Source-linked AI summary
Memcomputing: a computing paradigm to store and process information on the same physical platform
M. Di Ventra, Y. V. Pershin
TL;DR
Memcomputing addresses the separation of information storage and processing in conventional computers by using memory circuit elements that perform both functions on one physical platform. The paper defines design criteria for memcomputing architectures and demonstrates a memristive-network solution to the shortest-path problem, including healing after damage. It presents memelements as nanoscale devices supporting adaptive, massively parallel and biologically inspired computation.
Problem
Conventional computing separates storage from processing, creating data-transfer limitations, while an alternative should combine both functions in an intrinsically massively parallel architecture.
Method
The paper introduces memcomputing with memristive, memcapacitive, and meminductive elements, formulates implementation criteria, and applies a memristive network to shortest-path computation.
Results
The memristive network connects two selected nodes through memristive devices in the ON state, and a pulse produces a new shortest path around a damaged region.
Takeaways & Limitations
Memelements provide a platform for storing and processing information together, with collective dynamics supporting adaptive, massively parallel, and biologically inspired computing.
Abstract
from arXiv · showhide
In present day technology, storing and processing of information occur on physically distinct regions of space. Not only does this result in space limitations; it also translates into unwanted delays in retrieving and processing of relevant information. There is, however, a class of two-terminal passive circuit elements with memory, memristive, memcapacitive and meminductive systems -- collectively called memelements -- that perform both information processing and storing of the initial, intermediate and final computational data on the same physical platform. Importantly, the states of these memelements adjust to input signals and provide analog capabilities unavailable in standard circuit elements, resulting in adaptive circuitry, and providing analog massively-parallel computation. All these features are tantalizingly similar to those encountered in the biological realm, thus offering new opportunities for biologically-inspired computation. Of particular importance is the fact that these memelements emerge naturally in nanoscale systems, and are therefore a consequence and a natural by-product of the continued miniaturization of electronic devices. We will discuss the various possibilities offered by memcomputing, discuss the criteria that need to be satisfied to realize this paradigm, and provide an example showing the solution of the shortest-path problem and demonstrate the healing property of the solution path.
I. INTRODUCTION
Memcomputing is introduced as a paradigm that combines information storage and processing in the same physical platform, addressing data-transfer limits in conventional architectures. It uses nanoscale two-terminal memory elements whose adaptive, collective dynamics support massively parallel computation, illustrated through shortest-path optimization and path healing.
- Motivation: Conventional von Neumann systems separate memory from CPU processing, making data transfer between them a fundamental computational bottleneck.The architecture also imposes limits on performance and scalability.
- Memcomputing: Memcomputing uses memelements to realize information storage and processing on the same physical platform without relying on active elements as the main operating tools.The paradigm is motivated by biological systems in which processing, storage, collective operation, and adaptation coexist.
- Memelements: Memelements are two-terminal devices whose resistance, capacitance, or inductance retains information about prior dynamics and arise naturally from delayed electronic or ionic responses at the nanoscale.Their input may be voltage, charge, current, or flux, with the complementary constitutive variable as output.
- Architecture: Massively parallel processors combine arrays of memelements with conventional circuit components, using externally driven collective evolution to compute while storing intermediate information.The paper presents this architecture as a route to increased computational power for graph-theory optimization problems.
- Application: The paper applies memcomputing to the shortest-path problem and examines self-reinforcement, a dynamic feature associated with ant-colony-like collective behavior.It also demonstrates healing of the solution path, connecting the approach to biologically inspired computation.
II. MEMCOMPUTING CRITERIA
The paper formulates design criteria for memcomputing architectures that combine storage and computation in scalable, massively parallel memelements networks. These criteria address retention, initialization, collective dynamics, readout, and robustness, while the processor architecture uses memristive network units and switches.
- 1. Scalable massively-parallel architecture with combined information processing and storage: Memcomputing architectures should combine information processing and storage in scalable massively parallel collections of memristive, memcapacitive, or meminductive elements.Many memelements are expected to participate in parallel computation, creating a potential advantage over sequential processing.
- 1. Scalable massively-parallel architecture with combined information processing and storage: Combined processing and storage can reduce component counts, increase integration density, and support more complex circuit connectivity.Memristive devices can serve simultaneously as gates and latches, and combining memelements can improve logic-circuit performance.
- 2. Sufficiently long information storage times: Memelements should retain information for longer than the calculation time, preferably through non-volatile storage, while also supporting low power and short read/write times.Reported nanoionic examples include write speeds below 10 ns, endurance above 10^5 cycles, retention around 7 years, and scaling below 30 nm.
- 3. The ability to initialize memory states: Relevant memelements must be initialized before computation, either automatically for volatile devices or through an explicit initialization mechanism.Initialization may be unnecessary for elements already storing intermediate or final parameter values.
- 4. Mechanism(s) of collective dynamics, strong "memory content": Collective dynamics should make each memdevice state depend on several or all other devices, with strong memory content supporting reliable switching.In memristive logic, the state of one device can change according to the voltage and state of another.
- 5. The ability to read the final result (from relevant memelements): The final result should be readable without substantially modifying individual memelements, using an input that keeps their states constant or changes them only slightly.Threshold devices generally facilitate this non-destructive readout.
- 6. Robustness against small imperfections and noise: Memcomputing architectures should tolerate small fabrication imperfections and relatively small damages without significant loss of operation.This robustness criterion is motivated partly by the brain’s continued functionality despite neuron loss.
- Processor architecture: The illustrated memristive processor uses a grid network whose basic units contain two bipolar memristive devices connected in parallel and two switches.The switches access individual devices, while the parallel connection symmetrizes bipolar responses.
III. MEMCOMPUTING SCHEMES
The paper surveys memcomputing schemes based on memristive networks, synapses, logic, and cellular automata, while noting that memcapacitors and meminductors can extend these designs. These alternatives may offer low-energy operation and motivate architectures with brain-like computing capabilities.
- Existing schemes: Proposed memcomputing schemes include neuromorphic memristive synapses, massively parallel memristive networks, logic with memory elements, and memristive cellular automata.A cited memristive network applies the paradigm to the maze optimization problem.
- Alternative memelements: Memcapacitors and meminductors can replace memristors in these schemes, although their use requires modified implementations.Their possible use broadens the physical devices available for memcomputing architectures.
- Energy and future directions: Because memcapacitors and meminductors may consume little or virtually no energy, they could make memcomputing more energetically efficient than memristor-based implementations.The paper identifies brain-comparable computing capability and power consumption as an important milestone.
IV. APPLICATION OF MEMRISTIVE NETWORKS TO THE SHORTEST-PATH PROBLEM
A memristive network computes the shortest path by applying a voltage pulse to selected nodes and reading the connected low-resistance units. Its dynamics progressively reinforce the shortest route, quantify path concentration through decreasing network entropy, and can heal damage by finding a shortest detour.
- Algorithm: The processor initializes all devices OFF, applies a voltage pulse to two selected nodes, and reads the shortest path from units switched ON.The network is a square grid whose basic units contain two bipolar memristive elements and two switches.
- Shortest-path solution: A chain of ON memristive units connects the selected nodes, yielding the shortest-path solution after the applied pulse.The initial network has all devices OFF; the final state represents each basic unit through its memristance.
- Collective dynamics: Current initially flows through multiple routes, but the least-resistive path switches faster and attracts more current, reinforcing that route during computation.This feedback resembles trail reinforcement in ant-colony optimization.
- Collective dynamics: The solution emerges from both selected nodes and propagates toward the center, while network entropy decreases as the number of available current paths contracts.The entropy decrease is more pronounced for larger memory content, represented by R_off/R_on.
- Memory content: With low memory content, R_off/R_on = 1.25, many devices switch ON and the shortest path cannot be found exactly at the given bias.This behavior demonstrates the importance of strong memory content for accurate solutions.
- Healing: After three central grid points are removed, a single input-output pulse creates a new path around the damage along the shortest possible detour.The detour is reinforced because current flows through all possible routes, with the shortest receiving the strongest effect.
V. CONCLUSION
The paper presents memcomputing as storing and processing information on the same physical platform, with applications including digital logic and shortest-path computation with healing properties.
- Memcomputing combines information storage and processing on the same physical platform.
- The paradigm is described as bypassing several bottlenecks of present-day computing architectures.
- Memristors, memcapacitors, and meminductors are described as constitutive units already widely available for some applications.
- These elements emerge naturally with increasing miniaturization of electronic devices.
- The paradigm’s similarities to brain features and collective biological behavior motivate possible directions in neuromorphic architectures and biological studies.
Methods
The methods use threshold current-controlled bipolar memristive devices in an 11×11 network, with memristive states updated through a threshold switching model and network potentials solved from Kirchhoff’s current law.
- The numerical results use current-controlled bipolar memristive devices with threshold switching.The model uses a threshold current and bipolar device behavior.
- The memristive state update is governed by a threshold rule that changes the state when |IM| exceeds It.The rate parameter γ controls the memristance change above the threshold current It.
- All numerical results use an 11×11 memristive network.
- The simulations use RM off = 200 Ohms, γ = 106 Ohms/(s·A), and It = 10mA as model parameters.
- At each time step, grid-point potentials are obtained by solving Kirchhoff’s current-law equations with a sparse matrix technique.The corresponding memristive-state change is then computed from the model.