For decades, the progress of computing followed a predictable rhythm: transistors got smaller, chips got faster, and software grew more capable without developers needing to worry about the underlying hardware. This era, defined by Moore’s Law and Dennard scaling, is hitting a fundamental "thermal wall." As transistors approach the atomic scale, quantum tunneling causes electrons to leak through barriers, and the heat generated by moving charges becomes impossible to dissipate.
Researchers are now shifting from geometric scaling—simply making things smaller—to functional scaling, where the physics of the material itself performs the computation.
Moving Beyond Charge as a State Variable
Traditional computers represent information using the presence or absence of electrical charge. However, moving charge through resistive wires generates heat. To bypass this, researchers are exploring alternative "state variables" like electron spin or photons.
Kaushik Roy and colleagues have demonstrated that Spin Torque Devices (STDs) can use the magnetic orientation of electrons to perform logic. Because flipping a magnetic state requires far less energy than moving a stream of electrons, these devices offer a path toward ultra-low-power neuromorphic hardware. Similarly, Tuomo von Lerber has shown that all-optical logic gates can process information at the speed of light using laser injection locking, eliminating the resistive losses inherent in copper interconnects.
Shattering the Von Neumann Bottleneck
In a standard computer, data must constantly travel between the processor and the memory. This "Von Neumann bottleneck" accounts for the vast majority of energy consumption in modern AI tasks. The solution is In-Memory Computing (IMC), where the memory element itself performs the calculation.
Alvaro Velasquez and Rajendra Bishnoi describe using emerging non-volatile technologies like Resistive RAM (RRAM) and Ferroelectric FETs (FeFETs) to build crossbar arrays. In these systems, matrix multiplication—the core operation of deep learning—happens naturally through Kirchhoff’s laws: voltages applied to the rows result in currents at the columns that represent the product of the inputs and the stored weights. Qunsong Zeng has applied this to 6G baseband processing, achieving a million-fold improvement in energy efficiency by executing Fourier transforms directly within the memory array.
Embracing Stochasticity and Imperfection
At the nanoscale, devices become inherently "noisy" and unreliable. While traditional digital logic fights this randomness, a new wave of research treats it as a feature. Siddharth Gaba and Wei Lü argue for embracing the stochastic switching of silver-filament memristors to generate random bitstreams for stochastic computing.
This philosophy extends to Echo State Graph Neural Networks (ESGNN). Shaocong Wang and Yi Li developed a system that uses the intrinsic randomness of dielectric breakdown in resistor arrays to create fixed neural weights. By not trying to "fix" the hardware's imperfections, they reduced training costs by 99%, proving that "unreliable" hardware can be highly effective for probabilistic AI tasks.
The Shift to Hardware Software Co-Design
As the "free lunch" of automatic hardware speedups ends, the burden of performance is shifting to how we integrate materials, circuits, and algorithms. Cherry A. Murray and Sayeef Salahuddin advocate for a radical co-design approach. This means moving away from treating hardware as a static abstraction and instead designing algorithms that are aware of the underlying physics.
For example, Chen Shen developed a photonic chip that solves partial differential equations by mimicking physical processes like heat diffusion with light. This isn't a digital simulation; it is an analog mapping where the mathematical equivalence between Maxwell’s equations and the target PDE allows the chip to "solve" the problem at the speed of light.
Finally, Moritz Brunion suggests a "CMOS 2.0" transition. Instead of a flat "city" of transistors, future chips will be 3D "metropolises," using wafer bonding to stack specialized layers of logic, memory, and routing. This transition from 2D geometric scaling to 3D functional integration represents the next frontier in sustaining the growth of global computing power.