Heat as a Computer: MIT Engineers Build Silicon That Calculates Without Electricity

MIT researchers designed silicon microstructures that perform matrix-vector multiplication using heat flow instead of electrical signals. No transistors, no code — the physics of heat propagation is the computation itself.

📅

✍️ Gianluca

Heat as a Computer: MIT Engineers Build Silicon That Calculates Without Electricity

In January 2026, researchers at MIT published a remarkable study: they designed silicon microstructures that perform mathematical operations using heat flow rather than electrical signals. No transistors. No clock cycles. No code. The physics of heat propagation through a carefully shaped piece of silicon is the computation itself.

The work was led by Caio Silva, an undergraduate student in MIT's Department of Physics, with Giuseppe Romano, a research scientist at MIT's Institute for Soldier Nanotechnologies, as senior author. Their approach opens a new branch of analog computing where waste heat, traditionally the enemy of every chip designer, becomes the signal carrier.

Key results from the MIT study:

  • Matrix-vector multiplication performed with over 99% accuracy
  • Silicon structures sized roughly at the scale of a dust particle
  • Input data encoded as temperature sets using existing chip waste heat
  • Output collected as thermal power at a fixed thermostat temperature
  • No additional energy consumption required for the computation itself

What Is Thermal Computing?

In digital computing, information travels as electrons. Ones and zeros are encoded as voltage levels, and logic gates switch states to execute operations. In thermal computing, the signal carrier is entirely different: it is heat.

The core idea works like this. The input data is represented as a set of temperature differences applied to specific points on a silicon structure. The geometry of the structure (its internal channels, voids, and material densities) is designed so that heat propagates through it in a mathematically precise way. By the time the heat reaches the output points, it has been transformed into the result of an operation.

This is called analog thermal computation. Values are not binary bits but physical gradients. The computation does not run; it simply happens, as a consequence of physics.

Inverse Design: When AI Builds the Physics

The most challenging part of this approach is not the physics; it is the design. How do you shape a piece of silicon so that the way heat moves through it equals a specific mathematical operation?

Traditional engineering works forward: you propose a structure, simulate its behavior, and check whether it does what you want. For thermal computation, that direction is impractical. The search space of possible geometries is enormous, and intuition alone cannot find the right pattern.

The MIT team used a method called inverse design, which reverses the process entirely:

  • 1. Define the target behavior

    You specify the mathematical operation you want the structure to perform, for example a specific matrix-vector multiplication.

  • 2. Let an algorithm find the geometry

    An optimization algorithm using techniques like gradient descent, topology optimization, or physics-informed neural networks iterates over thousands of candidate geometries, simulating heat flow each time and comparing the output against the desired result.

  • 3. Converge on the optimal pattern

    Over many iterations, the algorithm converges on a porous silicon pattern, a microscopic arrangement of channels and voids, that produces exactly the required thermal transformation.

The result is a physical object that encodes a mathematical operation in its shape. You cannot see the math, but it is there, frozen in the geometry of the silicon.

How Heat Becomes a Matrix Multiplication

Matrix-vector multiplication is one of the most fundamental operations in mathematics and computing. It underlies neural networks, signal filtering, linear transformations, and control systems. The standard form is:

y = M · x

where x is the input vector, M is the matrix, and y is the output vector

In thermal computing, this maps directly onto physics:

Mathematical conceptPhysical equivalent
Input vector xSet of input temperatures applied to specific points on the structure
Matrix MThe geometry of the silicon structure: channels, porosity, thickness distribution
Output vector yThermal power collected at output points held at a fixed reference temperature
ComputationHeat diffusion through the structure, governed entirely by physics, with no active components

There is no code executing this operation. There is no clock ticking, no instruction pointer moving through memory. The linear transformation happens because the geometry makes it thermodynamically inevitable.

Why the Silicon Must Be Precisely Patterned

Heat does not flow randomly. It follows the diffusion equation:

∂T/∂t = α ∇²T

where T is temperature, t is time, and α is thermal diffusivity of the material

By modifying the local structure of the silicon, introducing pores, varying thickness, creating internal channels, the researchers change α at every point in the material. This in turn changes how heat propagates spatially. The spatial distribution of the output temperature is, mathematically, the result of a linear transformation applied to the input temperatures.

One important technical constraint: the laws of heat conduction allow only positive coefficients in the transformation. To handle matrices with negative entries, the team represented each target matrix as the difference of two positive-coefficient structures, computing both independently and subtracting the outputs.

Current Limitations

The results are impressive at small scale, but several fundamental constraints limit where thermal computing can go today:

  • Heat is orders of magnitude slower than electrons

    Thermal diffusion in silicon operates in microseconds to milliseconds. Electrical signals in modern chips switch in picoseconds. For any application requiring high throughput, such as real-time inference on large models, thermal computing at the speed of heat is simply too slow.

  • Each structure encodes one operation only

    The geometry of the silicon is fixed at fabrication time. A structure designed to multiply by matrix M cannot be reprogrammed to multiply by a different matrix. If you need a different operation, you need a different physical object. Unlike a CPU or GPU, there is no programmability.

  • Accuracy decreases with matrix complexity

    The 99%+ accuracy achieved by the team applies to simple matrices with two or three columns. Larger and more complex matrices introduce thermal noise and boundary effects that reduce precision. Deep learning models typically require thousands of matrix dimensions.

  • Scalability requires millions of tiled structures

    For large-scale applications, many such structures would need to be tiled and interconnected. Manufacturing and integration at that density is not yet practical. The bandwidth of the thermal channel is also inherently limited, requiring significant expansion for anything resembling a machine learning workload.

Where This Fits in the Landscape of Analog Computing

Thermal computing is one point in a broader emerging space where physical phenomena replace digital logic for specific operations. Each approach trades generality and precision for energy efficiency and physical compactness:

ApproachSignal carrierBest suited for
Thermal computing (MIT)Heat gradientsPassive sensing, thermal management, low-power linear operations
Memristor crossbarsElectrical current (analog)In-memory matrix multiplication, neural network inference
Optical computingLight (photons)Ultra-high-speed linear algebra, signal processing
Neuromorphic chipsSpiking electrical signalsEvent-driven inference, pattern recognition, low power edge AI
Computing in memory (CIM)Electrical (in SRAM/DRAM)Reducing data movement for matrix-heavy workloads

What makes the thermal approach genuinely novel is that it requires no active energy input for the computation itself. It harvests and redirects the heat already produced by surrounding electronics. This is as close to "free computation" as physics currently allows, within a very narrow problem domain.

A Reflection: Could Future Chips Ever Reconfigure Their Physical Geometry?

The fundamental limitation of thermal computing, and of most analog accelerators, is that the physical structure is fixed at manufacturing time. This raises an interesting question: could we ever build hardware where the geometry itself can change on demand, effectively selecting which mathematical operation to perform?

At first glance, this sounds like science fiction. Silicon is rigid, and modern transistors operate at nanometer scales where mechanical movement seems impossible. But looking at the trajectory of materials science, the idea is less distant than it appears, though the path there is anything but straightforward.

The core idea:

Instead of a fixed geometry encoding a single matrix, imagine a structure whose internal configuration can be altered electrically, changing which mathematical operation the heat flow performs. Not reprogramming in software. Reprogramming in physics.

Technologies That Are Already Moving in This Direction

Several existing research areas point toward a future of physically reconfigurable computation, each approaching the problem from a different angle:

  • Phase-Change Materials (PCM): GST and Related Compounds

    Materials like germanium-antimony-telluride (GST) can switch between amorphous and crystalline states with an electrical pulse. This changes their thermal and electrical conductivity. Each local region of a chip can be individually switched, effectively changing the thermal properties at that point, and therefore changing how heat flows through the structure. The geometry stays fixed, but the effective physics changes. GST is already used in experimental neuromorphic systems as analog weight storage.

  • Memristor Crossbar Arrays

    Memristors implement y = Wx directly via Ohm's law and Kirchhoff's current law, where the resistance of each device encodes a matrix weight. Crucially, those resistances are programmable: applying a voltage pulse changes the resistance, changing the matrix. This is already a working example of a reconfigurable physical operator, albeit electrical rather than thermal.

  • MEMS and Nano-Electromechanical Systems (NEMS)

    Micro-electromechanical systems can physically move structures at micron and sub-micron scales in response to electrical signals. In principle, a MEMS-based thermal router could open or close thermal channels, changing the effective geometry seen by heat. The engineering challenges are severe: thermal stability, actuation speed, and device density at chip scale. But foundational physics does not forbid it.

  • Electro-Active and Electro-Caloric Materials

    Electrocaloric materials change temperature in response to an electric field. Piezoelectric and electroactive polymers deform mechanically under voltage. At sufficient miniaturization, these could alter local geometry or thermal boundary conditions in a programmable array. The materials science is progressing, but integration at semiconductor manufacturing scales remains a significant open problem.

  • Programmable Metamaterials

    Thermal metamaterials are structures engineered to guide, concentrate, or block heat in ways that bulk materials cannot. Research groups have demonstrated thermal cloaks, concentrators, and rectifiers using layered or patterned media. The next step, active electrical control over which metamaterial behavior is active, is being explored, though not yet at chip integration scales.

The Realistic Near-Term Path

Changing macroscopic geometry at nanometer scales under electrical actuation, fast enough to be useful for computation, is an extraordinary engineering challenge. Silicon is not elastic at chip dimensions, and the thermal time constants of any moving structure would likely be too slow to matter.

The more realistic near-term direction does not change the shape. It changes the local physical properties of the material at specific points, effectively achieving the same result. A chip with a fixed geometric scaffold but electrically tunable thermal properties at each node would behave like a reconfigurable thermal operator without any moving parts.

This converges toward a concept already familiar in the analog AI accelerator community: a programmable analog physical operator, a structure where the weights (encoded in material properties rather than resistance values) can be updated, and the computation is performed passively by physics.

The plausible future architecture:

  • Fixed geometric scaffold etched into silicon or another substrate
  • Grid of phase-change or electrocaloric nodes distributed through the structure
  • Electrical control layer that sets the state of each node
  • Thermal computation layer that executes the operation passively
  • Digital readout layer that converts thermal output to digital values

Why Precision Remains a Hard Wall

Any physical analog computation, thermal, electrical, or optical, faces a fundamental ceiling on precision. Digital systems can achieve arbitrary accuracy by adding more bits. Analog systems are bounded by noise, thermal drift, fabrication variance, and material aging.

For applications in machine learning, where operations like attention and feed-forward layers tolerate quantization down to 4 or even 2 bits, this is acceptable. The precision is sufficient and the energy savings are real. For scientific computing or cryptographic applications, it is not.

The honest engineering summary is: analog thermal computation (and analog computation broadly) will likely become a complement to digital systems, not a replacement. A chip that uses heat to perform its linear algebra and electricity to handle control flow and precise arithmetic is a more realistic long-term picture than a chip that has abandoned transistors entirely.

Key Takeaways

  • MIT demonstrated matrix-vector multiplication at 99%+ accuracy using only heat flow through patterned silicon
  • Inverse design is the enabling technology: AI finds the geometry that makes the physics execute a specific operation
  • No energy input is required for the computation; it uses waste heat already present on the chip
  • Heat is fundamentally slow, limiting this to passive sensing, diagnostics, and low-throughput linear operations for now
  • Reconfigurable physical computation is a real research direction, most plausibly via phase-change materials and programmable metamaterials
  • The future is hybrid: digital control, analog physics-based operators, and thermal passivity working together on the same die

Conclusion

The MIT thermal computing work is a genuinely novel proof of concept: heat, properly directed through a designed structure, can execute a linear algebra operation with high accuracy and zero active power. It will not replace CPUs or GPUs. It will not run large language models. What it does is demonstrate that computation does not have to mean electricity, and that the physical world around a chip, including the heat it wastes every second, can be structured to do useful work.

The longer-term question raised by this research is even more interesting: as materials science advances, could we build structures whose physical operators can be electrically reconfigured? The honest answer is that we are not there yet, and the path is genuinely difficult. But between phase-change materials, programmable metamaterials, and memristor arrays, the building blocks of that future are already on the bench. What we are watching, in slow motion, is the beginning of hardware that computes with the laws of physics rather than in spite of them.

Resources and Links