Intel Foundry Demonstrates World's Thinnest GaN Chiplet at 19 Micrometers, Merging Power and Logic on a Single Die
Intel combines gallium nitride power transistors with silicon digital logic on an ultra-thin 19-micrometer chiplet harvested from 300mm wafers, targeting AI data center power delivery and 5G/6G infrastructure.
Overview
Intel Foundry has unveiled what it describes as the world’s thinnest gallium nitride (GaN) chiplet, measuring just 19 micrometers thick — roughly one-fifth the width of a human hair. The chiplet, harvested from a 300-millimeter GaN-on-silicon wafer, is notable not just for its dimensions but for what it integrates: both GaN power transistors and silicon-based digital logic circuits on a single die, eliminating the need for separate companion chips. The work was presented at the IEEE International Electron Devices Meeting (IEDM) and represents a research milestone in Intel’s effort to position itself as a systems foundry for the AI era, as detailed in a blog post by Intel Foundry.
What We Know
The chiplet combines two distinct transistor types on a single piece of silicon: GaN N-channel metal-oxide-semiconductor high-electron-mobility transistors (N-MOSHEMT) for power handling, and silicon p-channel metal-oxide-semiconductor field-effect transistors (Si PMOS) for digital control logic. According to HotHardware, this hybrid approach allows complex computing functions to be built directly into power chiplets, a significant departure from traditional manufacturing where GaN power electronics and silicon-based logic are kept on separate dies.
The technical specifications are striking. Transistor gate lengths reach as short as 30 nanometers, voltage blocking capability extends to 78 volts, and the chiplet’s RF operating frequencies exceed 300 GHz. Inverter circuits on the die switch in 33 picoseconds. Intel has also subjected the technology to four industry-standard reliability tests — time-dependent dielectric breakdown, positive bias temperature instability, high-temperature reverse bias, and hot-carrier injection — all of which it passed, according to Intel Foundry.
The 19-micrometer thickness is not merely a miniaturization achievement. At that scale, the chiplet can be sandwiched between layers in advanced 3D chip packages without significantly increasing overall height. This opens the door to point-of-load power delivery, where voltage regulation happens millimeters from the processor cores rather than through long routing paths that dissipate energy as heat. As HotHardware notes, GaN chiplets switch faster and lose less energy than silicon alternatives, enabling voltage regulators that are smaller, more efficient, and positioned closer to the processor.
The implications extend well beyond data centers. GaN’s high-frequency performance makes it a natural candidate for radio frequency frontend technology in 5G and 6G base stations, where efficient operation at frequencies exceeding 200 GHz is increasingly critical. Other potential applications include radar systems, satellite communications, and electric vehicle power stages.
What We Don’t Know
Intel has not announced a specific timeline for commercialization. The chiplet remains a research demonstration, and the company has not disclosed yield data, per-unit cost projections, or manufacturing scalability details. It is also unclear which Intel Foundry process node would eventually host production of GaN chiplets at scale, or whether the technology will first appear in Intel’s own products or be offered to external foundry customers.
The competitive landscape adds further uncertainty. Other semiconductor companies, including TSMC and GlobalFoundries, have their own GaN research programs, and it remains to be seen whether Intel’s integration of digital logic with GaN power transistors will translate into a durable manufacturing advantage.
Analysis
Power delivery has become one of the defining bottlenecks for AI infrastructure. As data center racks approach megawatt-scale power consumption, the energy lost in converting and routing electricity from rack-level supplies to individual processor cores represents a growing share of total operating cost. Moving from 48-volt rack distribution to 800-volt DC architectures can reduce required amperage from 18,500 to 1,250 amps for the same megawatt of delivered power, but the final conversion stages still rely on silicon-based voltage regulators that are bulky and inefficient at the frequencies modern processors demand.
Intel’s GaN chiplet addresses this problem at the physical level. By making the power conversion layer thin enough to embed inside the processor package itself, and by integrating the digital control logic that traditionally required a separate silicon die, the design could simplify power delivery architectures while reclaiming energy that currently dissipates as waste heat.
For Intel Foundry, the demonstration serves a strategic purpose beyond the technology itself. The company is actively competing for external foundry customers against TSMC, and the ability to offer heterogeneous integration — combining materials like GaN and silicon on a single chiplet — could differentiate its systems foundry pitch in a market where leading-edge logic node leadership alone may not be sufficient.