Computing Paradigms

A model of computation defines a computational step. An architecture organizes state, control, communication, and execution. A physical embodiment recruits a material process to realize one or more models. They belong in the same survey, but not in the same category.

“Models of Computation” would therefore be too narrow. Compute-in-memory and photonic computing, for example, are not complete formal models merely because they depart from the stored-program machine. “Computing Paradigms” covers models, architectures, and embodiments while keeping those distinctions visible.

The practical question is:

Which alternatives to stored-program, von Neumann-style computation have progressed from formal proposal to physical demonstration, integration, scale, external access, and recurring use?

Status was checked on 2026-08-11. Vendor milestones remain vendor reports unless independently reproduced.

Maturity scale

The matrix uses the same substrate-neutral stages for every family:

  1. Reference — an executable implementation exposes the semantics, usually on a conventional host.
  2. Physical — a purpose-built system realizes the relevant primitive rather than merely simulating it.
  3. Integrated — state, execution, communication, and control form a coherent device or system.
  4. Scaled — many elements operate through a demonstrated composition or interconnect.
  5. Access — outsiders can program or run the relevant implementation.
  6. Operational — it performs recurring work beyond a one-off demonstration.
demonstrated limited, restricted, or specialized no public evidence located opens claim, evidence, criticism, and sources

A mark records the strongest public implementation found. It does not transfer one project’s achievement to every machine using the same label.

Execution models

These families primarily change what causes a computational step.

Scroll horizontally to compare all six stages.

“Access” for interaction nets means access to software evaluators, not purpose-built hardware.

Architectural and physical paradigms

These families primarily change where state and work reside, or which physical dynamics perform the operation.

Scroll horizontally to compare all six stages.

Evidence and open tests

Each entry uses the same four fields. “Next test” means the result that would materially strengthen or weaken the implementation claim.

Spatial dataflow

Evidence
Dataflow graphs, Kahn process networks, and synchronous dataflow have mature software and FPGA toolchains. SambaNova and NextSilicon report operational reconfigurable dataflow processors.
Embodiments
CMOS ASICs, FPGAs, systolic arrays, and coarse-grained reconfigurable arrays.
Boundary
Irregular control, mutable pointer structures, reconfiguration cost, and dependence on a stored-program host.
Next test
A broad workload suite running without application-specific rewriting or a supervisory conventional processor.

Interaction nets

Evidence
HVM2 and Vine execute the model on CPUs and GPUs. Tendrils says it is building dedicated hardware, but has disclosed no physical demonstration.
Embodiments
Conventional digital processors currently executing a graph-rewriting abstract machine.
Boundary
Placement, allocation, routing, reuse, and readback of a dynamically changing graph.
Next test
A purpose-built core running nontrivial programs with measured energy, memory, and routing costs.

Functional graph reduction

Evidence
Heron is an FPGA processor core for pure non-strict functional languages; HAFLANG continues the hardware line.
Embodiments
FPGA fabric and conventional processors.
Boundary
Allocation, garbage collection, locality, and dynamically varying parallelism.
Next test
A many-core reducer whose end-to-end gains survive allocation, communication, and collection.

Cellular automata

Evidence
Universal variants are established, with purpose-built FPGA systems and memristive in-memory demonstrations.
Embodiments
FPGA, CMOS, and memristive arrays.
Boundary
Compiling ordinary algorithms into local rules while preserving locality; global communication and input/output.
Next test
A programmable system where useful algorithms remain local after compilation.

Neuromorphic and spiking

Evidence
Intel’s Hala Point integrates 1,152 Loihi 2 processors for research workloads.
Embodiments
Digital CMOS, mixed-signal circuits, and emerging memory devices.
Boundary
Programming portability, training, numerical comparison, and workload generality.
Next test
Independently reproduced gains on tasks whose event structure was not selected to flatter the hardware.

Compute-in-memory

Evidence
IBM fabricated a 64-core phase-change-memory inference chip.
Embodiments
SRAM, phase-change memory, ReRAM, memristive arrays, and mixed-signal CMOS.
Boundary
Conversion overhead, precision, endurance, variability, and incomplete control semantics.
Next test
End-to-end gains including conversion, calibration, memory updates, and host coordination.

Photonic analog computing

Evidence
A single-chip photonic neural network has integrated optical matrix operations and nonlinear activation.
Embodiments
Silicon photonics, lasers, modulators, detectors, and optical memory elements.
Boundary
Electronic input/output, precision, nonlinearity, storage, calibration, and thermal stability.
Next test
An externally programmable system whose advantage includes electronic interfaces and control.

Thermodynamic and probabilistic

Evidence
Extropic reports X0 silicon and the XTR-0 prototype platform. Normal Computing reports a 2025 CN101 tape-out. THRML provides a conventional reference environment.
Embodiments
Stochastic CMOS, with magnetic and other noisy devices also under investigation.
Boundary
Narrow sampling semantics, calibration, scaling, and limited independent measurement.
Next test
Reproducible application-level advantage on externally selected distributions and workloads.

Reversible and adiabatic

Evidence
Vaire reports an energy-recovery proof of concept; no externally accessible system has been disclosed.
Embodiments
CMOS adiabatic logic, with superconducting implementations also proposed.
Boundary
Clock and interconnect losses, area, latency, errors, and incomplete system-level energy accounting.
Next test
Net energy recovery at system scale after control, memory, communication, and error handling.

Quantum computing

Evidence
Circuit, measurement, and annealing models have physical systems and mature software stacks; IBM reports a decade of cloud-accessible quantum processors.
Embodiments
Superconducting circuits, trapped ions, neutral atoms, photons, and annealing systems.
Boundary
Error correction, control overhead, data loading, and useful advantage.
Next test
A reproducible application advantage including error correction and classical orchestration costs.

Molecular and chemical computing

Evidence
Chemical-reaction and strand-displacement calculi have physical realizations, including enzyme-powered DNA computing networks.
Embodiments
DNA, enzymes, molecular reactions, and surface-confined systems.
Boundary
Latency, reset, error, cascading depth, automation, and electronic readout.
Next test
A reusable molecular system executing several nontrivial programs without manual laboratory reconstruction.

Reading the result

A physical demonstration establishes that a primitive can exist. It does not establish programmability, scaling, economy, or superiority. Absence of a custom device does not refute a model; it identifies the next criticism the model has not survived.

The comparison is not about exoticness. The relevant question is whether an advantage survives representation, execution, communication, control, and readout.