High-resolution simulation computing
By using a cross-structure array and processor to implement multiple analog levels and dot product engines with memristors, the noise and variability problems of analog computing systems are solved, improving computational efficiency and reliability, and making it suitable for high-bandwidth AI and scientific computing.
Patent Information
- Application Number
- CN202580022228.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-03-21
- Filing Date
- 2025-03-20
- Publication Date
- 2026-11-10
AI Technical Summary
Analog computing systems face noise and variability issues when processing continuous data, leading to inaccurate calculations. Furthermore, equipment variability and manufacturing variations introduce additional errors, limiting their application in mainstream computing.
The analog computing system employs a cross-structure array and processor, utilizes memristors made of compound materials for precise control and addressing, achieves multiple analog levels by applying voltage signals, combines a dot product engine for efficient parallel computing, simplifies read and write operations, and optimizes component size and conductance variations.
It achieves high-resolution simulation computing, reduces computational complexity, improves computational efficiency and reliability, reduces the impact of noise, and provides a high-energy-efficiency computing solution suitable for high-bandwidth AI and scientific computing.
Smart Images

Figure CN123076873A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates generally to simulation calculations, and particularly to embodiments of molecular memristors for high-precision simulation calculations. Background Technology
[0002] Analog computing systems are a type of computing architecture that processes data by manipulating continuous electrical signals, in contrast to the discrete binary data used in digital computing systems. Analog computing systems are designed to handle tasks involving continuous information, performing computational and data processing operations using continuously changing physical phenomena, potentially offering advantages in speed, energy efficiency, and parallel processing for certain types of computing. In recent decades, analog computing systems have regained interest due to their potential to efficiently process continuous data, particularly in applications such as artificial intelligence (AI) and signal processing.
[0003] In modern times, analog computing systems are being integrated with AI in various ways, such as through neuromorphic computing, analog machine learning accelerators, and analog AI chips. Neuromorphic computing aims to simulate the neural structure of the brain by using analog circuits to represent neurons and synapses. On the other hand, analog machine learning accelerators are used to accelerate machine learning algorithms by performing operations such as matrix multiplication and convolution more efficiently. This integration of AI into analog computing systems is particularly beneficial in edge AI, robotics, and healthcare, where real-time processing and low power consumption are crucial. Attached Figure Description
[0004] Now, referring to the accompanying drawings and describing examples of systems and / or methods according to the content of this subject matter, wherein:
[0005] Figure 1 A schematic diagram of an analog computing system for high-resolution analog computing is shown as an example, based on the content of this topic.
[0006] Figure 2 Examples of multiple word lines and bit lines of a cross-structure array in a simulation computing system, based on the content of this topic, are shown.
[0007] Figure 3A The graph illustrates, as an example of the content of this topic, the change in conductance in a memristor as a function of the number of write pulses applied, ranging from the lowest to the highest conductance value;
[0008] Figure 3B The diagram illustrates, based on the content of this topic, a graph depicting the conductance measurements of a memristor between two pairs of consecutive states, labeled Level-2, Level-3 and Level 10k2, Level 10k3, at various levels of data points and counts.
[0009] Figure 3CA graph illustrating the probability distribution of conductance states in a memristor is shown as an example, based on the content of this topic.
[0010] Figure 3D A graph showing the statistical distribution of electrical conductivity levels in a memristor, based on examples from the content of this topic, is provided.
[0011] Figure 3E An example is shown, based on the content of this topic, of a graph illustrating the relationship between the write conductance level and the corresponding read conductance value in a memristor;
[0012] Figure 4A An example is shown, based on the content of this topic, depicting a graph of the endurance of a memristor across read and write cycles;
[0013] Figure 4B An example is shown, based on the content of this topic, of a graph representing the electrical conductivity of a memristor over time;
[0014] Figure 4C The diagram illustrates, based on the content of this topic, two different methods for programming the electrical conductivity levels in a memristor.
[0015] Figure 4D A three-dimensional graph illustrating the relationship between pulse voltage amplitude, pulse width, and resulting conductance difference in a memristor is shown as an example based on the content of this topic.
[0016] Figure 4E An example is shown, based on the content of this topic, of a graph representing the relationship between the transition time and the reciprocal of thermal energy between different subsequent conduction level pairs in a memristor;
[0017] Figure 4F A graphical representation of the conductance modes in a memristor is shown as an example, based on the content of this topic.
[0018] Figure 5A An example is shown, based on the content of this topic, of a graph comparing the Fourier transform (FT) coefficients obtained from a memristor-based hardware implementation (memristor) with the calculated FT;
[0019] Figure 5B The diagram illustrates, based on examples from this topic, graphs comparing the performance of matrix-matrix multiplication across different matrix types using memristor-based computing systems (analog computing systems);
[0020] Figure 5C A graph showing the conductance characteristics of a memristor as a function of the number of applied electrical pulses is shown as an example, based on the content of this topic.
[0021] Figure 6Examples based on the content of this topic are shown in two graphs representing the conductance of a memristor as a function of the number of applied electrical pulses for updating characteristics with 9000 weights from 900 cross points.
[0022] Figure 7 A set of four heatmaps based on examples from this topic compares DFT matrix write accuracy immediately after a write and after a seven-month period.
[0023] Figure 8 An example is shown, based on the content of this topic, illustrating a graphical representation of reconstructing the original waveform using an analog computing system via the inverse discrete Fourier transform (IDFT) process;
[0024] Figure 9 A series of sub-images depicting examples based on the content of this topic are provided, each sub-image consisting of a pair of diagrams.
[0025] It will be noted that throughout the accompanying drawings, the same reference numerals denote similar but not necessarily identical elements. These drawings are not necessarily drawn to scale, and the dimensions of some parts may be exaggerated for clarity of the examples shown. Furthermore, the drawings provide examples and / or embodiments consistent with the specification; however, the specification is not limited to the examples and / or embodiments provided in the drawings. Detailed Implementation
[0026] Typically, analog computing systems are integrated with memristors, non-volatile electrical components used to limit or regulate current flow in circuits. Memristors have become a promising component in analog computing systems due to their ability to retain information even without power. Integrating memristors into analog computing systems enhances performance and efficiency, paving the way for advancements in neuromorphic computing and other innovative applications.
[0027] Furthermore, analog computing systems are well-suited for processing continuous data, which is crucial in AI applications such as neural networks and signal processing. Compared to the discrete data used in digital computing systems, processing continuous data in AI applications provides a detailed and nuanced representation of real-world data.
[0028] Analog computing systems utilize continuous data to perform various computations. One of the key operations in such systems is vector-matrix multiplication (VMM). VMM is a fundamental operation in many computational algorithms, particularly in AI and machine learning algorithms. By using analog signals, analog computing systems can perform VMM with lower power consumption and higher speed compared to digital computing systems.
[0029] However, analog computing systems are highly sensitive to noise and variability, which can lead to inaccuracies in calculations such as Virtual Memory Models (VMMs). Errors in VMM operations can propagate and amplify, causing significant problems in large-scale computations. Furthermore, device defects (such as those found in memristors or resistive random access memory (RRAM)) can introduce additional errors due to manufacturing variations or operational wear. Compared to digital systems, analog computing systems exhibit limited flexibility, making reconfiguration for different computations challenging. Moreover, the physical size and power requirements of analog components can limit a computing system's ability to handle larger datasets or more complex tasks, thus reducing its scalability.
[0030] Traditionally, Dot Product Engines (DPEs) are configured to significantly reduce computation time by leveraging the parallel nature of analog computing to execute Virtual Machines (VMMs) within a single time step. Compared to digital computing systems, DPEs consume less power because they are designed to perform operations directly on stored data without repetitive read / write loops. This parallel processing capability—meaning that multiple operations can be executed simultaneously, rather than sequentially as in traditional digital systems—further enhances the efficiency of analog computing systems. DPEs are particularly beneficial in AI and machine learning for tasks such as neural network inference and real-time data processing, enabling faster and more efficient computation.
[0031] Despite these advancements, the adoption of digital physical devices (DPEs) and other analog computing elements in mainstream computing has been limited by challenges related to accuracy, noise, and device variability. For example, the electrical properties of memristors, the core of DPEs, often exhibit variability, leading to inconsistent performance. Integrating DPEs with existing digital technologies is complex and requires advanced manufacturing techniques. Furthermore, the cost of manufacturing and integrating memristors can be higher than that of traditional digital components, limiting their widespread adoption.
[0032] An example method for improving the performance and reliability of DPEs in analog computing is disclosed. In one example, an analog computing system is described. This analog computing system (hereinafter referred to as the "System") includes a crossbar array and a processor. The crossbar array includes a bottom electrode and a top electrode, wherein an active material is disposed between the bottom electrode and the top electrode. In one example, the active material is selected from the group of compounds, including but not limited to: ruthenium complexes of azo aromatic ligands, coordinated azo aromatic ligands, non-coordinated azo aromatic ligands, or combinations thereof. The bottom electrode, the top electrode, and the active material disposed between the top electrode and the bottom electrode are arranged in such a way that each junction acts as a memristor.
[0033] In one example, the cross-structure array consists of multiple word lines and bit lines. Each memristor in the array is formed at the intersection (junction) of the word line and bit line, creating a grid-like structure that allows for precise control and addressing of the individual memristors. For ease of array operation, the word lines are electrically coupled to the positive terminal, while the bit lines are coupled to the negative terminal. This configuration allows a voltage difference to be applied to a specific memristor within the array.
[0034] During operation, an input voltage signal can be applied to the cross-structure array. This application of the input voltage signal causes molecular transformations within each memristor. These molecular transformations involve complex electronic rearrangements and ionic recombination within the active material. The molecular transformations induced by the applied input voltage signal result in the formation of multiple different analog levels within each memristor. These analog levels can correspond to different conductance states, effectively allowing each memristor to store analog values of data.
[0035] In one example, multiple different analog levels involve at least 16,520 conductance states. These conductance states can span a wide range, from as small as approximately 200 nanosiemens (nS) to as large as approximately 5.9 millisiemens (mS). Each of these discrete states corresponds to a unique analog value that can be stored in a memristor, providing a method for dense and granular analog data storage.
[0036] In one example, the system can operate in a first scenario (read operation) and a second scenario (write operation). In the first scenario, the input voltage signal can act as a read voltage pulse, while in the second scenario, the input voltage signal can act as a write voltage pulse. A read voltage pulse lower than a predefined threshold voltage that will cause a molecular transformation can be applied. In one example, the read voltage pulse is less than 600 millivolts (mV). When a read voltage pulse is applied, the system receives a current signal corresponding to the current conductance state (data) of the memristor being read as a response.
[0037] On the other hand, a write operation can be performed when the write voltage pulse exceeds a predefined threshold required for molecular transformation. In one example, the write voltage pulse can range from approximately 850 mV to 1.3 V. The exact voltage used can be fine-tuned based on the desired change in conductance. By controlling the pulse width between 40 nanoseconds and 10 microseconds (µs) and the voltage between 850 mV and 1.3 V, any conductance change from minimum to maximum conductance at 16,520 levels can be achieved using a single pulse (any input voltage signal).
[0038] In one example, a method for fabricating a cross-structure array structure is also described. The method includes fabricating a bottom electrode on a substrate using, for example, but not limited to, standard photolithography and etching techniques. A molecular thin film can then be spin-coated on top of the substrate. The top electrode can be deposited using a mask or via standard photolithography and etching processes. The top and bottom electrodes can be composed of metals selected from, but not limited to, gold (Au), copper (Cu), platinum (Pt), and cobalt (Co).
[0039] In one example, to fabricate a memristor, a thin film of [RuIIL2](BF4)2 (L=2,6-bis(phenylazo)pyridine) is sandwiched between top and bottom gold electrodes. The film thickness can vary from 5 to 100 nm, and the electrode thickness can be in the range of about 10 to 80 nm.
[0040] It is noteworthy that conducting in-situ Raman spectroscopy under different simulated conditions to measure the molecular electronic states in memristors can facilitate molecular transformations (electron rearrangement and ion rearrangement) within the memristor. In one example, the spectral data (correlated with in-situ Raman spectroscopy) show that the transition from the lowest to the highest conductivity level is controlled by supramolecular charge transfer, from a charge-disproportionate (CD) electronic state to a symmetric state. In the broken symmetry (CD) state, adjacent molecules have three and one electrons respectively in their antibonding orbitals, while in the valence symmetry state, all molecules contain two electrons in their antibonding orbitals. The different antiion positions between these states make ion shifts an integral part of the transition. Individual 80 ns pulses induce selective molecular and antiion rearrangements, promoting the transition between CD and symmetric states.
[0041] This method offers several technical advantages. For example, it functions as a dot product engine, where input values can be applied as voltages to the word lines of a cross-structure array, while matrix values are stored as conductance states of memristors. The computation results can then be read as currents from the bit lines of the array. This allows for efficient parallel processing of certain types of mathematical operations, particularly those common in neural networks and machine learning applications.
[0042] Furthermore, this method facilitates reading unidirectional current flow within a voltage range, simplifying the reading process and reducing potential errors. Additionally, the memristor is designed to operate without the need for an additional selector device, which is often required in other memristor systems to prevent unwanted current flow in large arrays. This simplification contributes to the overall system efficiency and scalability.
[0043] Furthermore, the physical dimensions of the components (in the simulation system) are carefully controlled to optimize performance. The active material layer typically has a thickness ranging from 5 nanometers (nm) to 100 nanometers (nm), while both the top and bottom electrodes have thicknesses ranging from 20 nanometers (nm) to 80 nanometers (nm). This precise dimensional control contributes to the overall performance and reliability of the memristor.
[0044] Furthermore, this method provides linear changes in conductance without requiring iterative verification. The processor can apply a series of voltage pulses to the memristor, resulting in a predictable and linear change in conductance. This eliminates the need for complex write verification schemes common in many other memristor systems, thereby improving the speed and efficiency of write operations. A key feature of the memristors discussed in this topic is their linear and symmetric weight updates under 900mV pulses. This means that if a specific change in conductance is required, it can be achieved in an open-loop manner by applying an appropriate number of pulses. In contrast, existing analog memristors exhibit nonlinear weight updates, requiring expensive write and verification loops to achieve the desired conductance change. The ability to perform direct, predictable weight updates eliminates the need for such iterative corrections, significantly reducing the computational overhead of the write process.
[0045] In one example, the system implements in-memory analog computation, where matrices are stored as cross-structured conductance states and vectors are applied as input voltage signals, thereby enabling single-step vector-matrix multiplication in the analog domain, significantly reducing computational complexity from N² steps to a single step. The system is optimized for high-bandwidth artificial intelligence (AI) and scientific computing workloads, providing an energy-efficient and scalable alternative to traditional digital memory and processing architectures.
[0046] Furthermore, this method utilizes the physical properties of memristors within a cross-structure architecture (known as a dot product engine). For example, in vector-matrix multiplication, the input value is applied as a voltage to the top electrode (word line) of the cross structure, while the matrix value is stored as memristor conductance. The output is generated in a single step, and the result is read as the current at the bottom electrode (bit line). Unlike digital platforms, where the computational complexity of an n×n matrix is O(n^2), this method leverages the physical properties of memristors within a cross-structure architecture (known as a dot product engine). 2 The dot product engine reduces this complexity to O(1), thus achieving efficient simulation processing.
[0047] The above-described embodiments will be further described herein with reference to the accompanying drawings. It should be noted that the specification and drawings relate to exemplary embodiments and should not be construed as limiting the scope of the subject matter. It should also be understood that various arrangements can be designed, and although not explicitly described or shown herein, these arrangements embody the principles of the subject matter. Furthermore, all statements and specific examples recounting the principles, aspects, and embodiments of the subject matter herein are intended to cover their equivalents.
[0048] Figure 1 A schematic diagram 100 of an analog computing system 102 for high-resolution analog computing is shown as an example according to the content of this subject matter. The analog computing system 102 (hereinafter referred to as system 102) includes a processor 104 and a cross-structure array 106.
[0049] System 102 can be implemented as a combination of hardware and programming (e.g., programmable instructions that implement various functionalities). In the examples described herein, this combination of hardware and programming can be implemented in several different ways. Programmable instructions can be stored on a non-transient machine-readable storage medium that can be directly coupled to system 102 or indirectly coupled (e.g., via a network device). In the examples, processor 104 can be, for example, a single processor or a combination of processors to execute such instructions. In this example, the non-transient machine-readable storage medium can store instructions that implement system 102 when executed by processing resources. In other examples, system 102 can be implemented as electronic circuitry.
[0050] In one example, the cross-structure array 106 includes a bottom electrode and a top electrode, with an active material intermediated between the bottom and top electrodes. In one example, the active material may be selected from the group of compounds, including but not limited to: ruthenium complexes of azo aromatic ligands, coordinated azo aromatic ligands, non-coordinated azo aromatic ligands, or combinations thereof. The bottom electrode, the top electrode, and the active material intermediated between the top and bottom electrodes are arranged in such a way that each junction acts as a memristor 108. The cross-structure array 106 includes a mesh-like structure in which horizontal and vertical wires intersect at the junctions, forming a connected network. Figure 2 Explain the details related to the cross-structure array 106.
[0051] Therefore, system 102 utilizes the unique properties of cross structure 106 to achieve high-resolution simulation computation, enabling the execution of complex operations within a single time step, and providing significant advantages over traditional digital computing systems for certain types of computations such as vector-matrix multiplication.
[0052] Figure 2A schematic diagram of a cross-structure array 200 for implementation in an analog computing system (not shown), as an example according to the content of this subject matter, is shown. The cross-structure array 200 may be similar to the cross-structure array 106.
[0053] like Figure 2 As depicted, the cross-structure array 200 can be a mesh-like structure with multiple word lines 202 and multiple bit lines 204 intersecting each other. As previously described, the cross-structure array 200 includes a bottom electrode and a top electrode, wherein active material is intermediate between the bottom electrode and the top electrode. The bottom electrode, the top electrode, and the active material intermediate between the top electrode and the bottom electrode are arranged in such a way (mesh-like structure) that each junction acts as a memristor 210. In one example, each junction defined by the intersection of word lines 202 and bit lines 204 in the mesh-like structure forms a memristor 210.
[0054] Memristor 210 allows for precise control and addressing of individual memristors. To facilitate operation of the cross-structure array 200, multiple word lines 202 are electrically coupled to the positive terminal, while multiple bit lines 204 are coupled to the negative terminal. This configuration allows a voltage difference to be applied to specific memristors within the array.
[0055] During operation, an input voltage signal 206 can be applied to the cross-structure array 200. This applied input voltage signal can cause molecular transformations within each memristor 210. These molecular transformations involve complex electronic rearrangements and ionic recombination within the active material.
[0056] The molecular transformation caused by the applied input voltage signal 206 can result in the formation of multiple distinct analog levels within each memristor 210. These analog levels correspond to different conductance states, effectively allowing each memristor 210 to store analog values of data. A conductance state can refer to the conductivity in the active material at a given time. In one example, the multiple distinct analog levels involve at least 16,520 discrete conductance states. These conductance states can span a wide range, from a minimum of approximately 200 nanosiemens (nS) to a maximum of approximately 5.9 millisiemens (mS). Each of these discrete states corresponds to a unique analog value that can be stored in the memristor, providing a method for dense and granular analog data storage.
[0057] In one example, system 102 can operate in a first scenario (read operation) and a second scenario (write operation). In the first scenario, input voltage signal 206 can act as a read voltage pulse, while in the second scenario, input voltage signal 206 can act as a write voltage pulse. A read voltage pulse lower than a first predefined threshold voltage that will induce a molecular transformation can be applied. In one example, the read voltage pulse is less than 600 millivolts (mV) (the first predefined threshold voltage). When a read voltage pulse is applied, system 102 receives a current signal corresponding to the current conductance state (data) of the memristor being read as a response.
[0058] On the other hand, a write operation can be performed when the write voltage pulse exceeds a second predefined threshold required for molecular transformation. In one example, the write voltage pulse can be in the range of approximately 850 mV to 1.3 volts (V) (the second predefined voltage). The exact voltage used can be fine-tuned based on the desired change in conductance. By controlling the pulse width between 40 ns and 10 µs and the voltage between 850 mV and 1.3 V, any conductance change from minimum to maximum conductance at 16,520 levels can be achieved using a single pulse (any input voltage signal).
[0059] In one example, System 102 implements in-memory analog computation, where matrices are stored as cross-structure conductance states and vectors are applied as input voltage signals, thereby enabling single-step vector-matrix multiplication in the analog domain, significantly reducing computational complexity from N² steps to a single step. System 102 is optimized for high-bandwidth AI and scientific computing workloads, providing a highly energy-efficient and scalable alternative to traditional digital memory and processing architectures.
[0060] Experimental data
[0061] Figure 3A A graph 302, illustrating, as an example according to the content of this subject matter, depicts the change in conductance in memristor 210 as a function of the number of applied write pulses, ranging from the lowest to the highest conductance value. Conductance scaling factor 10. -3 The indicator in the upper left corner shows the conductivity value on the y-axis multiplied by 10. -3Figure 302 begins with an initial conductance reference of 1 at the origin, where the initial pulse count of 0 marks the start of the enhancement process. The conductance increases sharply with the number of pulses. The inset shows magnified cross-sections (indicated by the green line) capturing conductance ranges spanning 3341.5 and 3342 mS for pulse counts of 9316 and 9318, respectively. The midpoint pulse count of 16500 and the endpoint pulse count of 33000 are marked along the x-axis to define the extent of the enhancement and suppression phases, respectively. From 0 to 16520, the enhancement process is captured, while for pulse counts between 16521 and 33040, suppression is triggered. Notably, the slopes of the enhancement and suppression processes are equal in magnitude but opposite in direction to each other.
[0062] In some cases, memristor 210 enables voltage-driven ion rearrangement, facilitating 16,520 different analog levels. This feature allows system 102 to achieve a wide range of conductance states, providing high flexibility and control during computation. In other aspects, memristor 210 uses 16,520 square pulses with an amplitude of 900 mV and a duration of 80 ns for positive weight updates and the same number of pulses with an amplitude of -750 mV and a duration of 65 ns for negative weight updates. This approach allows for precise control of conductance changes within the device, enabling linear and symmetrical updates or writes to analog levels individually within a single time step.
[0063] Figure 3B Figure 304 illustrates, according to the content of this topic, the conductance measurements of a memristor between two pairs of consecutive states, labeled Level-2, Level-3 and Level 10k2, Level 10k3, at various levels of data points and counts.
[0064] The average conductance G is depicted at level -10k2. 0 10k2 and its standard deviation σ 10k2 In the -10k3 level section, the average conductivity level is -10k3 G. 0 10k3 Its corresponding standard deviation σ 10k3 Shown together. ΔG 10k3 Indicates the electrical conductivity level G 0 10k3 and G 0 10k2 The difference lies in the fact that the memristor 210 allows for linear and symmetrical updates or writes to analog levels individually within a single time step. This feature enables precise control over conductance changes within the device, providing a high degree of flexibility and control during computation.
[0065] Similarly, at level -2, the average conductance is G. 02. The standard deviation is σ². Level -3 indicates that the average conductance is G. 0 3. The standard deviation is σ3, and ΔG3 represents G. 0 2 and G 0 The difference between 3. This provides a detailed view of the conductivity characteristics and statistical changes at this measurement level, highlighting the accuracy and resolution of the conductance state of memristor 210. Ensure that the difference between any two subsequent conductivity levels is at least six times the standard deviation at each level.
[0066] Figure 3C A graph 306, illustrating the probability distribution of conductance states in memristor 210, is shown as an example according to the content of this topic. Graph 306 shows the probability distribution of each conductance level around its average value. Different conductance levels are denoted as G. n And their respective average values are G. 0 n , where n varies from 1 to 16520.
[0067] In some cases, the color gradient from blue to yellow represents the progression of the conductance state from the initial state (n=1) to the final state (n=16520). This indicates that all different memristor states have a noise floor of less than 70 nS.
[0068] In other respects, this graph can be used to analyze the accuracy and repeatability of the conductance states of a memristor. The distribution of conductance states can provide insights into the performance of the device under different operating conditions, such as the number or amplitude of applied pulses. A detailed view of the conductance characteristics and statistical variations at different measurement levels allows the memristor 210 to achieve a wide range of conductance states, providing a high degree of flexibility and control during the calculation process.
[0069] Figure 3D A graph 308 shows the statistical distribution of conductance levels in memristor 210, as an example based on the content of this topic. The x-axis represents the conductance at each level, scaled by 10^-3 Siemens (S), while the y-axis shows two distinct quantities: the conductance change ΔG between two consecutive levels. n (S) and the standard deviation σ of each level n (S), both scaled to 10^-7 Siemens (S). The green area represents the conductance difference ΔG between any two consecutive levels in the range of 1 to 16520. n (S). The red area below represents the standard deviation σ for each level. n (S). This shows that the difference in conductivity between any two levels is at least 6 times its standard deviation.
[0070] The chart demonstrates the accuracy with which each conductivity level can be distinguished within the memristor 210, highlighting the device's high resolution and analog accuracy.
[0071] In other respects, System 102 can achieve a signal-to-noise ratio greater than 73 dB while consuming significantly less energy than traditional digital computers. This energy efficiency (combined with the high computational resolution provided by the system) addresses some of the challenges associated with the use of analog computing elements in mainstream computing, such as accuracy, noise, and device variability.
[0072] Figure 3E A graph 310, illustrating the relationship between write conductance levels and corresponding read conductance values in memristor 210, is shown as an example according to the content of this topic. The x-axis represents the write conductance level, while the y-axis represents the read conductance value; both axes are scaled to 10^-3 Siemens (S). The graph is characterized by two distinct curves: one for readings at -0.5 V and the other for readings at +0.5 V.
[0073] The negative reading voltage, shown in red, demonstrates a low read conductance across all write conductance levels. This feature can be particularly useful in blocking negative currents through the device, enabling selector-less crossover structures.
[0074] In contrast, the +0.5V reading curve, depicted in green, shows a linear increase in the read conductance value corresponding to the write conductance level. This indicates that the positive voltage reading accurately reflects the write conductance state.
[0075] The illustration plots conductance on a logarithmic scale, further emphasizing the linearity of the positive readout voltage and the blocking effect in the negative voltage.
[0076] In other respects, the memristor 210 can be read using pulses of less than 600mV of any duration without disturbing its resistance state. This feature allows for efficient and lossless reading of the memristor's conductance state, potentially enhancing the overall performance and reliability of computing systems.
[0077] Figure 4A A graph 402, illustrating the endurance of memristor 210 across read and write cycles, is shown as an example based on the content of this topic. The x-axis represents the endurance cycles, marked by specific cycle counts, including 10^3 for thousands of cycles, 10^5 for hundreds of thousands of cycles, 10^7 for tens of millions of cycles, and 10^9 for billions of cycles. The y-axis indicates the conductance of the memristor, scaled in 10^-3 Siemens (S).
[0078] Figure 402 shows a series of peaks and troughs corresponding to the conductance changes during the enhancement and suppression phases of memristor operation. The enhancement phase refers to the process of increased conductance of the memristor, while the suppression phase refers to the process of decreased conductance. In some cases, these phases can be controlled by applying a series of electrical pulses to the memristor, where the number and amplitude of the pulses determine the conductance level.
[0079] The pattern of conductance variation repeats consistently across the durability cycle range, indicating that the device can maintain a stable conductance level across multiple cycles, from a single cycle to billions of cycles. This feature can be particularly beneficial in applications requiring long-term stability and repeatability of conductance states, such as in non-volatile memory applications or in systems where precise control of analog states is critical.
[0080] In other respects, the memristor 210 exhibits high durability, capable of withstanding a large number of cycles without performance degradation. This high durability can help improve the overall reliability and lifespan of computing systems, potentially enhancing their suitability for a wide range of computing applications.
[0081] Figure 4B A graph 404, illustrating the electrical conductivity level of memristor 210 over time, is shown as an example according to the content of this topic. The graph is divided into several horizontal bands, each representing the device at 85°C. 0 Different conductance levels when subjected to a 500mV DC voltage. The darkest blue conductance band at the bottom represents the 8th conductance level, while the lowest conductance level is counted as level -1, and the conductance increases with the count. The yellow conductance level represents the 6005th conductance level. In between, there are additional bands indicating intermediate conductance states: dark blue level 9, blue level 10, light blue level 1003, green level 1004, light green level 1005, orange level 6003, and dark yellow level 6004. Each level is labeled with its own conductance level, and the data point within each band shows the readout noise of the conductance at that level over a time period represented on the x-axis. Notably, no drift was observed in these arbitrary conductance levels.
[0082] In other respects, the data points within each band can show the stability and consistency of the conductance at that level over a time period represented on the x-axis. This provides a visual representation of the memristor 210's ability to maintain a stable conductance level over time, potentially highlighting the device's reliability and repeatability in analog state representation. A detailed view of the conductance characteristics over time allows the memristor 210 to achieve a wide range of conductance states, providing a high degree of flexibility and control during calculations.
[0083] Figure 4CA graph 406, illustrating, as an example according to the content of this topic, depicts programming the electrical conductance level in memristor 210 using two different methods. The graph shows the conductance difference on the y-axis, ranging from 0 (~200 ns) to 6 ms, and the programming method on the x-axis. A pulse train characterized by a pulse train duration of 80 ns and an amplitude of +900 mV can be used to write different conductance states, where the number of pulse counts determines the written state.
[0084] In some cases, the diagram also illustrates single-pulse programming, where applying a single pulse of the same amplitude Vp1 or Vp2 directly achieves the desired conductance change level, skipping intermediate steps. This indicates a more efficient programming method because it allows direct access to any conductance level within a single time step, thereby simplifying the weight update process and enhancing the computational efficiency of the memristor 210.
[0085] The maximum conductivity difference shown in the graph is 6 mS, demonstrating the range of conductivity levels achievable through these programming methods. This wide range of conductivity levels allows the memristor 210 to represent a variety of different analog states, providing a high degree of flexibility and control during computation. This feature can be particularly beneficial in applications requiring high-resolution analog computation, such as signal processing, neural network training, and natural language processing.
[0086] Figure 4D A three-dimensional graph 408, illustrating the relationship between pulse voltage amplitude, pulse width, and resulting conductance difference in memristor 210, is shown as an example based on the content of this topic. Graph 408 shows the pulse voltage amplitude (0.6–1.5V) on the x-axis and the range on the y-axis from 80ns to 3.5 microseconds (x10⁻¹⁰). -6 The pulse width (in seconds) and the logarithm of the conductance difference caused by the pulse applied on the z-axis.
[0087] The color gradient in the graph represents the conductance difference amplitude, with a scale range from 10^-9 S to 5 x 10^-3 S. This color gradient provides a visual representation of the conductance difference amplitude, allowing for easy identification of conductance levels associated with different pulse widths and amplitudes.
[0088] The inset on the left of the chart shows the root mean square error (RMSE) associated with the conductance difference. This inset provides a detailed view of the accuracy of the memristor's conductance variation in response to different pulse widths and voltage amplitudes, highlighting the precision and repeatability of the memristor's conductance state.
[0089] In other respects, this graph can be used to analyze the performance of memristor 210 under different operating conditions, such as varying pulse width and voltage amplitude. This detailed view of the conductance characteristics as a function of pulse width and voltage amplitude allows memristor 210 to achieve a wide range of conductance states, providing a high degree of flexibility and control during the calculation process.
[0090] Figure 4E A graph 410, illustrating the relationship between the transition time and the reciprocal of thermal energy between different subsequent conduction level pairs in memristor 210, is shown as an example according to the content of this topic. The x-axis represents the Boltzmann constant multiplied by the reciprocal of temperature (1 / K). B The y-axis represents the natural logarithm of the reciprocal of the transition time (ln(1 / transition time)), in the reciprocal of the second (s^-1). The graph is characterized by a linearly fitted line indicating the activation energies (Ea) of the transitions between seven pairs of conductance levels: level -1 to 2, level -501 to 502, level -2001 to 2002, level -5001 to 5002, level -10001 to 10002, level -15001 to 15002, and level -16501 to 16502. The constant transition rates and activation energies of all these different transitions suggest that this explains the observed linear zero-order transition.
[0091] Figure 4F A graphical representation 412 of the conductance modes in memristor 210, as exemplified by the content of this topic, is shown. The figure displays two identical images side-by-side, representing the real and imaginary DFT matrices. Half (64×32) of the crossover structure captures the [real part (DFT matrix)]. 64×32 The other half captures the [imaginary part (DFT matrix)]. 64×32 Due to the structural symmetry in the DFT, half of the columns are sufficient to capture all frequency components at the output.
[0092] The color gradient across the loop corresponds to increasing conductance values, with darker colors representing lower conductance and lighter colors representing higher conductance. This color coding representation can provide a visual indication of the conductance state of the memristor 210, potentially aiding in the interpretation and analysis of the device's conductance properties.
[0093] The minimum conductivity value is 0, while the maximum conductivity value of 5.9 mS is marked in the lower right corner of the graph. This value represents the maximum conductivity state achieved by memristor 210, corresponding to the brightest color in the color gradient. The maximum conductivity value of 5.9 mS can be associated with a specific conductivity level, which can be the maximum conductivity level that memristor 210 can achieve under given operating conditions.
[0094] The bottom of the figure provides a conductivity scale in millisiemens (mS) to quantify the conductivity level depicted in the pattern. The conductivity scale ranges from 0 mS, corresponding to the initial pulse count of 0, to 5.9 mS, corresponding to the maximum conductivity value of 5.9.
[0095] Figure 5A A graph 502, illustrating an example based on the content of this topic, compares the Fourier transform (FT) coefficients obtained from a memristor-based hardware implementation (memristor 210) with the calculated FT. The graph above shows the input waveform, where the y-axis represents the voltage (V) and the x-axis represents the word line number, ranging from 1 to 64. This waveform can represent the input signal processed by memristor 210. The waveform can be a time-domain signal, such as an acoustic wave or an electrical signal, converted to a frequency-domain representation through a Fourier transform process.
[0096] The bottom chart shows the corresponding FT from the hardware and the calculated FT, with the column current (mA) on the y-axis and the bit line number on the x-axis, ranging from 1 to 64. Bit line labels 16, 32, 48, and 64 are marked to correspond to specific points along the x-axis.
[0097] Two overlaid graphs illustrate the accuracy of the memristor hardware in replicating the Fourier transform of the input signal generated by the software. This feature can be particularly beneficial in applications requiring high-resolution signal processing, such as neural network training, natural language processing, and other core computational tasks. The memristor 210's ability to accurately perform the Fourier transform process can contribute to its high computational resolution, potentially enhancing the overall performance and efficiency of the computing system.
[0098] In other respects, the memristor 210 is capable of performing core computational tasks such as signal processing, neural network training, and natural language processing with high computational resolution. This capability is facilitated by the device's ability to implement various analog levels and accurately reproduce the Fourier transform of the input signal. The high computational resolution of the memristor 210 can help make it suitable for a wide range of computing applications, potentially driving progress in the field of analog computing.
[0099] Figure 5B A graph 504, illustrating an example based on the content of this topic, compares the performance of matrix-matrix multiplication across different matrix types using a memristor-based computing system (analog computing system 102). The x-axis categorizes matrices into various types, including triangular matrices, Hermitian matrices, asymmetric matrices, symmetric matrices, random matrices, and double random matrices. Each category can represent a specific type of matrix that can be processed by memristor 210, where each type has its own characteristics and computational requirements.
[0100] The y-axis of the graph is divided into two metrics. The upper half measures the root mean square error (RMSE) in nanoamperes (nA), while the lower half measures the signal-to-noise ratio (SNR) in decibels (dB). These metrics provide a quantitative assessment of the performance of the memristor 210 in handling matrix-matrix multiplication tasks. RMSE indicates the accuracy of the calculation, with lower values representing higher accuracy. SNR measures the quality of the signal output by the memristor 210, with higher values representing better signal quality.
[0101] Above the x-axis are violin plots showing the distribution of RMSE values for each matrix type. The real (Re) and imaginary (Im) components of the complex matrix are represented. These plots provide a visual representation of the distribution of RMSE values, highlighting the variability in accuracy across different matrix types. The real and imaginary components represent two parts of a complex number that can be used in calculations involving complex matrices.
[0102] The colored circles depicted below the x-axis indicate the SNR for each matrix type. These circles are scaled from 12 bits to 13.3 bits of precision. This scale represents the precision of the signal output by memristor 210, with higher bit values indicating higher precision. The colored circles provide a visual representation of the SNR for each matrix type, highlighting the noise performance of memristor 210 when processing various types of matrices.
[0103] Memristor-based computing systems are capable of performing matrix-matrix multiplication tasks with high accuracy and low noise across a wide range of matrix types. This provides the best SNR to date among any neuromorphic crossover architecture. This capability can enhance the versatility and applicability of systems in a variety of computing applications, potentially driving advancements in the field of analog computing.
[0104] Figure 5C A graph 506 illustrates, as an example of the content of this subject matter, the conductance characteristics of a memristor 210 as a function of the number of applied electrical pulses. The x-axis of this graph can represent the number of pulses, in increments of 10. 4 Scaling, while the y-axis can represent conductance in Siemens (S) units, according to 10 -3 Zoom. The graph can be characterized by a series of lines where conductivity increases with the number of pulses, peaking at different conductivity levels before returning to the baseline, forming a triangular pattern.
[0105] The color gradient from blue to yellow in the graph indicates the progression of conductivity level as the number of pulses increases. Blue can represent a lower conductivity level, while yellow can represent a higher conductivity level. This color coding representation can provide a visual indication of the conductivity state of the memristor 210, potentially aiding in the interpretation and analysis of the device's conductivity properties.
[0106] In other respects, this graph can demonstrate the memristor's ability to achieve and maintain different conductance states over a wide range of pulse counts. This feature highlights the accuracy and repeatability of the device in analog state programming. By varying the number of pulses applied to the device, it is possible to control the conductance level, thereby enabling the device to represent a variety of different analog states with high precision. This detailed view of the conductance characteristics as a function of pulse count allows the memristor 210 to achieve a wide range of different analog levels, providing a high degree of flexibility and control during calculations.
[0107] Figure 6 Examples based on the content of this topic are shown in two graphs representing the conductance characteristics of memristor 210 as a function of the number of applied electrical pulses for 9000 weight updates from 900 cross points. Figure 6 (a) and Figure 6 (b)). These two charts share the same x-axis, representing the x-axis by 10. 4 The number of pulses scaled. For each graph, the conductance characteristic is represented differently on the y-axis. The left graph shows conductance in Siemens (S) units on the y-axis, while the right graph shows the logarithm of conductance on the y-axis. 10 (Conductivity)
[0108] In some cases, Figure 6 The left chart (a) shows the linear and symmetrical modes of conductance change. This mode can be observed as the conductance increases during the enhancement phase and then decreases during the suppression phase. The enhancement phase refers to the process of increasing conductance of the memristor, while the suppression phase refers to the process of decreasing conductance. These phases can be controlled by applying a series of electrical pulses to the memristor, where the number and amplitude of the pulses determine the conductance level.
[0109] In other respects, Figure 6 The right-hand chart (b) in the diagram presents a similar pattern on a logarithmic scale. This scale can emphasize the precision and dynamic range of the memristor's response to electrical impulses. The logarithmic scale provides a more detailed view of conductance variations, especially for small variations that may not be easily discernible on a linear scale. This detailed view enhances the understanding of the memristor's conductance properties and its response to electrical impulses.
[0110] Furthermore, the individual data points of the 9000 weighted update characteristics obtained from the 900 cross points are superimposed on the curve to demonstrate the consistency of the memristor's behavior. These data points can represent individual measurements of the memristor's conductance at different time points or under different operating conditions. The consistency of these data points can indicate the reliability and repeatability of the memristor's conductance state, potentially enhancing the performance and reliability of the memristor 210 in analog computing applications.
[0111] Figure 7 This presents a set of four heatmaps (700) as examples based on the content of this topic. Figure 7 (a)-(d) in the table compares the DFT matrix writing accuracy immediately after writing and after a seven-month period. In some aspects, the computational system may include mechanisms for writing the Discrete Fourier Transform (DFT) matrix with high accuracy, such as... Figure 7 The charts are depicted in the diagram. The x-axis represents the column index, and the y-axis represents the row index, both ranging from 0 to 60. The color scale on the right side of each chart represents the electrical conductivity level, scaled at 10^-3 Siemens (S), where blue represents lower conductivity and yellow represents higher conductivity.
[0112] In some cases, marked as Figure 7 The first graph in (a) “Real Part - Write” shows the real part of the DFT matrix after an immediate write. This graph provides a visual representation of the conductance level in the real part of the DFT matrix, where color intensity indicates the magnitude of the conductance. (Labeled as...) Figure 7 The second chart in (b) "Imaginary Part - Write" also shows the imaginary part of the DFT matrix. Similar to the first chart, this chart can provide a visual representation of the electrical conductance levels in the imaginary part of the DFT matrix.
[0113] In other respects, marked as Figure 7 The third chart in (c) "Real Part - 7 Months Later" shows the real part of the DFT matrix after 7 months. This chart can demonstrate the stability and retention of the electrical conductivity level in the real part of the DFT matrix over the extended time period. (Labeled as...) Figure 7 The fourth chart (d) "Imaginary Part - 7 Months Later" depicts the imaginary part of the DFT matrix after the same time period. Similar to the third chart, this chart can show the stability and retention of the electrical conductance level in the imaginary part of the DFT matrix over time.
[0114] Consistency between the "write" and "7-month later" graphs for both the real and imaginary parts demonstrates the stability and persistence of the DFT matrix over time. This feature can enhance the reliability and repeatability of computing systems, potentially driving advancements in the field of analog computing.
[0115] Figure 8 An example is shown, according to the content of this topic, illustrating a graphical representation 800 of the original waveform reconstructed using an analog computing system 102 via an inverse discrete Fourier transform (IDFT) process. The diagram shows a graph where the x-axis represents row numbers, which can indicate discrete positions or time intervals. The y-axis on the left side of the graph can show the voltage (V) of the original signal, while the y-axis on the right side can show the strength of the signal reconstructed via vector-matrix multiplication (VMM).
[0116] This chart compares two curves: the blue curve, which represents the original signal, and the orange curve, which represents the reconstructed signal. These two curves are very close, demonstrating the accuracy of the IDFT process in reconstructing the original waveform from its frequency components. The close alignment of the two curves across all row numbers indicates the effectiveness of a memristor-based computing system in performing the IDFT operation.
[0117] In some cases, memristor-based computing systems can be capable of performing the IDFT accurately on a wide range of waveforms. This allows the system to handle waveforms of varying complexity and duration, potentially enhancing its versatility and applicability across diverse computing applications.
[0118] In other respects, memristor-based computing systems can utilize different methods or algorithms for waveform reconstruction, not just DFT or IDFT. These methods can include, for example, cosine transform, wavelet transform, or other suitable signal processing techniques. The choice of reconstruction method can depend on the specific requirements of the computation process, such as desired computational speed, waveform reconstruction accuracy, or system energy efficiency.
[0119] Figure 9 A series of 900 sub-images are shown, each consisting of a pair of graphs. In some aspects, the computational system may be capable of performing Discrete Fourier Transform (DFT) on different representative waveforms, such as... Figure 9 As depicted in the diagram. The top graph in each pair of graphs can show the time-domain signal waveform, where the signal strength is represented on the y-axis and the row number is represented on the x-axis. This graph can provide a visual representation of the input signal in the time domain, potentially aiding in the interpretation and analysis of the signal's temporal characteristics.
[0120] The bottom chart in each pair of charts displays the corresponding DFT result, with the real and imaginary components of the DFT represented on the y-axis and column numbers on the x-axis. The real component chart is labeled "Real Part," and the imaginary component chart is labeled "Imaginary Part," providing a clear distinction between the two components of the DFT. DFT results computed by a cross-structured vector matrix multiplication (VMM) system are depicted in orange, while DFT results computed by a software-based FFT function are depicted in blue. This color-coded representation provides a visual indication of the accuracy of the cross-structured VMM system in capturing the frequency domain representation of the input signal, potentially aiding in the interpretation and analysis of the signal's spectral characteristics.
[0121] In certain situations, cross-structure VMM systems can accurately compute the DFT of various time-domain signals, highlighting their ability to perform complex signal processing tasks. The accuracy of cross-structure VMM systems is illustrated by the close alignment of the orange and blue lines in the bottom chart, indicating a high correlation between the DFT results computed by the cross-structure VMM system and software-based FFT functions. This feature enhances the reliability and repeatability of cross-structure VMM systems in performing DFT computations, potentially driving advancements in analog computing.
[0122] Various implementations have been described. However, it should be understood that various modifications can be made without departing from the spirit and scope of this disclosure. Therefore, other embodiments are within the scope of the appended claims.
Claims
1. A simulation computing system, comprising: A cross-structure array, comprising a bottom electrode, a top electrode, and an active material between the top electrode and the bottom electrode, such that each junction of the cross-structure array acts as a memristor, wherein the active material is selected from the group consisting of ruthenium complexes of azo aromatic ligands, coordinated azo aromatic ligands, non-coordinated azo aromatic ligands, and combinations thereof. A processor, operatively coupled to the cross-structure array, is used for: An input voltage signal is applied to the cross-structure array to induce molecular transformations in each memristor. These molecular transformations indicate electronic rearrangements and ionic recombination of the active material, forming multiple different analog levels, including multiple conductance states, to store analog values of data.
2. The analog computing system of claim 1, wherein the cross-structure array comprises multiple word lines and multiple bit lines, wherein each memristor formed at each junction of the cross-structure array is located at the intersection of the word lines and the bit lines.
3. The analog computing system according to claim 2, wherein the plurality of word lines are electrically coupled to the positive terminal and the plurality of bit lines are electrically coupled to the negative terminal.
4. The analog computing system of claim 1, wherein the plurality of different analog levels comprises at least 16,520 conductance states, and wherein each memristor is configured to be any one of the 16,520 conductance states.
5. The simulation computing system of claim 4, wherein the conductance states span a range from a minimum conductance of about 200 nanosiemens nS to a maximum conductance of about 5.9 millisiemens mS, wherein each conductance state corresponds to a unique analog value that can be stored in the memristor.
6. The analog computing system according to claim 1, wherein the input voltage signal includes a read voltage pulse and a write voltage pulse.
7. The simulation computing system according to claim 6, wherein the processor is used for: The readout voltage pulse is applied to the cross-structure array, the readout voltage pulse having a value lower than the threshold voltage required for the molecular transformation; and Receive a current signal in response to the read voltage pulse from the cross-structure array, wherein the current signal corresponds to the conductance state of the memristor.
8. The analog computing system according to claim 7, wherein the read voltage pulse is less than 600 millivolts mV.
9. The simulation computing system according to claim 6, wherein the processor is used for: The write voltage pulse is applied to the cross-structure array to update the conductance state of the memristor, the write voltage pulse having a value higher than the threshold voltage required for the molecular transformation.
10. The analog computing system of claim 9, wherein the write voltage pulse is in the range of about 850 mV to 1.3 V.
11. The analog computing system of claim 9, wherein the processor controls the pulse width of the write voltage pulse to be between approximately 40 nanoseconds (ns) and 10 microseconds (µs) to achieve a predefined conductance change.
12. The simulation computing system of claim 1, wherein one or more matrices are stored as cross-structure conductance states to enable simulation computing in memory.
13. The analog computing system of claim 1, wherein one or more vectors are applied to the cross-structure array as the input voltage signal.
14. The simulation computing system according to claim 1, wherein the simulation computing system is capable of implementing single-step vector multiplication in the simulation domain, thereby reducing the computational complexity from NT. 2 The number of steps is reduced to a single step.
15. The analog computing system of claim 1, wherein the analog computing system is optimized to provide high-bandwidth artificial intelligence (AI) and scientific computing workloads, and wherein the optimization is for providing a highly energy-efficient and scalable architecture.
