Probabilistic computing is here; here is how it works

Quantum and neuromorphic computing have attracted billions in investment and entered the public zeitgeist on the promise of orders-of-magnitude gains in computing. Even though these technologies are still decades away from commercial availability, they’ve been buoyed by a general acceptance that once they come to fruition, the world will never be the same.

What many people don’t realize, however, is that another novel technology could also offer large gains over classical digital computing and is commercially available today: probabilistic computing. With industry leaders such as Boeing and Bosch already using the technology in production, probabilistic computing is the next computing revolution that most people have never heard of—until now.

This article is a primer for those new to the field, explaining how the technology works, the major players, and why it is the next revolution in computing.

Deterministic approaches in an uncertain world

Every classical computing solution—whether from Intel, AMD, or Arm—follows the same pattern: take an exact value that has no other metadata, process it, and return another single, exact value. Underneath the hood, the transistors do not know whether the number they’re multiplying is the mass of an aircraft component measured to four decimal places or the output of a neural network that is 73% confident. To the silicon, both look the same: a scalar, with no attached metadata of how certain or uncertain that number actually is.

But in the real world, values are almost never that clean. Sensors drift, measurements and manufacturing carry tolerances, and neural networks output probabilities. Put bluntly: Every meaningful input to a real engineering problem is, at heart, a distribution (i.e., a range of possible values with a shape).

As a result, the classical computing approach accepts that software must first flatten that distribution into a single number before the hardware can do anything with it. Or, developers can represent that distribution with many points, requiring many calculations to carry the information forward at each computational step. Either approach is far from ideal.

To regain a real-world representation, engineers must implement algorithms that recover the full range of plausible outcomes. The most common technique is Monte Carlo simulation, which runs the same calculation thousands of times against slightly varied inputs and aggregates the results. Invented in the 1940s, it’s still the dominant (and unavoidable) way to recapture the information that today’s computing systems throw away by default.

But Monte Carlo is algorithmically expensive, with its accuracy improving only as the square root of the number of samples. That means that 100× more iterations buy 10× better accuracy, and doubling the accuracy on a single calculation requires four times the compute and energy. The cost of precision grows quadratically, and, at production scale, that growth is a wall.

At large investment banks, for example, Monte Carlo compute alone can consume hundreds of millions of dollars per year on workloads that have to run overnight, every night, across every position in the book. In industrial AI, such calculations are too expensive to run at the edge, so sensor data must travel to the cloud and back, slowing real-time control, consuming additional energy, and exposing data to security risks.

Probabilistic computing solves these same problems in hardware rather than through algorithms.

In general, two distinct approaches to probabilistic computing have gained traction: thermodynamic and distribution-extended (UxHw) computing. And, while the two share the same high-level goal, they differ in implementation, performance, and commercial availability.

The thermodynamic approach to probabilistic computing seeks to harness noise and the properties of electrical circuits as dynamical systems. In most cases, the noise is either harnessed from the thermal noise inherent in all electronic systems or purposefully injected as part of the process of solving a computational problem.

In practice, companies pursuing this idea design analog circuits whose electrical behavior (e.g., voltages, currents) naturally mirrors a target probability distribution in the presence of circuit noise. Once the circuit reaches equilibrium, capture its state at a given location to obtain one sample from the solution distribution. Leave the circuit to run, and it accumulates these samples over time, effectively achieving what digital Monte Carlo methods do on a traditional processor.

Whereas digital processors would require thousands of Monte Carlo iterations limited by the clock frequency, thermodynamic computing completes operations much more quickly: the rate at which voltage and current change in a circuit.

Some of the leading companies in this category include:

The appeal of this approach is genuine, as, if the physics works at scale, the energy efficiency could be impactful. In practice, the challenge is that it’s extremely difficult to manufacture analog circuits with the predictability that digital workloads demand. To our knowledge, no company in this category has shipped a commercial product or published benchmarks against state-of-the-art digital processors yet.

The distribution-extended hardware (UxHw) path starts from the opposite premise. Rather than designing analog circuits and sampling distributions, UxHw keeps the silicon digital and standards-compatible while redefining what a “value” is at the architectural level.

In a UxHw processor, a variable is represented both as its usual single value and with additional metadata. The metadata is a distribution over all possible values it could take, with a weighted likelihood assigned to each. For example, a sensor reading of “10 with some noise” might be stored both as the value that would be used in a traditional processor, as well as with metadata that indicates that there is a small probability of being 9.8 or 10.2, a higher probability of being 10.0, and so on across the full range of plausible values.

When a program asks the processor to add two such variables, the hardware adds both components, considering all possible pairings. It then computes the sum for each pairing and weights each sum by the likelihood of that pairing. The output is a new distribution shape, or the distribution of every value the sum could take, with the correct probability attached to each one. Multiply, subtract, divide, and every other operation works the same way. Each applies its arithmetic to every plausible combination of inputs, yielding the resulting distribution of outputs.

The holy grail is to represent these metadata distributions as compactly as possible and to perform arithmetic on them efficiently.

In addition to always providing the same answers as traditional processors, UxHw also produces the distribution metadata. Importantly, this happens in a single pass, as the hardware performs exact arithmetic on the distribution shapes themselves, rather than drawing random samples and aggregating them toward an answer. And because the math is deterministic, the same input distributions always produce the same output distribution on every run and chip, without sampling, iteration, or worrying about convergence.

The appeal of UxHw is its practicality. Because UxHw silicon is standard digital CMOS, chips share the manufacturability, repeatability, and standards compatibility of every existing modern processor and can run software that developers can already write. The technology is also scalable, from energy-constrained sensor systems and mobile robots to high-performance computing for engineering and quantitative finance workloads.

For this reason, UxHw is already commercially available. Currently, Signaloid has three commercial form factors: a cloud platform integrated with AWS (with named customers including Boeing and CERN), a microSD-form-factor edge hardware module available through Mouser Electronics, and a sub-10-watt ASIC taped out with TSMC in May 2026.

As the first probabilistic computing solution to reach the market, UxHw has already proven real-world results. So far, benchmarks against Intel Xeon-class hardware at the same power dissipation show the following speedup

Every prior compute paradigm shift moved through the same phase. GPUs were first bought to run video games before they were used to run physics simulations or neural networks. TPUs existed inside Google for years before most engineers knew the acronym. CUDA was a niche programming model before it became the basis of an entire industry.

Probabilistic computing is in that phase now. UxHw technology is already available for deployment in server racks, on factory floors, and custom UxHw ASICs are in fabrication at TSMC.

COMPANIES: BOEING , CERN , GOOGLE , TSMC

Phillip Stanley-Marbell is the founder and CEO of Signaloid. Up until September 2025, he was a full professor (chair of physical computation) in the Electrical Engineering Division of the Department of Engineering at the University of Cambridge, where he led the Physical Computation Laboratory. He completed his Ph.D. at Carnegie Mellon and has held full-time positions across industry and academia, including at Bell Labs, IBM Research Zurich, Apple, and MIT. Follow Phillip on LinkedIn

Probabilistic computing represents a pivotal shift in how we approach complex computations. By integrating uncertainty directly into the hardware, companies like Signaloid are not just improving efficiency but redefining what's possible in fields like finance and AI. Exciting times ahead!

You must Register or Login to post a comment.

This site uses Akismet to reduce spam. Learn how your comment data is processed.