In the quest for greater processing power, engineers have spent decades shrinking transistors to fit billions onto computer chips. However, in the age of artificial intelligence, the industry faces a critical question: Is there a limit to how small these components can become, and is further miniaturization still the primary goal?
To understand the stakes, one must first grasp the function of the transistor. According to Suman Datta, a professor of electrical and computer engineering at Georgia Tech, modern devices rely on tiny silicon switches etched into chips. These switches toggle on and off billions of times per second, orchestrating the flow of data that powers everything from simple searches to complex AI data centers.
While transistors were once visible to the naked eye, they are now far smaller than the width of a human hair. Researchers have already created proof-of-concept devices where a single atom controls electron flow, representing the absolute smallest scale physically possible for the switching mechanism. Recently, Anton Persson of Chalmers University of Technology and Tara Peña of UCLA, along with their colleagues, published research in Nature Nanotechnology demonstrating nanoribbon transistors with channel widths of just 25 nanometers. They achieved this using two-dimensional semiconductors made of tungsten disulfide, materials that are only one or a few atoms thick.
These atomically thin materials allow for more precise control of electrical current than traditional silicon, potentially enabling even smaller devices. However, Persson and Peña note that while silicon transistors will continue to shrink over the next decade, the pace will slow. Moving beyond silicon to materials like two-dimensional semiconductors may eventually be necessary to achieve further reductions in size.
Size, however, is not the only metric. Cost and manufacturing reliability are significant hurdles. Datta compares the research and development costs in the semiconductor industry to those of pharmaceutical companies developing a single drug. Persson and Peña emphasize that producing billions of functional transistors at scale is far more difficult than creating a working prototype in a laboratory.
Energy efficiency has also become a paramount concern. Datta points out that simply adding more transistors without improving individual efficiency doubles the power budget. For instance, Nvidia’s latest data center GPUs consume approximately 1.4 kilowatts, with upcoming models expected to draw nearly 500 watts more, half of which is wasted as heat. “It’s not just smaller, faster, cheaper, but also more energy efficient,” Datta said, noting that the industry must balance all four vectors.
To overcome physical space limitations, engineers are increasingly turning to three-dimensional designs. Rather than making each transistor only smaller, the industry is stacking them vertically to fit more components within the same chip area. This approach allows for significant gains in computing power even if individual transistors shrink only modestly.
The drive for more powerful chips remains urgent due to the explosive growth of AI, but the benefits extend to consumer electronics as well. Smaller, more efficient transistors lead to faster devices, longer battery life, and lower costs. As Persson and Peña observe, this historical trend has enabled modern smartphones to outperform the room-sized supercomputers of previous decades.
Ultimately, the pursuit of miniaturization continues as long as the physics can be controlled and the economics remain viable. “We will do everything to make it as small as possible as long as we can control the physics,” Datta said. “And as long as we can, we will find a way to make the economics work.”
Leave a Reply