Much has been made of the excessive power demands of AI, but solutions are sparse. This has led engineers to consider completely new paradigms in computing: optical, thermodynamic, reversible—the list goes on. Many of these approaches require a change in the materials used for computation, which would demand an overhaul in the CMOS fabrication techniques used today.
Over the past decade, Hector De Los Santos has been working on yet another new approach. The technique would require the same exact materials used in CMOS, preserving the costly equipment, yet still allow computations to be performed in a radically different way. Instead of the motion of individual electrons—current—computations can be done with the collective, wavelike propagations in a sea of electrons, known as plasmons.
De Los Santos, an IEEE Fellow, first proposed the idea of computing with plasmons back in 2010. More recently, in 2024, De Los Santos and collaborators from University of South Carolina, Ohio State University, and the Georgia Institute of Technology created a device that demonstrated the main component of plasmon-based logic: the ability to control one plasmon with another. We caught up with De Los Santos to understand the details of this novel technological proposal.
How Plasmon Computing Works
IEEE Spectrum: How did you first come up with the idea for plasmon computing?
De Los Santos: I got the idea of plasmon computing around 2009, upon observing the direction in which the field of CMOS logic was going. In particular, they were following the downscaling paradigm in which, by reducing the size of transistors, you would cram more and more transistors in a certain area, and that would increase the performance. However, if you follow that paradigm to its conclusion, as the device sizes are reduced, quantum mechanical effects come into play, as well as leakage. When the devices are very small, a number of effects called short channel effects come into play, which manifest themselves as increased power dissipation.
So I began to think, “How can we solve this problem of improving the performance of logic devices while using the same fabrication techniques employed for CMOS—that is, while exploiting the current infrastructure?” I came across an old logic paradigm called fluidic logic, which uses fluids. For example, jets of air whose direction was impacted by other jets of air could implement logic functions. So I had the idea, why don’t we implement a paradigm analogous to that one, but instead of using air as a fluid, we use localized electron charge density waves—plasmons. Not electrons, but electron disturbances.
And now the timing is very appropriate because, as most people know, AI is very power intensive. People are coming against a brick wall on how to go about solving the power consumption issue, and the current technology is not going to solve that problem.
What is a plasmon, exactly?
De Los Santos: Plasmons are basically the disturbance of the electron density. If you have what is called an electron sea, you can imagine a pond of water. When you disturb the surface, you create waves. And these waves, the undulations on the surface of this water, propagate through the water. That is an almost perfect analogy to plasmons. In the case of plasmons, you have a sea of electrons. And instead of using a pebble or a piece of wood tapping on the surface of the water to create a wave that propagates, you tap this sea of electrons with an electromagnetic wave.
How do plasmons promise to overcome the scaling issues of traditional CMOS logic?
De Los Santos: Going back to the analogy of the throwing the pebble on the pond: It takes very, very low energy to create this kind of disturbance. The energy to excite a plasmon is on the order of attojoules or less. And the disturbance that you generate propagates very fast. A disturbance propagates faster than a particle. Plasmons propagate in unison with the electromagnetic wave that generates them, which is the speed of light in the medium. So just intrinsically, the way of operation is extremely fast and extremely low power compared to current technology.
In addition to that, current CMOS technology dissipates power even if it’s not used. Here, that’s not the case. If there is no wave propagating, then there is no power dissipation.
How do you do logic operations with plasmons?
De Los Santos: You pattern long, thin wires in a configuration in the shape of the letter Y. At the base of the Y you launch a plasmon. Call this the bias plasmon, this is the bit. If you don’t do anything, when this plasmon gets to the junction it will split in two, so at the output of the Y, you will detect two equal electric field strengths.
Now, imagine that at the Y junction you apply another wire at an angle to the incoming wire. Along that new wire, you send another plasmon, called a control plasmon. You can use the control plasmon to redirect the original bias plasmon into one leg of the Y.
Plasmons are charge disturbances, and two plasmons have the same nature: They either are both positive or both negative. So, they repel each other if you force them to converge into a junction. And by controlling the angle of the control plasmon impinging on the junction, you can control the angle of the plasmon coming out of the junction. And that way you can steer one plasmon with another one. The control plasmon simply joins the incoming plasmon, so you end up with double the voltage on one leg.
You can do this from both sides, add a wire and a control plasmon on either side of the junction so you can redirect the plasmon into either leg of the Y, giving you a zero or a one.
Building a Plasmon-Based Logic Device
You’ve built this Y-junction device and demonstrated steering a plasmon to one side in 2024. Can you describe the device and its operation?
De Los Santos: The Y-junction device is about 5 square [micrometers]. The Y is made up of the following: a metal on top of an oxide, on top of a semiconducting wafer, on top of a ground plane. Now, between the oxide and the wafer, you have to generate a charge density—this is the sea of electrons. To do that, you apply a DC voltage between the metal of the Y and the ground plane, and that generates your static sea of electrons. Then you impinge upon that with an incoming electromagnetic wave, again between the metal and ground plane. When the electromagnetic wave reaches the static charge density, the sea of electrons that was there generates a localized electron charge density disturbance: a plasmon.
Now, if you launch a plasmon by itself, it will quickly dissipate. It will not propagate very far. In my setup, the reason why the plasmon survives is because it is being regenerated. As the electromagnetic field propagates, you keep regenerating the plasmons, creating new plasmons at its front end.
What is left to be done before you can implement full computer logic?
De Los Santos: I demonstrated the partial device, that is just the interaction of two plasmons. The next step would be to demonstrate and fabricate the full device, which would have the two controls. And after that gets done, the next step is concatenating them to create a full adder, because that is the fundamental computing logic component.
What do you think are going to be the main challenges going forward?
De Los Santos: I think the main challenge is that the technology doesn’t follow from today’s paradigm of logic devices based on current flows. This is based on wave flows. People are accustomed to other things, and it may be difficult to understand the device. The different concepts that are brought together in this device are not normally employed by the dominant technology, and it is really interdisciplinary in nature. You have to know about metal-oxide-semiconductor physics, then you have to know about electromagnetic waves, then you have to know about quantum field theory. The knowledge base to understand the device rarely exists in a single head. Maybe another next step is to try to make it more accessible. Getting people to sponsor the work and to understand it is a challenge, not really the implementation. There’s not really a fabrication limitation.
But in my opinion, the usual approaches are just doomed, for two reasons. First, they are not reversible, meaning information is lost in the computation, which results in energy loss. Second, as the devices shrink energy dissipation increases, posing an insurmountable barrier. In contrast, plasmon computation is inherently reversible, and there is no fundamental reason it should dissipate any energy during switching.
Instagram Engineering
Shopify Engineering