As GPUs and accelerators scale up to support larger AI workloads, traditional power regulator designs that live on the motherboard are starting to hit their current density and transient response limits. To counter this, Intel is planning to integrate voltage regulation (IVR) directly into the chip’s packaging using its Foveros-B variant, which is said to allow the creation of 5,000W GPUs by 2027.
If you thought that the RTX 5090’s 575W was plentiful, wait until you see next-gen Vera Rubin Ultra and its successor Feynman Ultra accelerators, which are expected to gobble 2,300W and 4,000W+, respectively. Looking at this, Intel’s 5kW GPUs no longer sound unrealistic. To achieve this level of power density, new methods of delivery and cooling have been developed to ensure reliable operation of these beasts.
IVRs move the final DC-to-DC conversion right next to the silicon chip, hidden inside the packaging. This approach allows a high current at low voltages to travel along the PCB/interposer traces, reducing resistance losses and improving transient response as the regulator is physically close to each power domain. Understandably, manufacturing IVRs will be more complicated, requiring advanced 2.5D/3D assembly techniques, leveraging interposers, silicon bridges, and micro-bumps to route power vertically through the package.
That said, IVRs aren’t a magic bullet as they bring some downsides. For instance, IVRs have some inefficiencies like any voltage regulation system, but this time the resulting heat is right beside the main chip, meaning that it needs to be dealt with by the same cooling system. They also add cost and complexity compared to regular board-level VRMs.
This is what happened with Intel’s earlier attempts at this technology, which was used on the Haswell-generation processors. The IVR performed the final DC-to-DC conversion on the die and produced multiple internal voltages for the CPU cores, iGPU, and uncore. Unfortunately, the additional heat generation limited the CPU’s ability to boost higher. This should be less of a concern with the new design, as it puts the IVRs into the packaging, but we can’t be sure until products reach the market.
What is certain is that cooling solutions have advanced drastically lately. From Frore Systems’ LiquidJet, which is expected to handle up to 4,400W chips, to Microsoft’s etched liquid channels on silicon, cooling high-wattage chips is less of a limiting factor.
Intel expects production of this solution to debut by 2027, allowing customers to evaluate its potential for the following multi-kilowatt designs. The company is scheduled to reveal further details during the ISSCC conference in February 2026.
