PartsCube Global

Back to Home
Semiconductor Engineering

AI Accelerators Drive Changes In Power Delivery Networks

Semiconductor Engineering

Key Takeaways:

  • AI accelerators and GPUs consume thousands of amperes at increasingly low voltages, causing a shift to vertical power delivery that brings regulation closer to loads.
  • Advances in integrated voltage regulators and passive components are enabling new power architectures.
  • Power delivery is becoming a system-level co-design challenge, expanding simulation needs and the role of OSATs.

Power delivery is now a significant performance limiter in data center chips. The demands of AI and high-performance computing chips running near 0.7 V for a 600W GPU are forcing current to rise to 1,000 amps and beyond, compounding resistive losses and voltage droop across the PDN. Multi-chiplet constructions only compound these issues by further increasing current density and signal congestion.

To solve these challenges, designers are turning to vertical power delivery networks, where the voltage regulator modules are relocated much closer to the processors from their position on the device periphery — to the PCB landside (under the PCB), inside the accelerator package, or inside the silicon interposer. Together with the growing customization of AI accelerator modules, there are many approaches on offer with varying design and manufacturing complexity.

The vertical PDN trend began nearly a decade ago with the integration of backside power delivery networks (BPDNs) on the wafer backside, enabling fine-tuned power delivery more directly to the transistors while reducing IR drop across the tens of interconnect levels on-chip. That approach is being extended with increasingly sophisticated on-die decoupling capacitors (DCaps) formed in deep trenches as part of the BPDN. While this addresses high-frequency decoupling at the chip level, recent changes inside packages, interposers, substrates, and across data center racks are becoming just as important.

Add to that recent developments in the design of compact passive devices, like high-performance inductors, and new architectural solutions become possible, moving the VRMs and DC-DC converters into the interposer or package.

But these architectural changes are neither straightforward nor simple because they involve system co-design of the PDN, interposer, power management ICs (PMICs), integrated voltage regulators (IVRs), DC-DC converters, capacitors, and other passive devices, alongside custom HPC, memory devices, and interconnects that perform today’s elaborate AI tasks.

“In the most advanced semiconductors, people are moving from discrete PMICs to distributed IVRs. Those IVRs are great because they are placed much closer to the chip, which reduces the IR drop to the chip’s own power delivery network, so it is way more efficient. In addition, closed and fast control loops between the chip power management unit to the IVR reduce the reaction time to voltage droop events from milliseconds to microseconds. Because this reaction is so fast you can tune voltage very accurately, with a high degree of granularity even on different areas of a chip,” said Nir Sever, senior director of business development at proteanTecs.

Reconfiguring the PDN is part of a much larger drive to improve energy use. “Energy efficiency becomes the most important performance index for AI-infused platforms moving forward,” said Lipen Yuan, senior director of TSMC’s Advanced Technology Business Development, during a 2024 IEDM short course.

Power delivery challenges are most important at points of load. “AI systems are placing unprecedented demands on board-to-chip power delivery, making the final stage of voltage regulation increasingly important,” said CP Hung, vice president of corporate R&D at ASE.


Fig. 1: By regulating voltage at the point of load, OSATs enable fine-tuned voltage control. Source: ASE

“The move to multi-die architectures is bringing power integrity to the forefront, alongside thermal, stress, and warpage challenges. But the underlying power-delivery problem already existed in advanced single-die designs,” said Sojan Philips, principal product manager of Redhawk-SC at Synopsys. “AI accelerators are now operating at currents approaching the kiloampere range, where even milliohm-level resistance in the power-delivery network can result in significant IR drop. That loss of voltage margin can directly affect performance and power targets, particularly maximum operating frequency and Vmin.”

Several developments illustrate these trends:

  • Backside PDN: Intel Foundry’s 18A was the first to move backside wafer processing in leading-edge high-volume manufacturing. Samsung and TSMC plan to implement backside PDNs at future advanced nodes, starting next year.
  • Integrated voltage regulators: ASE, Marvell, Empower, and others are moving voltage regulation closer to the processor.
  • Embedded capacitors and inductors: Integration of compact passive components is becoming a key differentiator in high-density power delivery.
  • Co-optimization of package and on-chip power: Power integrity can no longer be optimized independently at the PCB, substrate, interposer, and transistor levels.

As a result, OSATs are playing an increasing role in system co-design, electrical validation, and power integrity (PI) simulation. “The OSAT sits where die, substrate or interposer, passives, and board meet, so we are brought in earlier to model the package-level PDN and co-simulate it with the customer’s die power data,” said Alex Kizeev, technical director of signal and power integrity at Amkor’s design optimization team.

Kizeev breaks down the supply chain and responsibilities among different companies. “The best results come from a PDN budget partitioned across the die, package, and board. The foundry and fabless teams own the on-die PDN, including the die power map and the bump-map allocation. The substrate supplier defines what stackup and via/plane technology is possible. The OSAT optimizes the package-level pieces, including stackup choice, via and ball assignment, DCap selection and placement, warpage and reliability, and validation.”

“Backside power delivery can significantly improve power delivery by reducing IR drop and freeing up frontside routing resources for signals, but it doesn’t eliminate power-integrity challenges,” said Philips. “It also changes the thermal landscape. As power delivery and other structures move to the backside, the paths available for heat removal become more complex. This makes co-optimization of power integrity and thermal management increasingly important.

The overall trend is moving toward closer integration in advanced packages and dies, including the integration of inductors, capacitors, and power-conversion components into substrates and interposers.

Why power losses scale with increasing power
A simple illustration shows how the power efficiency drops as the industry scales from a 1kW GPU to a 5kW GPU (see figure 2). While losses from the voltage regulator itself stay constant at around 13%, resistive losses jump from 8% to 31%, and voltage droop rises from 17% to 31%. As a result, increasing power levels must be supplied to meet all the chip-level needs.


Fig. 2: At higher power levels, the PDN efficiency using motherboard voltage regulators degrades significantly due to VR losses, resistive losses, and droop. Source: Intel Foundry

With this lateral power delivery approach, power integrity (PI) becomes increasingly difficult to maintain as compute loads rise. As AI accelerators integrate increasing levels of compute, HBM, and interconnects between devices, the low supply voltage (0.7-0.9V) of devices means multiple kiloamp levels of current are needed to meet the high-power needs of modern GPUs. As if managing kiloamp currents is not enough, fast workload transients tend to magnify voltage droop, so designers apply voltage guardbands that further increase power consumption and reduce system efficiency. [1]

Next-gen data centers pull in 800V
For next-generation data centers built to meet AI demands, 800V DC power delivery is starting to replace traditional AC power distribution. The 800V DC power distribution coexists with multiple distributed regulators located much closer to individual loads.

To lower transmission losses from the grid to the microprocessor, next-generation PDNs begin at 800V DC and are stepped down in stages using DC-to-DC converters to 48V, 12V, and then 0.7V, for instance. One of the most critical junctures occurs at the point of load, where integrated voltage regulators supply the AI multi-chiplet package. By limiting the number of power conversions, system-level efficiency can improve while reducing thermal cooling requirements.

The voltage regulator module can be positioned underneath the PCB, a logical first step that is commonly referred to as vertical power delivery. However, chipmakers are embracing recent developments that shrink the size of various voltage regulator components, especially capacitors and inductors, to move voltage regulation inside the main processor package and below the SoC, called integrated power delivery. Finally, integrating voltage regulator components into the interposer itself is called fully integrated power delivery.

One step at a time
The approach companies take to voltage regulation integration depends on the company’s expertise, its partners, and the complexity of the system involved.

For this reason, companies are introducing suites of solutions. For example, ASE is offering a range of power delivery capabilities to enable PCB backside integration, which considers the limited space inside computer trays. [2] The company’s PowerSiP options can include double-sided connected capacitors sandwiched between coreless substrates; substrateless designs, solder-free cap array modules with exposed terminals for embedded substrate or PCB solutions; high-aspect-ratio 4:1 copper pin integration for lower packaging impedance; and high-density active and passive component integration using single- or double-sided molding.

“Power efficiency is becoming a defining challenge for next-generation AI systems,” said Harrison Chang, vice president of corporate R&D at ASE. “By shortening the high-current transmission path, our solution can reduce parasitic routing losses by 50% or more, translating directly into meaningful power savings at the system level,” he said in the release.

In June 2025, Marvell introduced its Package Integrated Voltage Regulator (PIVR) technology for AI and HPC. The PIVR replaces the conventional arrangement of discrete voltage-regulator components, often several inches away on the PCB, with compact silicon power converters and passive components tightly coupled to the processor. It’s designed to reduce the amount of low-voltage, high-current routing between the regulator and the processor. The PIVR can sit on the underside of the processor package, shortening the path from voltage conversion to the processor’s power connections. It also can sit next to the processor’s package or inside the advanced package.

Marvell is partnering with multiple companies on integration, including Infineon, which makes power semiconductors and packaged devices; Ferric, which specializes in designing integrated voltage regulators using thin-film magnetic inductors; Monolithic Power Systems, a developer of high-current-density power modules; and Empower Semiconductor, a developer of compact integrated power converters.

In July of this year, Analog Devices completed the acquisition of Empower Semiconductor, expanding ADI’s AI compute power delivery capabilities with integrated voltage regulators and silicon capacitors. “By enabling power conversion closer to the processor, the combined solution shortens the power delivery path and improves efficiency to support higher-performance, higher-density systems. Building on its leadership in high-performance power management, ADI is investing in its system-level platform to deliver a step-change in performance, density, and efficiency from grid to core.” [3]

In addition, on Nov. 2, onsemi announced it will acquire the assets of Ferric, which develops integrated voltage regulators, for up to $775 million. [4]

To understand these important acquisitions, it helps to take a closer look at IVRs.

What is an IVR?
An integrated voltage regulator is typically a 3D stack arrangement that replaces off-chip voltage regulator modules and bulky capacitors and inductors on the PCB with integrated magnetics and passives mounted onto the silicon or package using thin-film technology.

“The IVR is implemented as a 3D stack, bringing together the deep-trench capacitor, PMIC, and inductor, and embedding them directly into the substrate or interposer,” said Veshal Sridhar, product engineering architect, 3D IC at Synopsys. “Companies are moving it closer to the SOC for obvious reasons such as reduced signal parasitics between IVR and SOC die, but that introduces new thermal effects and stress effects because you’re bringing another 3D stack in and embedding it inside an interposer. There’s also potential noise coupling into the other dies from the IVR, because even for a simple 3D stack you are going to have noise coupling.”

In what Intel Foundry refers to as a Super High Density MIM (metal-insulator-metal), capacitance density can be improved by 10× compared to conventional MIM capacitors. [5] This may offer response times on the order of nanoseconds. As part of Intel’s collaboration with UMC for 12nm finFETs, the capacitors feature three metal layers for routing in a scalable design. “This broad adaptability highlights the HDMIM as a universal, highly scalable passive component, capable of heterogeneous integration with logic and power chips manufactured across diverse process nodes from various foundries.” [5]

Inductors have traditionally been one of the greatest passive bottlenecks due to their large footprint, but by integrating them into the package using copper-plated through-hole technology and advanced magnetic materials (see Figure 3), Intel is touting a high resistance-to-capacitance ratio while minimizing stray magnetic fields that could adversely affect surrounding circuits. [5]


Fig. 3: This in-package inductor uses copper-plated through holes embedded with high-permeability magnetics to improve the resistance-to-capacitance ratio and limit magnetic parasitic effects. Source: Intel Foundry

Intel’s FIVR platform uses switched-capacitor voltage regulators and hybrid buck converters to accommodate nanosecond transient spikes by computing cores without exacerbating voltage droop. FIVRs that allow a continuous switching range are superior to fixed SCVRs when it comes to delivering low-voltage inputs (e.g., 1.8 to 3V).

Finally, Intel’s low-voltage GaN MOSHEMT-on-CMOS devices are designed for >12V applications, where high density and high efficiency are essential, while minimizing the overall number of power conversions. By integrating both low voltage and high voltage MIM capacitors with active devices, it minimizes PDN impedance from VR to load. By monolithically integrating the GaN power switches with GaN nMOS drivers, the company can achieve even faster response to transients along with aggressive power density.

“Integrated voltage regulators provide unprecedented speed and granularity in voltage control, but realizing their full power-saving potential requires equally precise visibility into the silicon itself,” said proteanTecs’ Sever. “Traditional power management methods leave significant timing margin untapped. By continuously measuring timing margins across the full chip, cycle by cycle, under actual operating conditions, our embedded monitoring technology identifies where performance is truly limited and where voltage can be safely reduced. Combined with the fine-grained control of IVRs, this enables more aggressive power optimization without compromising performance or reliability.”

Can the system handle the power?
Multiphysics simulation plays an important role in power delivery planning. “A designer may have a simple stack in mind, interposer, base die, top die, and substrate. And it starts with a prototype grid and listing the power required for each die,” explained Synopsys’ Sridhar. “So there is a power budget that comes from the devices, and number of C4 bumps, number of microbumps, TSVs, etc. You want to determine whether your power delivery system can deliver the power needed from the C4 bumps through the top of the stack. Starting with a simple static IR-drop analysis, we can then leverage 3DIC Compiler together with signoff solutions like RedHawk-SC to help bridge the gap between early bump planning and PSE (power sourcing equipment) planning.”

In cases where the PDN cannot handle the power required of the system, design changes such as increasing the interposer size may come into play.

Sridhar noted how one customer brought system-level PI (power integrity) analysis into the early architecture stage, closing the loop between implementation and system analysis using 3DICC plus RHSC, turning iterations that once took days into hours and giving designers more freedom to explore before the architecture is frozen.

At the same time, the package assembly design kits that OSATs compile are becoming more detailed. “Kits are moving beyond geometric design rules to include electrical models, material properties, and validated reference designs. Teams now must consider bump pitch against current density, metal layers against cost, and DCap placement against thermal path early in the design,” said Amkor’s Kizeev.

Conclusion
Power integrity and power efficiency have already become key differentiators in advanced AI accelerator-based data center designs. Companies are racing to incorporate leading power delivery solutions, including integrated voltage regulators, in positions closer to the GPU or AI accelerator modules where they can be most effective. Advanced power delivery began with the integration of wafer backside PDNs, but package-level and interposer-level solutions are coming fast.

Coordination among major companies in the supply chain is critical during early-stage PDN design. “Some of these design houses are not building the entire system, for instance, like HBM, where even the base die and the stack may be coming in from other design houses,” noted Synopsys’ Phillips. They want to make sure that the grid they’re designing in-house works with the other models. So they want a model exchange between these two, which is performed as part of a feasibility analysis.”

References

  1. Radhakrishnan, V. Javvaji and N. Butzen, “Driving Power Delivery Innovations for the AI Data Center Era,” Intel Foundry, https://www.intel.com/content/dam/www/central-libraries/us/en/documents/2026-08/foundry-power-delivery-white-paper.pdf
  2. “ASE Expands PowerSiP Platform for Next-Generation AI Power Delivery,” press release, October 13, 2026, https://ase.aseglobal.com/press-room/ase-expands-powersip-platform-for-next-generation-ai-power-delivery/
  3. https://www.analog.com/en/newsroom/press-releases/2026/5-19-2026
  4. “onsemi to Acquire Ferric, Establishing Leadership in Power Delivery for Next-Generation AI Processors,” November 2, 2026. https://www.onsemi.com/company/newsroom/news-and-insights/onsemi-to-acquire-ferric-establishing-leadership-in-power-delivery-for-next-generation-ai-processors/
  5. Yu et al., “Vertical Power Delivery for 5 kW AI Systems: From Advanced Passives to GaN IC Integration,” 2026 IEEE 38th International Symposium on Power Semiconductor Devices and ICs (ISPSD), Las Vegas, NV, USA, 2026, pp. 337-340, doi: 10.1109/ISPSD64561.2026.11553768.

Related Articles
Backside Power Delivery Creates Fab Tool, Thermal Dissipation Barriers
Moving the power delivery network to the backside of a chip reduces congestion, but it introduces new challenges for fabs.

GaN Power Devices Go Vertical
Why new designs and process flows could help overcome manufacturing challenges.

New Approaches To Power Decoupling
Strategic placement of decoupling capacitors and regulators is critical to keeping power stable, but no one approach is sufficient.

The post AI Accelerators Drive Changes In Power Delivery Networks appeared first on Semiconductor Engineering.