In conventional digital computing, transistors act as binary switches. We force them to be either completely "on" or completely "off," abstracting away the underlying physics to build a predictable world of 1s and 0s. To compute a mathematical function like an exponential or a logarithm, a digital processor must execute a series of logical operations through thousands of these switches, consuming significant time, silicon area, and energy.
But what if we didn't abstract the physics away? What if we let the physical behavior of the silicon do the math for us?
When a metal-oxide-semiconductor field-effect transistor (MOSFET) is biased below its threshold voltage, it enters the subthreshold or weak inversion regime. In this state, the transistor is technically "off" by digital standards, but a tiny, sub-microampere current still flows through the channel. Crucially, this current scales exponentially with the gate voltage. By operating in this regime, we can perform complex mathematical transformations directly through the native electron dynamics of the hardware.
The Physics as the Math
In the subthreshold region, the current flowing through a transistor channel is governed by carrier diffusion rather than drift. Electrons must overcome an energy barrier determined by the gate voltage. According to statistical mechanics, the number of charge carriers with enough thermal energy to cross such a potential barrier follows a Boltzmann distribution. This manifests as an inherent exponential relationship between the gate voltage ($V_g$) and the drain current ($I_d$).
As Lazzaro et al. and Chiueh et al. early on recognized, this physical property provides a "free" exponential function. If we invert the relationship—measuring the voltage required to sustain a specific current—we get a natural logarithmic encoding over several orders of magnitude.
This simple mechanism serves as the foundation for executing complex arithmetic functions with minimal hardware overhead:
- Multiplication and Division: By arranging subthreshold transistors into translinear loops, we can add or subtract gate voltages. Because voltages map exponentially to currents, adding voltages equates to multiplying their corresponding currents ($e^A \cdot e^B = e^{A+B}$). Chakrabartty et al. utilized these translinear properties to compute complex kernel operations for support vector machines at sub-microwatt levels, while Liu et al. leveraged them for high-speed cosine similarity searches directly within memory arrays.
- Summation: By connecting multiple transistor drains to a single wire, Kirchhoff’s current law performs instantaneous addition ($I_{\text{total}} = \sum I_n$) without requiring any logic gates, a principle highlighted by Weller et al..
- Integration and Diffusion: Combining the subthreshold resistance of a transistor with a capacitor creates a continuous-time integrator. Cauwenberghs demonstrated how grids of these components can solve complex differential equations for image diffusion directly via current flow.
Emulating Biological Dynamics
The biological brain is an exceptionally low-power computer, operating on roughly 20 watts. It achieves this because its underlying mechanisms—such as the opening and closing of voltage-gated ion channels—rely on thermodynamic barriers similar to those in subthreshold transistors. Neuromorphic engineers exploit this parallel to build hardware that directly emulates biology rather than simulating it numerically.
For instance, Millner et al. used the subthreshold exponential current-voltage relationship to replicate the rapid, exponential rise of action potentials in the Adaptive Exponential Integrate-and-Fire (AdEx) neuron model. To capture the temporal aspects of the nervous system, Liu et al. and Whatley et al. showed that weak-inversion circuits can natively implement biological time constants for membrane integration, leak, and refractory periods.
Synaptic dynamics also map cleanly to this regime. Park et al. demonstrated that subthreshold transistors can linearly emulate biological conductance within the log-domain, providing a massive dynamic range for synaptic signals while keeping energy consumption down to a mere 22 pJ per spike. For learning, Indiveri designed compact subthreshold circuits that implement spike-timing-dependent plasticity (STDP) while solving the problem of analog memory leakage through bistable states.
From Physics to Algorithms
Beyond mimicking neurons, subthreshold computation can be mapped directly to probabilistic and statistical algorithms.
Consider the softmax function, ubiquitous in machine learning for normalizing scores into probabilities. Elfadel et al. derived the softmax function from statistical mechanics and showed how an analog circuit element—the "entropic resistor"—computes parallel exponentials and sums within a single transistor time constant.
Similarly, Lazzaro demonstrated that the exponential current-voltage curve can be directly mapped to the logarithmic calculations required for Hidden Markov Model (HMM) state decoding. By replacing power-hungry digital arithmetic units with subthreshold circuits, the author built a wordspotting state decoder that operates in the nanowatt range, making always-on voice interfaces highly efficient.
For spatial and visual processing, Koch et al. used subthreshold transistors as non-linear resistors in massively parallel networks to compute optical flow in real-time. Harrison built a bio-inspired visual collision detector mimicking a fly's motion detection system, consuming only 18 nanowatts per pixel by mapping the underlying mathematical correlation directly onto subthreshold transistor nonlinearities. Liu et al. further showed that these exponential characteristics can natively approximate Gaussian curves to perform real-time Difference-of-Gaussian feature extraction on flexible thin-film substrates.
Taming the Imperfections of Analog Silicon
If subthreshold analog computing is so energy-efficient, why hasn't it completely replaced digital processors? The primary challenge lies in device mismatch and environmental sensitivity. Because the subthreshold current depends exponentially on physical parameters like the threshold voltage and temperature, even microscopic variations during manufacturing can cause two identical-looking transistors to behave differently. Furthermore, as Ribar et al. noted, subthreshold circuits are highly sensitive to temperature fluctuations, which can drastically alter circuit behavior.
To make these systems practical, researchers design adaptive, self-calibrating hardware loops. For example, Iizuka et al. introduced a frequency-divided feedback loop into a Winner-Take-All network to automatically compensate for manufacturing defects and fluctuating signal levels. Harris et al. developed a bio-inspired adaptation circuit that enforces a "constant average statistics constraint" to continuously correct for gain and offset errors in sensor arrays in real-time. On the software side, Weller et al. demonstrated that "variation-aware" training algorithms can anticipate these hardware imperfections, allowing printed analog neuromorphic systems to maintain high accuracy despite the high variability of low-cost manufacturing.
Key Foundational Work
Computing Motion Using Resistive Networks (1987) — contributed 2 concept elements to this direction.
Winner-Take-All Networks of O(N) Complexity (1988) — contributed 2 concept elements to this direction.