared times the loss.
The second is a useful identity. Loss divided by delivered power is I²R divided by VI, which reduces to IR divided by V. University of Minnesota authors state this directly in a 2026 preprint: resistive loss as a fraction of delivered power equals the supply droop as a fraction of the rail voltage. A 2% droop is therefore a 2% loss, and you can read the efficiency cost straight off an IR drop budget.
The loss audit
Here is a comparison of three layouts feeding the same load. The load current and voltage are illustrative choices. The resistances are vendor claims from Vicor, which sells vertical modules: about 200 µΩ for a lateral path in a May 2024 article, about 70 µΩ for a close-coupled lateral layout and about 10 µΩ for a design that pushes most current through bottom-side modules. Vicor quoted those last two at about 800 A and at 1,200 A, and we apply them all at 1,000 A, assuming that resistance stays constant.
The converter efficiencies are also assumptions. For lateral we use 90%, the figure UIC assumes for its PCB-level reference. For vertical we use 88%, which Infineon showed at APEC 2025 for a 54 V to 0.7 V module at 300 A. That is a vendor claim at a different voltage and current from our model, so treat it as a placeholder. The regulator is assumed to raise its output so the load always sees 0.75 V.
| At 1,000 A into 0.75 V (750 W) | Lateral, long path | Lateral, close | Vertical |
|---|---|---|---|
| Path resistance (vendor claim) | 200 µΩ | 70 µΩ | 10 µΩ |
| I²R loss in the path | 200 W | 70 W | 10 W |
| Voltage drop | 200 mV (26.7% of 0.75 V) | 70 mV (9.3%) | 10 mV (1.3%) |
| Regulator output power | 950 W | 820 W | 760 W |
| Converter efficiency (assumed) | 90% | 90% | 88% |
| Input power at 48 V | 1,055.6 W | 911.1 W | 863.6 W |
| Conversion loss | 105.6 W | 91.1 W | 103.6 W |
| Total loss | 305.6 W | 161.1 W | 113.6 W |
| End-to-end efficiency | 71.1% | 82.3% | 86.8% |
To check a column, take the vertical one. The loss is 1,000 squared times 10 µΩ, which is 10 W. Add 750 W of load and the regulator must supply 760 W. Dividing by 0.88 gives 863.6 W at the input, so the total loss is 863.6 minus 750, or 113.6 W. The 48 V side barely registers: at about 22 A or less, an illustrative 1 mΩ path loses under 0.5 W in every column, and we have left it out.
Now compare the vertical column with the close lateral one. I²R falls by 60 W, or about 86%. Conversion loss rises by about 12.5 W. The net saving is about 47.5 W, or about 29% of the total. Against the long lateral path, the saving is about 192 W, or about 63%. So the answer to "how much does vertical save?" depends mainly on which lateral design you are replacing.
The model also shows where the loss ends up. In the close lateral case, the path accounts for about 43% of the total. In the vertical case it is about 9%, and the converter takes the other 91%.
TDK's February 2026 Electronic Design article shows the same general shape, listing a 70% cut in I²R loss and a 13.6% cut in system power loss for vertical against standard lateral delivery. It gives no test conditions, so this is a vendor claim too.
What vertical converters achieve today
Because conversion decides the result, the efficiency you assume matters more than any other input. Here is what we could trace. Boundaries differ between the rows, so they should not be ranked against one another.
| Source | Design | Reported efficiency | Status |
|---|---|---|---|
| UC Berkeley, APEC 2024 | Single-stage 48 V to 1 V, tested to 1,500 A | 92.7% peak at 395 A and 85.7% at 1,500 A, gate drive included | Measured on a lab prototype |
| Infineon, APEC 2025 plenary | 54 V to 0.7 V module, GaN plus silicon | 88% at 0.7 V and 300 A | Vendor claim on a slide |
| UIC, 2026 preprint | 70 distributed regulators, 48 V to 1 V, 1 kW into a 500 mm² load | 84% end to end, using 54% of the area under the load | Modelled, covers conversion and distribution |
The Berkeley row shows why the operating point matters. Efficiency falls 7 points between peak and full load, and at 85.7% the converter sheds about 250 W of heat at 1,500 A (1,500 W divided by 0.857, minus 1,500 W). Swap that 85.7% into our model and the vertical total rises from 113.6 W to about 136.8 W. The saving against close lateral falls from about 29% to about 15%. Berkeley's conditions differ from ours, as its output is 1 V, so this is a sensitivity test and not a prediction.
The UIC work reaches a similar conclusion from another direction. Its 2023 paper finds that all the proposed architectures lose under 10% in the distribution network and over 10% in the converters, and it names converter efficiency as the thing to improve.
How the answer changes with current
Keep the same assumptions (70 µΩ at 90% against 10 µΩ at 88%, into 0.75 V) and vary the current. This is illustrative. Real resistance and efficiency both change with load and temperature, and our model holds them fixed. The last column gives the vertical converter efficiency at which the two totals become equal.
| Current | Lateral I²R | Vertical I²R | Lateral total | Vertical total | Vertical saves | Break-even efficiency |
|---|---|---|---|---|---|---|
| 300 A | 6.3 W | 0.9 W | 32.0 W | 31.7 W | 0.3 W (about 1%) | 87.9% |
| 500 A | 17.5 W | 2.5 W | 61.1 W | 54.0 W | 7.1 W (about 12%) | 86.6% |
| 1,000 A | 70 W | 10 W | 161.1 W | 113.6 W | 47.5 W (about 29%) | 83.4% |
| 2,000 A | 280 W | 40 W | 477.8 W | 250.0 W | 227.8 W (about 48%) | 77.9% |
The break-even column works as a quick decision tool. At 1,000 A, for instance, a vertical converter below about 83.4% would lose more than a good 90% lateral design with 70 µΩ. Against a 200 µΩ lateral path, the threshold falls to 72%, so the baseline chosen changes the answer a great deal.
Vertical delivery does not end the square law either. A Google Research paper in IEEE T-CPMT (2026) describes accelerator current demand approaching and exceeding 10,000 A, according to its abstract. At that current, the same illustrative 10 µΩ path would lose 1 kW and drop 100 mV, or 13.3% of a 0.75 V rail.
What a loss table leaves out
Heat lands in a new place. The loss that was once spread across a board now concentrates under the die. UIC's 1 kW substrate-embedded model puts junction temperature above 110 °C without thermal management, with loss rising to 225 W and efficiency falling to 81.6%. With micro-channel cooling at 1.6 g/s, it reports about 84 °C, 210 W and 82.6%. Berkeley's prototype, which has no heat sink, ran continuously at 1,300 A on air, and the authors say that going higher needs better heat sinking or liquid cooling.
Area is the hard limit. UIC's 2026 and 2023 studies both work to 2 A/mm², with 1 kA into a 500 mm² die as the reference load. Berkeley's prototype measures 202.5 by 25.8 mm, or about 5,225 mm². By our arithmetic, 1,500 A over that box footprint is about 0.29 A/mm². That is a fair result for a board-level prototype with a different aim, and it shows how much shrinking separates board-level hardware from a die-shadow target. UIC's higher-efficiency variants need nearly 75% of the load die area, and its authors raise concerns about fabrication reliability and thermal stress at that level.
Vias and bumps open up. This one helps vertical. In UIC's 2023 model, a PCB-level reference needs an unreasonably large 1,200 mm² die for its 1 kA, while a vertical approach can feed a 500 mm² die using about 1% of the BGAs and 2% of the C4 bumps. For related packaging context, see our post on advanced semiconductor packaging.
Transients need separate analysis. UIC's 2026 abstract gives a peak steady-state drop of 2.7% and a peak transient drop of 9% without decoupling capacitors. A DC audit will not show you that second number.
Mechanics count as well. The Google abstract lists thermal management, warpage and manufacturing as open challenges alongside power and signal integrity, and it treats input voltage and module granularity as design choices. Choose these with the package team from the start. Our chiplet and advanced packaging lifecycle on The Semiconductor Hub shows where interposer and substrate decisions fall in a programme.
Running this audit on your own design
- Start from your real lateral design. Extract path resistance at temperature. Using someone else's 200 µΩ will overstate the gain if yours is 70 µΩ.
- Use your own current profile. A chip that spends most of its time at 400 A should be judged mainly at 400 A, even if its peak is much higher.
- Ask for efficiency at your operating point. Get it at your voltage, current and temperature, with gate drive included, and be clear about what the boundary covers.
- Calculate your break-even efficiency. Divide the vertical regulator's output power by the lateral design's input power. If the quoted module is below that number, keep the lateral layout or shorten its path.
- Check that the heat has somewhere to go. Rerun the loss table at the temperature your cooling can actually hold.
University of Minnesota authors offer a useful yardstick for the whole chain. If a first stage reaches 98%, the rest of the system must achieve 91.8% to meet a 90% end-to-end goal, and they treat 90% as a level that is within reach but needs several open problems solved. None of the figures traced in this post reaches 90% at full load, so the case for vertical today rests on high-current designs, where the copper loss is large enough to pay for the converter.
Sources
- Zhu, Zou and Pilawa-Podgurski (UC Berkeley), "A 1500-A/48-V-to-1-V Switching Bus Converter for Next-Generation Ultra-High-Power Microprocessors", APEC 2024, 25 to 29 February 2024.
- Krishnakumar and Partin-Vaisband (University of Illinois Chicago), "Vertical Power Delivery for Emerging Packaging and Integration Platforms - Power Conversion and Distribution", arXiv:2309.10141, 18 September 2023.
- Krishnakumar et al., "A Comprehensive Design Framework for Vertical Power Delivery in High-Performance Computing", arXiv:2606.28837, 27 June 2026.
- Yue et al. (University of Minnesota), "Toward Multi-kW Power Delivery Methodologies for Advanced 3D Heterogeneous Integration", arXiv:2609.24904, 21 September 2026.
- Gan et al. (Google Research), "Towards 10kAmp AI Chip: A System Perspective from Lateral to Vertical Power Delivery", IEEE Transactions on Components, Packaging and Manufacturing Technology, 2026. Abstract only.
- Neyer (Infineon), "Improving cloud computing and AI power delivery efficiency with a holistic semiconductor approach", APEC 2025 plenary slides. Vendor claims.
- Wood (Vicor), "GenAI: Helping to tame the most power-intensive computing application on Earth", May 2024. Vendor claims.
- Jain (Vicor), "Vertical Power Delivery Enables Cutting-Edge Processing", undated on Vicor's site. Vendor claims.
- Pelletier (TDK), "What's the Difference Between Lateral and Vertical Power Delivery?", Electronic Design, 13 February 2026. Vendor claims.
The worked calculations use illustrative load figures and assumed converter efficiencies, and they hold resistance and efficiency constant. Vendor figures are the companies' own claims, and preprint results are modelled unless stated otherwise.

Your comments will be moderated before it appears here.