ETH Zurich Simulates 42,000-Atom Nanoribbon on Supercomputers, Advancing Transistor Design

Scientists at ETH Zurich and MARVEL have achieved a breakthrough in transistor design by simulating a 42,000-atom nanoribbon – a fundamental component of next-generation transistors – on supercomputers in Switzerland and the USA. This feat, accomplished using new software called QuaTrEx, pushes the boundaries of what’s possible in nanoscale materials modelling and addresses the growing need for more powerful microchips driven by advances in AI and cloud computing. The team combined density functional theory, the GW approximation, and the non-equilibrium Green’s function in a novel way to overcome previous computational limitations. “A brute force approach to simulate such a physical system would not run even on the largest supercomputer,” says Mathieu Luisier, highlighting the ingenuity required for this achievement, which earned an Honorable Mention at the 2025 ACM Gordon Bell Prize.

QuaTrEx Software Combines DFT, GW, and NEGF Methods

QuaTrEx (Quantum Transport Simulations at the Exascale and Beyond) represents a significant leap forward in the computational modeling of nanoscale materials, promising to accelerate the development of next-generation transistors. This combination allows researchers to move beyond simulating only a handful of atoms, achieving simulations of devices containing over 42,000 atoms – a scale crucial for realistic transistor modeling. The impetus behind QuaTrEx stems from the limitations of classical physics at the nanometer scale, where quantum interactions become dominant. Existing computational methods struggled to accurately model these interactions in devices of practical size.

The team’s breakthrough isn’t simply combining established techniques, but refining them for this specific purpose. A key innovation lies in calculating boundary conditions and, crucially, performing the GW approximation using open boundary conditions – a first in the field. Furthermore, the researchers implemented a novel parallel algorithm to efficiently distribute the computational load across multiple GPUs, maximizing supercomputer performance. While a complete transistor simulation remains within reach, the team prioritized validation using a 42,240-atom device mirroring industry-fabricated structures.

Luisier notes future improvements could involve “using what people call mixed-precision,” potentially increasing speed tenfold while carefully balancing accuracy, or leveraging machine learning to bypass computationally expensive initial steps.

42,240-Atom Nanoribbon Simulation Validates Computational Approach

Developed by a team led by Mathieu Luisier at ETH Zurich and MARVEL, the package successfully simulated a nanoribbon—a core component of next-generation transistors—comprising over 42,240 atoms. This feat signifies a major step forward in accurately predicting the behavior of materials at the quantum level, crucial for continued advancements in microchip technology and the relentless pursuit of Moore’s scaling law. The simulation ran concurrently on supercomputers in both Switzerland and the United States, leveraging the Alps system and the Frontiers machine. The breakthrough isn’t simply about scale; it’s about refining established techniques.

Specifically, calculating the ‘W’ component of the GW approximation using open boundary conditions—allowing energy and particles to flow through the simulation edges—was a previously unattempted step. This allows for more realistic modeling of real-world device operation. The team achieved a sustained performance exceeding one exa (10¹⁸) floating point operations per second. While a full transistor simulation remains within reach, the 42,240-atom nanoribbon served as a vital validation test, closely mirroring structures currently fabricated by the semiconductor industry.

Luisier notes potential future improvements, including exploring mixed-precision calculations and integrating machine learning to predict DFT inputs, potentially increasing speed by a factor of ten.

Frontiers & Alps Supercomputers Achieve Exascale Performance

This achievement marks a significant leap forward in our ability to design the next generation of transistors, essential for advancements in artificial intelligence, robotics, and cloud computing. Furthermore, a novel parallel algorithm was implemented to distribute the simulation across multiple GPUs, maximizing the utilization of the supercomputers’ capabilities. The simulations focused on nanoribbon field-effect transistors, devices with cross-sections measured in just a few nanometres. “We could not simulate the entire transistor because we did not have enough hours on the machine, though technically we could have done it,” Luisier clarifies, adding that the team focused on validation tests using realistic device geometries.

Future Work: Mixed-Precision & Machine Learning Integration

The success of QuaTrEx in simulating nanodevices with over 42,000 atoms isn’t the endpoint for the team at ETH Zurich and MARVEL, but rather a springboard for further innovation. Mathieu Luisier envisions significant speed gains through the adoption of “mixed-precision” computing, a technique already common in machine learning. “Instead of representing all numbers with 64 bits, for example, you could represent some of them with 32 bits and speed up the calculation,” he explains, acknowledging the trade-off between computational velocity and accuracy.

Current graphics processing units (GPUs) optimized for artificial intelligence already operate with 16 or even 8 bits, suggesting a potential tenfold increase in QuaTrEx’s processing speed—provided the loss of precision can be mitigated. This delicate balance will be crucial for future development, demanding careful calibration to maintain the reliability of simulations. Beyond precision adjustments, the researchers are exploring the integration of machine learning directly into the simulation process.

Currently, QuaTrEx relies on density functional theory (DFT) to initially calculate the Hamiltonian of the device, a computationally intensive step. “That is expensive and can become a bottleneck if you want to consider more atoms,” Luisier points out. The team hopes to bypass this initial calculation entirely, leveraging machine learning algorithms to predict the necessary DFT inputs. This would represent a paradigm shift, moving from computation to prediction, and further accelerating the simulation of increasingly complex nanoscale systems.

The ultimate goal, according to the team’s recent work presented at the International Conference for High Performance Computing, Networking, Storage, and Analysis in November 2025, is to model entire logic gates—complex circuits capable of performing fundamental operations— paving the way for the next generation of microchips.