Operating system and method
By designing a combination of multi-memory systems and floating gate transistors, using search coding and feature coding, the problems of high power consumption and low memory density in large-scale data and artificial intelligence processing are solved, and efficient search and comparison operations are achieved.
Patent Information
- Application Number
- CN202311065916.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-05-10
- Filing Date
- 2023-08-23
- Publication Date
- 2024-11-12
AI Technical Summary
The existing three-state content addressable memory (TCAM) technology has problems with high power consumption and low memory density when performing large-scale data and artificial intelligence processing, and the nonvolatile execution of TCAM technology requires paired memory cells. This leads to limited search and data storage efficiency.
A multi-memory system is designed to use selectively coupled sense amplifiers, matching lines and source lines, combined with series-connected floating gate transistors to achieve high-performance perfect and approximate matching operations through search encoding and feature encoding. Supports parallel operations for increased efficiency.
It realizes low latency, high resolution, and high content density search functions, suitable for large-scale data and artificial intelligence processing, improves search, comparison and sorting performance, reduces power consumption, and improves memory density.
Smart Images

Figure CN118942512A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to a three-dimensional in-memory searching (IMS), which may be performed using a three-dimensional NAND memory and may be applicable to big data and / or artificial intelligence (AI) processing. Background Art
[0002] The growth of big data and artificial intelligence (e.g., for artificial intelligence hardware accelerators) is increasing the importance of searching, comparing, and / or sorting data. Conventional systems use ternary content addressable memory (TCAM) technology to perform concurrent searching.
[0003] Conventional TCAM technology may be implemented using static randomly accessible memory (SRAM) technology, and thus has relatively low memory density (eg, 16 transistors to form a single TCAM cell) and relatively high power usage.
[0004] Recently proposed TCAM technology can be implemented using non-volatile memory technology, such as two transistor two resistor (2T2R) technology and two ferroelectric field-effect transistor (2FeFET) technology. However, non-volatile implementation of TCAM technology requires paired memories for a single TCAM cell (or paired memory cells to perform a search bit or a data bit), thus limiting the efficiency of search and data storage.
[0005] Example techniques for using search functionality in NAND flash-based memories that enable full match and approximate match operations are described in In-Memory-Searching Architecture Based on 3D-NAND Technology with Ultra-high Parallelism, 2020 IEDM by PHTseng et al. and In-Memory Approximate Computing Architecture Based on 3D-NAND Flash Memories, 2022 VLSI by PHTseng et al.
[0006] Therefore, there is an urgent need for technologies that can perform high-performance searching, comparing and / or sorting, for example in connection with large data and / or artificial intelligence processing. Summary of the invention
[0007] The system of one or more computers is configurable to perform specific operations or actions because software, firmware, hardware, or a combination thereof is installed on the system, which causes or enables the system to perform these actions during operation. The storage of one or more computers is configurable to perform specific operations or actions because it contains instructions that, when executed by a data processing device, cause the data processing device to perform these actions.
[0008] The first aspect of the present disclosure includes a computing system having a plurality of memories. Each memory selectively includes a respective match line coupled to a respective sense amplifier, a respective source line, and a respective plurality of memory strings. Each memory string is coupled to a respective match line and a respective source line. Each memory string selectively includes a respective plurality of series-connected floating-gate transistors, each of which has and responds to a respective configuration state and a respective control input. The computing system also includes a plurality of word lines. Each word line is coupled to a respective one of the respective control inputs of each series-connected floating-gate transistor. The computing system also includes a search encoder, which is capable of receiving a search and driving the word line according to a search code. Wherein, each memory string is enabled to couple a respective match line and a source line, and a respective memory string is coupled to a respective match line and a respective source line by responding to the configuration state of the respective memory string to fully match the control inputs of the series-connected floating-gate transistors of the respective memory string. Furthermore, each sense amplifier is enabled to generate a respective indication of whether the match line to which it is coupled is coupled to the source line via the found matching impedance. Furthermore, the respective memory string is coupled to the respective match line and the respective source line by an unfound matching impedance that does not completely match the control inputs of the series-connected floating-gate transistors of the respective memory string in response to the configuration state of the respective memory string. Furthermore, each sense amplifier is enabled to generate a respective indication of whether the match line to which it is coupled is coupled to the source line via the unfound matching impedance.
[0009] Other embodiments of the first aspect include corresponding computer systems, apparatuses, and computer storage recorded on one or more computer storage devices, each configured to perform the actions of the computing method.
[0010] Some computing systems optionally include one or more of the following features. In the computing system, word lines are operated in pairs according to search codes, and configuration states are managed in pairs according to feature codes. Each configuration state corresponds to a respective threshold voltage. Each respective control input is coupled to a respective gate of one or more series-connected floating gate transistors. The threshold voltage can be set by one or more storage operations. Each memory string optionally includes a respective memory string enable device that selectively enables or disables the respective memory string independently of the word line in response to an enable control input of the respective memory string. The described techniques can optionally be performed by hardware, method or program, or computer software on a computer accessible medium.
[0011] A second aspect of the present disclosure includes an operation method, the operation method including receiving one or more search parameters; encoding the search parameters according to a specific one of a plurality of search codes to generate a plurality of pairs of word line voltages, the specific search code searching for a match matching a feature stored in a pair of floating gate transistors according to a specific one of the feature codes, and the specific search code and the specific feature code can represent 1, 0, and 1 or 0; driving word line pairs of a plurality of serially connected instances of the paired floating gate transistors according to the generated word line voltages, the serially connected instances being organized as a plurality of memory strings operable in parallel to affect a plurality of shared match lines, each shared match line coupled to two or more memory strings, and each word line pair coupled to at least two instances of the paired floating gate transistors; sensing voltage changes on the shared match lines to determine which of the shared match lines has a voltage change greater than a predetermined sensing threshold; and providing the determined result as an indication of a relative match between the search parameters and the stored features. Other embodiments of the second aspect include corresponding computer systems, apparatuses, and computer storage recorded on one or more computer storage devices, each configured to perform the actions of the computing method.
[0012] Some methods of operation optionally include one or more of the following features. In the method of operation, at least one of the search encodings or at least one of the feature encodings is in accordance with at least one of an expansion encoding and a one-hot encoding. At least one of the search encodings or at least one of the feature encodings is in accordance with a reverse expansion encoding. At least one of the stored features corresponds to a range stored using one-hot encoding. At least one of the paired word line voltages is in accordance with a range encoded as a one-hot encoding. At least one of the search encodings or at least one of the feature encodings is in accordance with a low effect encoding. The described techniques may optionally be performed by hardware, method or program, or computer software on a computer accessible medium.
[0013] A third aspect of the disclosure includes a computing system, the computing system including an interface circuit for enabling a host agent to provide a search and receive one or more results of the search. The computing system includes a block of a plurality of memory strings, each block including a plurality of individual memory strings, each memory string receiving a same set of word lines having voltages determinable in response to the search. The computing system includes a plurality of sense amplifiers, each sense amplifier coupled to a corresponding one of the memory strings from each block. The computing system includes a priority encoder circuit enabled to receive a match indication from the sense amplifiers. The sense amplifiers are enabled to determine a match between characteristic information stored about the memory strings coupled to the sense amplifiers and word lines received by the memory strings coupled to the sense amplifiers, determine a mismatch, and determine at least a match. Some computing systems include corresponding computer systems, apparatus, and computer programs recorded on one or more computer storage devices, each configured to perform the actions of a method. And, the priority encoder circuit is enabled to indicate at least a portion of the results as a highest priority match of any matches identified by the sense amplifiers based on a determination of the sense amplifiers.
[0014] The computing system optionally includes one or more of the following features. The computing system also includes a host agent. Based on the determination made by the sense amplifier, the priority encoder circuit is enabled to indicate at least a portion of the results as a second highest priority match of any match identified by the sense amplifier. The feature information is stored in the memory strings by configuring the threshold voltages of multiple floating gate transistors of the memory strings. Two or more of the blocks are operable in parallel. A feature encoding is used to encode at least some portions of the feature information, and a search encoding is used to encode at least a portion of the search. The feature encoding is at least one of an expansion encoding, an inverse expansion encoding, a one-hot encoding, and a low-effect encoding. The computing system is part of a system-on-a-chip (SOC). The described techniques may optionally be implemented by hardware, methods or programs, or computer software on a computer-accessible medium. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1A , Figure 1B and Figure 1C Collectively, various aspects of an exemplary three-dimensional (search) engine are shown.
[0016] Figure 1D An exemplary two-dimensional (search) engine and its use is shown.
[0017] Figure 1E An example of the use of expansion coding of a two-dimensional (search) engine for a three-dimensional (search) engine is shown.
[0018] Figure 1F An example of the use of one-hot encoding for a two-dimensional (search) engine of a three-dimensional (search) engine is shown.
[0019] Figure 2 An exemplary expanded encoding for a three-dimensional (search) engine is shown.
[0020] Figure 3 An exemplary one-hot encoding for a three-dimensional (search) engine is shown.
[0021] Figure 4A An exemplary expanded encoding and a corresponding exemplary low-effect encoding are shown.
[0022] Figure 4B An example unfolded encoding with a Hamming distance compared to input 1 is shown.
[0023] Figure 4CExemplary threshold voltages and word line voltages are shown that enable comparison of 1's, 0's, and don't cares.
[0024] Figure 4D Icons related to overdriving the high and low gates to high and low cell current capabilities are shown.
[0025] Figure 4E The floating gate transistor impedance is shown as a function of the search information versus the stored data.
[0026] Figure 5A An exemplary inverse expansion encoding of a search input for a three-dimensional (search) engine is shown.
[0027] Figure 5B An exemplary Hamming distance as a function of stored data values and search input values for an inverse unfolded code is graphically illustrated.
[0028] Figure 6 An exemplary Hamming distance as a function of stored data values and search input values for an unfolded code is graphically illustrated.
[0029] Fig. 7A An example of one-hot encoding with storage range is shown.
[0030] Figure 7B An example of one-hot encoding on the input range is shown.
[0031] Figure 8 An exemplary operation flow of a three-dimensional (search) engine is shown.
[0032] Fig. 9 A three-dimensional engine apparatus including an exemplary three-dimensional engine device is shown.
[0033] Fig.10 An exemplary hardware system having a three-dimensional (search) engine is shown.
[0034] Description of Reference Numerals
[0035] 100, 119, 129, 180, 190: word line
[0036] 130, 131, 138, 139, 430: Memory string
[0037] 134, 135: IMS data bits
[0038] 136D, 136E: Series current
[0039] 140, 148, 149: bit lines
[0040] 144: Sensor output
[0041] 145: Shared source line
[0042] 150, 158, 159: 2D (search) engine
[0043] 169: Ground selection line
[0044] 170, 179: string selection line
[0045] 251, 351, 451, 461, 761: value
[0046] 252, 452, 462, 552: Expand code
[0047] 352, 762: One-hot encoding
[0048] 404: Search Input
[0049] 450, 750: Input
[0050] 453, 553, 599, 699: Hamming distance
[0051] 463: Low effect coding
[0052] 464: Storage Features
[0053] 470: Memory status 0
[0054] 471: Memory status 1
[0055] 472: Value V1
[0056] 473: Value V2
[0057] 474: Value V3
[0058] 480: Search 0
[0059] 481: Search 1
[0060] 482: Wildcard
[0061] 483: High-side gate overvoltage drive
[0062] 484: Low-side gate overvoltage drive
[0063] 485: High cell current
[0064] 486, 487: Low cell current
[0065] 489: Enter search information
[0066] 490: Storage 0
[0067] 491: Storage 1
[0068] 492: Random Item
[0069] 493: Storage invalid
[0070] 499, 590, 690: storage data value
[0071] 506, 606: Search input value
[0072] 550: Search value example
[0073] 551: Storage value
[0074] 555: Search value
[0075] 554: Reverse expansion encoding
[0076] 751: Value range
[0077] 752: One-hot encoding of storage range
[0078] 753, 763: Match?
[0079] 760: One-hot encoding of input range examples
[0080] 800: Engine Process
[0081] 801, 802, 803, 804, 805, 806, 807: Steps
[0082] 940: 3D engine equipment
[0083] 941: Device Interface
[0084] 942: Host / Bus Coupling
[0085] 950: 3D engine device
[0086] 951: Controller
[0087] 953: Engine Interface
[0088] 954: Analog signal source
[0089] 1000: 3D engine system
[0090] 1000E, 1000F, 1133: Feature 0
[0091] 1001E, 1001F: Feature 1
[0092] 1010: CPU
[0093] 1020: GPU
[0094] 1030: RAM
[0095] 1050: SSD
[0096] 1060: HDD
[0097] 1070: I / O
[0098] 1080: Bus / Interface
[0099] 1510E, 1510F: Feature 510
[0100] 1511E, 1511F: Feature 511
[0101] 4900, 4901, 4902, 4910, 4911, 4912, 4920, 4921, 4922, 4930, 4931, 4932: Unit
[0102] 9000: 3D engine
[0103] 9001: Features
[0104] 9003: Feature Encoder
[0105] 9004: Feature Input
[0106] 9005: Search encoder
[0107] 9006: Search input
[0108] 143,9007: Sense amplifier
[0109] 9008: Page buffer, cache area, priority encoding area
[0110] 9009: Output matching results
[0111] 9010: Word Line
[0112] 9030: Memory string
[0113] 9040: Matching line
[0114] 9050: 2D engine
[0115] 9099: Results
[0116] 9900: Storage characteristics DETAILED DESCRIPTION
[0117] Will refer to Figures 1A to 10 DETAILED DESCRIPTION A detailed description of techniques related to the architecture of a three-dimensional search engine.
[0118] One or more flow charts are described herein. The processing described by the flow chart can be performed and / or directed using a processor, the processor is programmed using a computer program stored in a memory accessible to a computer system and executable by the processor, using specialized logic hardware (including field programmable integrated circuits), and using various combinations thereof. Various actions are combinable, can be performed in parallel, and / or can be performed in different orders without affecting the processing performed. In some cases, the rearrangement of the actions can only obtain the same result if some other changes are also made. In other cases, the rearrangement of the actions can only obtain the same result if some conditions are met. In addition, for clarity, some flow charts in this article omit some actions that are not required for understanding the disclosed technology. Various additional actions can be performed before, after, and / or between the actions shown.
[0119] Selected acronyms, mnemonics and abbreviations that will be used in the description are listed below.
[0120]
[0121]
[0122] An example of a memory string is a plurality of memory devices connected in series. An example of a memory device is an element capable of storing information indicating one of at least two mutually exclusive states of the memory device. These states can be set by programming the memory device and can be read by activating a control input of the memory device. In some types of memory devices, programming is performed by configuring the threshold voltage of the memory device. Configuration sometimes refers to programming the memory device (e.g., programming to 1 or 0). Configuration sometimes simply refers to storing a value (e.g., 1 or 0).
[0123] 3D (search) engine concept
[0124] A novel three-dimensional search engine architecture that can be executed using three-dimensional NAND memory is disclosed herein, which can achieve high-performance search, comparison and / or sorting, such as those related to large data and / or artificial intelligence processing. This architecture provides low latency, high resolution, high content density, multi-block functionality and robust reliability. This architecture is suitable for fields such as large data search, artificial intelligence hardware accelerators and / or classifiers, approximate computing, associative memory, few-shot learning, SSD data management, DNA matching, data filtering, hyper-dimensional computing, and other applications that benefit from IMS and can be used for long search words and large data capacities. Because this architecture supports high-performance comparison, sorting and search, the term "search" is bracketed in "three-dimensional (search) engine" and "two-dimensional (search) engine".
[0125] A novel three-dimensional (search) engine receives a search for word lines of a non-volatile memory array. The three-dimensional (search) engine uses two word lines for each bit of search information. The three-dimensional (search) engine uses respective bit lines of the non-volatile memory array as respective match lines for the search. Respective memory strings (such as NAND memory strings) of the non-volatile memory array are used to store respective data characters, such as corresponding to the characteristics to be searched. Respective plurality of memory strings are coupled in parallel to respective shared bit lines.
[0126] In some use cases, read latency of a 3D (search) engine is reduced by operating multiple blocks in parallel (at the potential expense of relatively large current transients). Concurrent operation is accomplished by switching multiple enable controls, such as string select lines (SSLs).
[0127] In some usage scenarios, a three-dimensional (search) engine is used with expansion encoding and / or one-hot encoding to facilitate, for example, optional range search capabilities.
[0128] The 3D (search) engine has particular applicability to large data and / or artificial intelligence, such as for various search, comparison and / or sorting operations. A host agent directs the features stored in the 3D (search) engine and then provides the 3D (search) engine with a search to determine a match with the stored features. An indication of a match is returned to the host agent.
[0129] A three-dimensional (search) engine uses multiple two-dimensional (search) engines to perform parallel operations across two-dimensional (search) engines and may be referred to as a "search cube" or simply a "cube". A three-dimensional (search) engine is capable of performing multiple searches in parallel using multiple search parameters across multiple stored features. In some usage contexts, parallel searches are performed in each two-dimensional (search) engine. For example, each two-dimensional (search) engine is responsible for performing a unique search operation. In some usage contexts, parallel searches are performed in a specific two-dimensional (search) engine. For example, different parts of a specific two-dimensional (search) engine are responsible for performing a unique search operation. In some usage contexts, parallel searches are parallelized across multiple two-dimensional (search) engines and in one or more two-dimensional (search) engines. Match indications between search parameters and stored features are provided by the two-dimensional (search) engine for processing. For example, such processing includes determining the overall similarity between the search parameters and the stored features and / or identifying which of the one or more stored features are similar to the search parameters, such as by buffering, caching and / or one or more priority encodings.
[0130] Each two-dimensional (search) engine used by the three-dimensional (search) engine includes multiple memory devices to detect whether there is a match between the search information provided to the two-dimensional (search) engine, as determined by the search parameters and features and stored in the two-dimensional (search) engine. The two-dimensional (search) engine generates a match indication in response to the detection, and provides a match indication for buffering, caching, priority encoding and output for use in the system. Features are stored according to a feature storage code. Search information is encoded according to a search code. Storage codes and feature codes are, for example, expansion codes, unique hot codes, low-effect codes, and reverse expansion codes. The three-dimensional (search) engine can selectively perform full matches and approximate matches, for example, according to storage codes and / or feature codes. Full matches correspond to full bit matches, for example, full bit matches of the search parameters match all bits of the value stored in the memory string. Approximate matches correspond to less than full bit matches, for example, full bit matches except for one bit match, full bit matches except for two bit matches, or partial bit matches, for example, 90% bit matches (such as 9 bits out of 10 bits match).
[0131] A control agent of a three-dimensional engine selectively operates two-dimensional engines in parallel. Parallel operation allows for parallel matching determinations to be made for the same set of search information provided by the two-dimensional engines operating in parallel. Each two-dimensional engine generates a matching indication based on the search information and features stored in the two-dimensional engine. A priority encoder of the three-dimensional engine processes the matching indication to determine relative matching between two-dimensional engines and / or within one or more two-dimensional engines.
[0132] The two-dimensional (search) engine has some memory strings that can be operated in parallel. The memory strings of the two-dimensional (search) engine are coupled in parallel between the match lines and the source lines of the two-dimensional (search) engine. Each memory string can be used to store a corresponding feature, each feature having multiple bits and selected to be stored according to one of the feature codes. Each memory string can be used to compare the (encoded) search input with the (encoded) feature stored in the memory string. Before searching, the control agent performs pair-wise programming of the features in the memory device according to the selected feature code. Pair programming is based on using a pair of serially connected memory devices for each bit of a feature. In order to search, the control agent drives the control input in pairs according to the search information and the selected search code. The search information includes a search pattern and optional mask information and / or range information for codes that can be searched according to a range. The search results are provided as a match indication for multiple subsequent selectively buffered, cached and / or prioritized codes.
[0133] 3D (search) engine example implementation
[0134] An exemplary three-dimensional (search) engine embodiment uses floating gate transistors as memory devices. The floating gate transistors are instantiated in NAND series coupled memory strings. Therefore, the three-dimensional (search) engine is sometimes referred to as a three-dimensional NAND search engine or more broadly referred to as a search system or an operation system. The memory strings are sequentially instantiated in parallel between shared bit lines and shared source lines. The memory strings coupled to the shared bit lines are sometimes referred to as a two-dimensional (search) engine or more broadly referred to as a memory.
[0135] The floating gate transistor is programmed to store a bit of a characteristic according to at least two mutually exclusive states. The programming configures a threshold voltage of the floating gate transistor. The programmed floating gate transistor conditionally responds (e.g., by a change in impedance) to indicate a match between the stored characteristic and a search provided on a character line coupled to a control gate of the floating gate transistor.
[0136] The word lines can be used to provide one or more search inputs to search for stored features in floating gate devices programmed into the memory string. Each memory string can be used to compare a search provided via a word line coupled to the memory string with stored features programmed into the memory string.
[0137] Each shared bit line may be used as a match line to indicate zero, one, or more matches between one or more search inputs respectively provided to each memory string coupled to the shared bit line and the stored characteristics of the respective memory string.
[0138] The match indications provided by the shared bit lines are prioritized (eg, according to a predetermined priority scheme) and one or more matching results are indicated to one or more other agents in the system. Thus, multiple two-dimensional (search) results are combined into one or more three-dimensional (search) results.
[0139] Figures 1A to 1C Various aspects of an exemplary 3D (search) engine are collectively shown, namely 3D engine 9000, which is, for example, a NAND-based 3D (search) engine.
[0140] Figure 1A shows an overview of an exemplary three-dimensional (search) engine, Figure 1B Shown in different details Figure 1A Some aspects of an exemplary three-dimensional (search) engine of Figure 1C More details are shown Figure 1A Some aspects of an exemplary three-dimensional (search) engine.
[0141] Refer back to Figure 1A , this diagram has been simplified for clarity.
[0142] The 3D engine 9000 includes multiple (eg, 128K) identical (or substantially similar) 2D (search) engines, such as 2D engine 9050. For clarity, any details of all engines except one 2D (search) engine have been omitted.
[0143] Each of the two-dimensional (search) engines of the two-dimensional engine 9050 has a plurality of memory strings (in principle represented by memory string 9030) to store feature information provided to the three-dimensional engine 9000. The stored features are in principle represented by stored features 9900. The memory string 9030 is also used to determine a match between the stored feature information and the search information provided to the three-dimensional engine 9000 by conditionally influencing respective match lines of the match line 9040. For example, the conditionally influencing includes a match between a portion based on the search input and a portion of the stored features.
[0144] Feature encoder 9003 receives feature 9001 provided to 3D engine 9000 (e.g., from a host agent). In response, feature encoder 9003 generates feature input 9004 suitable for application to a memory string of a 2D (search) engine for storage as part of stored features 9900 by encoding the selected feature. Each memory string of each 2D (search) engine is capable of storing a unique feature, so feature input 9004 is represented as having a unique coupling with each 2D (search) engine of 2D engine 9050. The unique coupling can be implemented via a separate signal for each 2D (search) engine, a set of shared signals (e.g., a bus) that share a multitasking time, or any other suitable communication mechanism that can execute each memory string of each 2D (search) engine for storing a unique feature.
[0145] The search encoder 9005 receives a search 9002 (e.g., from a host agent) provided to the three-dimensional engine 9000. In response, the search encoder 9005 generates a search input 9006 for the memory string of the two-dimensional (search) engine suitable for application to the two-dimensional (search) engine 9050 through the selected search code to determine a match (if any) between the search input 9006 and the stored features 9900. Each two-dimensional (search) engine receives the same search information, so the search input 9006 is represented as having parallel coupling between each two-dimensional (search) engine. This is in contrast to a unique coupling for distributed feature information. Parallel coupling can be achieved by separately buffering signals to each two-dimensional (search) engine, a set of coupling signals (such as a bus), or any other suitable communication mechanism that enables each two-dimensional (search) engine to receive the same search information. Each memory device of each memory string of a particular two-dimensional (search) engine receives a unique word line (such as provided by the search input 9006), such as Figure 1A The word lines conceptually shown in FIG. 1 are denoted as WLs 9010. Although each two-dimensional (search) engine receives the same search information, each memory string in the two-dimensional (search) engine receives unique search information.
[0146] In response to the search input 9006 and based on the storage characteristics 9900 of each two-dimensional (search) engine, each two-dimensional (search) engine bit line (such as each match line 9040) provides respective match information to respective sense amplifiers (SAs) 9007 to determine the match indication of each two-dimensional (search) engine. The match indication is then processed by the page buffer, cache area, priority encoding area 9008 and provided as an output match result 9009, and transmitted to the host agent as a result 9099, for example.
[0147] Reference to Figure 1B, which shows in more detail Figure 1A Some aspects of an exemplary three-dimensional (search) engine. Figure 1B and Figure 1A The components with the same component numbers are the same components. Figure 1A , for clarity, Figure 1B Simplified.
[0148] Three of the two-dimensional (search) engines (such as 2D (search) Engine_0150, 2D (search) Engine_128K-2158 and 2D (search) Engine_128K-1159 in the figure) are shown as having memory strings that can be operated in parallel. Figure 1A As summarized, the three-dimensional engine 9000 includes a circuit of sense amplifiers. Each two-dimensional (search) engine has a sense amplifier, collectively referred to as SAs 9007. For clarity, three sense amplifiers are shown, each coupled to a respective one of the two-dimensional (search) engines shown. The three-dimensional engine 9000 further includes post-sensing circuitry, collectively referred to as page buffer, cache area, priority encoding area 9008 to process the output of each sense amplifier of SAs 9007 to produce the result of the matching search as output matching result 9009.
[0149] Each 2D (search) engine is shown in the X dimension and the Z dimension. In some variations of building an integrated circuit based on planar processing, the hierarchy of the integrated circuit is built in the Z dimension. The 2D (search) engines are arranged in the Y dimension.
[0150] Reference to Figure 1C , which shows in more detail Figure 1A and Figure 1B Some aspects of an exemplary three-dimensional (search) engine. Figure 1C , Figure 1B and Figure 1A The components with the same component numbers are the same components. Figure 1A and Figure 1B , for clarity, Figure 1C Simplified.
[0151] Each 2D (search) engine includes a separate bit line. Specifically, 2D (search) Engine_0, 150, 2D (search) Engine_128K-2, 158 and 2D (search) Engine_128K-1, 159 include BL_0, 140, BL_128K-2, 148 and BL_128K-1, 149 bit lines, respectively. Figure 1AIn the figure, the bit lines are collectively referred to as match lines 9040, because each bit line can be used to indicate a match between search and storage information. Each two-dimensional (search) engine includes multiple memory strings (such as 512 memory strings). For clarity, 4 memory strings are shown. Each memory string is selectively coupled to the bit line of the two-dimensional (search) engine through a transistor controlled by a separate SSL input. The transistor and the SSL input jointly selectively enable (or disable) the memory string based on the enablement (or disablement) of the SSL input. Each SSL input is coupled to the memory string of each two-dimensional (search) engine. Therefore, each SSL input can selectively enable (or disable) the memory string based on the enablement (or disablement) of the SSL input.
[0152] For example, the string selection line 170 of SSL0 is enabled to selectively couple the first memory string of the two-dimensional (search) engine 150 of 2D (search) Engine_0 to the bit line 140 of BL_0, the first memory string of the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2 to the bit line 148 of BL_128K-2, and the first memory string of the two-dimensional (search) engine 159 of 2D (search) Engine_128K-1 to the bit line 149 of BL_128K-1, all through respective transistors. The string selection line 170 of SSL0 can also selectively couple the first memory string (not shown) of the two-dimensional (search) engine 128K-3 between the two-dimensional (search) engine 150 of 2D (search) Engine_0 and the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2 to the bit lines therein, respectively, through respective transistors.
[0153] For another example, the string selection line 179 of SSL511 is enabled to selectively couple the 512th memory string of the two-dimensional (search) engine 150 of 2D (search) Engine_0 to the bit line 140 of BL_0, couple the 512th memory string of the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2 to the bit line 148 of BL_128K-2, and couple the 512th memory string of the two-dimensional (search) engine 159 of 2D (search) Engine_128K-1 to the bit line 149 of BL_128K-1, all through respective transistors. The string selection line 179 of SSL511 can also selectively couple the 512th memory string (not shown) of the two-dimensional (search) engine 128K-3 between the two-dimensional (search) engine 150 of 2D (search) Engine_0 and the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2 to the bit lines therein through respective transistors.
[0154] Similar to the bit lines, each of the two-dimensional (search) engines includes a respective source line (shown but not labeled for clarity). Each memory string is selectively coupled to the source line of the two-dimensional (search) engine through a transistor controlled by the ground select line input of the respective GSL.
[0155] Similar to the bit line, for example, the ground selection line 160 of GSL0 is enabled to selectively couple the first memory string of the two-dimensional (search) engine 150 of 2D (search) Engine_0 to the source line of the two-dimensional (search) engine 150, couple the first memory string of the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2 to the source line of the two-dimensional (search) engine 158, and couple the first memory string of the two-dimensional (search) engine 159 of 2D (search) Engine_128K-1 to the source line of the two-dimensional (search) engine 159, all through respective transistors. The ground selection line 160 of GSL0 can also selectively couple the first memory string (not shown) of the two-dimensional (search) engine 150 of 2D (search) Engine_0 and the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2 to the source line therein through respective transistors. All source lines are grounded.
[0156] Similar to the bit line, for another example, the ground selection line 169 of GSL511 is enabled to selectively couple the 512th memory string of the two-dimensional (search) engine 150 of 2D (search) Engine_0 to the source line of the two-dimensional (search) engine 150, couple the 512th memory string of the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2 to the source line of the two-dimensional (search) engine 150, and couple the 512th memory string of the two-dimensional (search) engine 159 of 2D (search) Engine_128K-1 to the source line of the two-dimensional (search) engine 150, all through respective transistors. The ground selection line 169 of GSL511 can also selectively couple the 512th memory string (not shown) of the two-dimensional (search) engine 128K-3 between the two-dimensional (search) engine 150 of 2D (search) Engine_0 and the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2 to the source lines therein through respective transistors.
[0157] The string select line input of the SSL and the ground select line input of the GSL are coupled in parallel to the respective transistors in each respective memory string in each two-dimensional (search) engine. Since there are 512 memory strings in each two-dimensional (search) engine, there are 512 string select line inputs of the SSL and 512 ground select line inputs of the GSL. Note that Figure 1C , for clarity, the coupling between the ground selection line 160 of GSL0 and the transistors of the two-dimensional (search) engine 159 of 2D (search) Engine_128K-1 is omitted, but a dotted line extension of the ground selection line 160 of GSL0 is shown.
[0158] In the 3D engine 9000, each memory string is 128 floating gate transistors of equal length (not including the transistors of the source line coupling each memory string to the bit line). In the planar integrated circuit process implementation of the 3D engine 9000, some number of levels are used to implement the memory strings. As the length of the memory string increases, the number of levels used to implement the memory string also increases. Therefore, doubling the length of the memory string doubles the number of levels used to implement the memory string.
[0159] Regarding word lines, in principle, each two-dimensional (search) engine is coupled to an identical two-dimensional array of word lines. In the X dimension, there are 512 rows of word lines, which correspond to 512 memory strings one by one. In the Z dimension, there are 128 columns of word lines, which correspond to the floating gate transistors of each memory string one by one. Therefore, the two-dimensional array of word lines is 512 rows and 128 columns or 512*128 word lines. Each word line is coupled to a single floating gate transistor of a single memory string in each two-dimensional (search) engine. It should be noted that in Figure 1C , for clarity, the coupling between the word line 100 of WL0_0 and the floating gate transistor of the two-dimensional (search) engine 159 of 2D (search) Engine_128K-1 is omitted, but a dotted line extension of the word line 100 of WL0_0 is shown.
[0160] For example, the word line 100 of WL0_0 is coupled to the floating gate transistor of the first memory string of the two-dimensional (search) engine 150 of 2D (search) Engine_0, the floating gate transistor of the first memory string of the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2, and the floating gate transistor of the first memory string of the two-dimensional (search) engine 159 of 2D (search) Engine_128K-1. The word line 100 of WL0_0 is also coupled to the floating gate transistor of the first memory string (not shown) of each two-dimensional (search) engine of 128K-3 between the two-dimensional (search) engine 150 of 2D (search) Engine_0 and the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2.
[0161] For another example, the word line 109 of WL0_511 is coupled to the floating gate transistor of the 512th memory string of the two-dimensional (search) engine 150 of 2D (search) Engine_0, the floating gate transistor of the 512th memory string of the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2, and the floating gate transistor of the 512th memory string of the two-dimensional (search) engine 159 of 2D (search) Engine_128K-1. The word line 109 of WL0_511 is also coupled to the floating gate transistor of the first memory string (not shown) of each two-dimensional (search) engine 128K-3 between the two-dimensional (search) engine 150 of 2D (search) Engine_0 and the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2.
[0162] In this document, word lines are sometimes represented by block numbers and level numbers in a two-dimensional environment. In principle, a block exists in the Y and Z dimensions, and a level exists in the X and Y dimensions. In principle, the block number corresponds to the row number, and the level number corresponds to the column number.
[0163] One block corresponds to a memory string from each two-dimensional (search) engine. For example, a memory string coupled between ground selection line 169 of GSL511 and string selection line 179 of SSL511 corresponds to block number 511. For another example, a memory string coupled between ground selection line 160 of GSL0 and string selection line 170 of SSL0 corresponds to block number 0.
[0164] A level corresponds to floating gate transistors from each memory string of all two-dimensional (search) engines. For one specific example, floating gate transistors coupled to word line 109 of WL0_511 and word line 100 of WL0_0 are part of level number 0. For another specific example, floating gate transistors coupled to word line 119 of WL1_511 are part of level number 1. Similarly, floating gate transistors coupled to word line 129 of WL2_511, word line 180 of WL126_0, and word line 190 of WL127_0 are part of level numbers 2, 126, and 127, respectively. Levels also correspond to levels with respect to planar integrated circuit processes, but the floating gate transistor layers correspond to one or more levels of planar integrated circuit processes.
[0165] Therefore, word lines can be identified as "WL" in the Z dimension, "<level>" and "<block>" in the X dimension, respectively. Therefore, the word line 100 of WL0_0 has a level number of 0 and a block number of 0. The word line 109 of WL0_511 has a level number of 0 and a block number of 511.
[0166] Figure 1D Shown in Figure 1A and Figure 1B130, MemStr1 131, MemStr510 138, and MemStr511 139. The floating gate transistors of each memory string are arranged and operated in pairs. One example of these pairs is IMS data bit 134. Another example is IMS data bit 135. In principle, each of the word lines of even-numbered levels is coupled to one of the floating gate transistors in a pair. For example, the word lines of WL0_0, WL0_1, ..., WL0_510 and WL0_511 of level 0 are respectively coupled to respective ones of a pair, such as WL2_0, ..., WL2_511 of level 2, and WL126_0, ..., WL126_511 of level 126. In addition, each word line of an odd-numbered level is coupled to the other one of a pair of floating gate transistors. For example, the word lines of WL1_0, WL1_1, ..., WL1_510 and WL1_511 of level 1 are respectively coupled to the other floating gate transistors in the pairs to which the word lines of WL0_0, WL0_1, ..., WL0_510 and WL0_511 of level 0 are respectively coupled.
[0167] Figure 1D also shows a specific use case of a two-dimensional (search) engine.
[0168] Each memory string can be used to store a feature. Therefore, the 512 memory strings of the two-dimensional (search) engine 150 of 2D (search) Engine_0 can be used to store up to 512 features. For clarity, one feature is labeled as feature 0, 1133. Each pair of floating gate transistors can be used to store a single bit of one of the features. There are 64 pairs of floating gate transistors in each memory string. Therefore, each memory string can be used to store up to 64 bits of features.
[0169] Each memory string is operable to compare a signature stored therein with a search input provided via a word line coupled to the memory string. Each floating gate transistor of the memory string responds to a word line voltage (e.g., a threshold voltage of the floating gate transistor configured by programming) according to a signature bit stored in the floating gate transistor. For example, when operating in pairs, a pair of floating gate transistors indicates a match to a search input by a found match impedance (e.g., a relatively low impedance), which is provided to the pair of floating gate transistors via a pair of word lines. The pair of floating gate transistors indicates a mismatch by a not-found match impedance (e.g., a relatively high impedance). Thus, in the entire memory string, in response to the stored signature bit pairs of all floating gate transistors matching a multi-bit search input provided via a corresponding word line pair, the entire memory string indicates a corresponding memory string match-found impedance. In response to less than all paired matches, the entire memory string indicates a corresponding memory string impedance greater than the memory string found impedance. In response to no paired matches, the entire memory string indicates a corresponding memory string no-match-found impedance. Between the extremes of all bit matches and no bit matches, the memory string indicates a corresponding memory string impedance that decreases (increases) monotonically with more (fewer) bit-wise matched storage signature bits compared to the applied word line. The string current 136D indicates the current in the memory string based on the indicated memory string impedance.
[0170] For example, the memory string 130 of MemStr0 is coupled to the word line 140 of BL_0 via activation of the string select line of SSL0, and is coupled to the shared source line 145 of CSL via activation of the ground select line of GSL0. The feature 0, 1133 stored in the memory string 130 of MemStr0 is programmed to be all 0 bits except for the single 1 bit in the floating gate transistor pair of IMS data bit 135.
[0171] Consider a first operating scenario for an example, in which all word line pairs (WL0_0, WL1_0, ..., WL126_0, and WL127_0) coupled to the memory string 130 of MemStr0 indicate a search for all 0 bits in pairs. In response, all floating gate transistor pairs (except the IMS data bit 135 pair) of the memory string 130 of MemStr0 indicate that a matching impedance is found. The IMS data bit 135 pair indicates that a matching impedance is not found (due to a mismatch between the word line search input and the storage feature bit). Therefore, the memory string 130 of MemStr0 as a whole indicates an impedance greater than the memory string matching impedance found. The sense amplifier of SA 143 senses the impedance greater than the memory string matching impedance found and indicates a mismatch between the word line pair and the feature 0, 1133 via the sense output 144.
[0172] Consider a second operational scenario for the example, where the word line pair (WL2_0 and WL3_0) coupled to the IMS data bit 135 indicates a search for a 1 bit, and the rest of the word line pairs (WL0_0, WL1_0, WL4_0, WL5_0, ..., WL126_0, and WL127_0) coupled to the memory string 130 of MemStr0 indicate a search for a zero bit in pairs. In response, all floating gate transistor pairs of the memory string 130 of MemStr0 (including the IMS data bit 135 pair) indicate a found match impedance. Therefore, the memory string 130 of MemStr0 as a whole indicates an impedance equal to the memory string found match impedance. The sense amplifier of SA143 senses this impedance as the memory string found match impedance and indicates via the sense output 144 that there is a match between the word line pair and the feature 0, 1133.
[0173] The sense amplifier of SA 143 senses and generates sense output 144, for example, by detecting a voltage change greater than (or less than) a preset sense threshold on word line 140 of BL_0 during a sense time interval. In some variations, the voltage change is due to the discharge of the memory string 130 of MemStr0 at a precharge value on word line 140 of BL_0 according to the indicated memory string impedance. In response to a first operating scenario where the impedance is greater than the memory string found matching impedance, the memory string 130 of MemStr0 discharges the word line 140 of BL_0 more slowly than in a second operating scenario where the impedance is equal to the memory string found matching impedance. The sense threshold and / or sense time interval are set to be able to distinguish between found matching impedance and non-found matching impedance based on sensing speed (such as bandwidth and / or latency of the engine) and / or noise immunity (such as resilience of the engine).
[0174] continue Figure 1A , Figure 1B and Figure 1C In the context of the first or second operation scenario, the sense output 144 is provided to the page buffer, cache, priority encoding area 9008 along with the corresponding sense amplifier information of other two-dimensional engines. The page buffer, cache, priority encoding area 9008 buffers, caches and / or prioritizes the sense output 144 in the context of the sense output of the entire two-dimensional engine.
[0175] In some applications, a controller activates multiple blocks of a two-dimensional (search) engine to enable parallel searching of multiple blocks. For example, multiple SSL inputs are activated in the same search operation to search multiple blocks during the same search operation.
[0176] For example, the controller activates the string selection lines of SSL_511 and SSL_510 in the same search operation to enable parallel search of two blocks in each of the 128K two-dimensional (search) engines (i.e., the two-dimensional (search) engine 150 of 2D (search) Engine_0, ..., the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2, and the two-dimensional (search) engine 159 of 2D (search) Engine_128K-1). Therefore, the controller enables the three-dimensional engine 9000 to search 2*128K blocks in parallel. Alternatively, the controller activates any two SLL inputs in the same search operation to search any two blocks in parallel. For other examples, the controller activates any 4, 6, 8, 32, 64, 128, 256, or 512 SSL inputs in the same search operation to enable parallel search of corresponding multiple blocks. For example, the controller enables all 512 SSL inputs in the same search operation, so the controller enables the 3D engine 9000 to search 512 blocks in parallel, including 512*128K memory strings. Optionally, the controller is responsive to a control register that specifies whether to search multiple blocks in parallel and to what extent.
[0177] Figure 1A , Figure 1B and Figure 1CAn exemplary three-dimensional (search) engine having a specific number of two-dimensional (search) engines is shown, each of which has a specific number of memory strings, each of which has a specific number of IMS data bits (each as a separate floating gate transistor pair). Specifically, the three-dimensional engine 9000 has 128K two-dimensional (search) engines (i.e., two-dimensional (search) engine 150 of 2D (search) Engine_0, ..., two-dimensional (search) engine 159 of 2D (search) Engine_128K-1), but some variations have fewer or more two-dimensional (search) engines. Each two-dimensional (search) engine has 512 memory strings (such as memory string 130 of MemStr0, ..., memory string 139 of MemStr511), but some variations have fewer or more memory strings. Each memory string has 64 IMS data bits (eg, word line pairs coupled to WL0_1 and WL1_0, ..., WL126_0, and WL127_0), but in some variations each memory string has fewer or more IMS data bits.
[0178] Some variations of the two-dimensional (search) engine and associated sense amplifier circuitry may enable the sense amplifier to resolve a single bit (e.g., match / mismatch). Some variations may enable the sense amplifier to resolve multiple bits (e.g., code mismatch, a match, or two or more matches). Some variations may enable the sense amplifier to make analog indications rather than digital indications. For example, a sense amplifier output is a monotonically increasing function based on a memory string current, and the memory string current is a monotonically increasing (or decreasing) indicator of similarity between the search input and the stored signature.
[0179] Some variations of a three-dimensional engine are capable of identifying at most one feature (stored in a single memory string) as a match for a search. Some variations are capable of identifying at most multiple features (stored in multiple memory strings) in parallel as matches for a search. Some variations are capable of identifying at most one two-dimensional engine in parallel as having one or more features stored therein as matches for a search. Some variations are capable of identifying multiple two-dimensional engines in parallel, each having one or more features stored therein as matches for a search.
[0180] Some variations of the three-dimensional (search) engine can be prioritized according to preset two-dimensional (search) engine priorities (such as a preset two-dimensional (search) engine is the lowest priority, and another preset two-dimensional (search) engine is the highest priority). Therefore, the match indication from the preset highest priority two-dimensional (search) engine corresponds to a highest priority match. In addition, the match indication from the preset second highest priority two-dimensional (search) engine corresponds to a high priority match. Some variations are prioritized according to how many storage feature bits in the memory string match the respective search input bits (such as zero matches, one match, two matches, or three or more matches). Therefore, the highest priority match corresponds to the maximum number of storage feature bits in the memory string that match the respective search input bits, and the second highest priority match corresponds to the second largest number of storage feature bits in the memory string that match the respective search input bits. Some variations are prioritized according to how many memory strings match in a particular two-dimensional (search) engine (such as zero matches, one match, two matches, or three or more matches). Thus, the highest priority match corresponds to the largest number of matches of the memory string in the particular two-dimensional (search) engine, and the second highest priority match corresponds to the second largest number of matches of the memory string in the particular two-dimensional (search) engine.
[0181] Some variations enable preferential encoding to identify a single matching stored signature. Some variations enable preferential encoding to identify multiple matching stored signatures. Some variations enable preferential encoding to identify a single "best matching" stored signature having, for example, a maximum number of stored signature bits that match a search, where the maximum number is less than a maximum stored signature length. Some variations enable preferential encoding to identify multiple "best matching" stored signatures, which are ranked, for example, according to the number of stored signature bits that match a search for each of a plurality of stored signatures.
[0182] In various usage scenarios (e.g., due to features programmed into the two-dimensional (search) engine, the search information provided, and / or its encoding), various numbers of two-dimensional (search) engines in a three-dimensional (search) engine detect a match between one or more features stored in one or more memory strings of the two-dimensional (search) engine and the search information of the two-dimensional (search) engine provided to the three-dimensional (search) engine.
[0183] In some use scenarios, zero, one, or more memory strings of a two-dimensional (search) engine detect matches, and the sense amplifier circuit encodes the match results into a plurality of bits. For example, the sense amplifier circuit encodes a match of zero memory strings, a match of one memory string, a match of two memory strings, and a match of more than two memory strings into four mutually exclusive codes. Some variations prioritize encoding based on one or more factors, such as a two-dimensional (search) engine priority, a number of matches of a storage feature bit, and a number of memory strings (e.g., features).
[0184] For example, in response to no match from any of the two-dimensional (search) engines 150 of 2D (search) Engine_0, ..., 2D (search) Engine_128K-1, 159, the sense amplifiers of SAs 9007 provide a vector of all zeros to the page buffer, cache, and priority coding area 9008. In response, the page buffer, cache, and priority coding area 9008 perform priority encoding to encode the all-zero vector into a binary value, which indicates a zero match, and the page buffer, cache, and priority coding area 9008 perform priority encoding to provide this binary value by outputting the match result 9009.
[0185] Based on the above example, in response to detecting a matching single memory string of the two-dimensional (search) engine 150 of 2D (search) Engine_0, the sense amplifier of SAs 9007 provides a vector having a least significant bit (LSB) corresponding to the two-dimensional (search) engine of 2D (search) Engine_0 as 1 and all other bits as 0. In response, the page buffer, cache area, priority encoding area 9008 encodes this vector into a non-zero binary value, and this non-zero binary value identifies the two-dimensional (search) engine 150 of 2D (search) Engine_0 as having a matching memory string.
[0186] Based on the above example, in response to multiple memory strings of the two-dimensional (search) engine 150 of 2D (search) Engine_0 detecting a match and a single memory string of the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2 detecting a match, the sense amplifiers of SAs 9007 provide a vector having a bit corresponding to the two-dimensional (search) engine 150 of 2D (search) Engine_0 as 1, another bit corresponding to the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2 as 1, and all other bits as 0. In response, the page buffer, cache area, priority encoding area 9008 encodes this vector into a non-zero binary value, and this non-zero binary value identifies the two-dimensional (search) engine 158 of 2D (search) Engine_128K-2 as the memory string with the highest priority match.
[0187] 3D (search) engine coding
[0188] The searching, comparing and / or sorting performed by a three-dimensional (search) engine is based on a selected feature encoding (for feature storage in a memory string of the three-dimensional (search) engine) and a selected search encoding (for searches performed by the memory string). The features and search encoding are selected in conjunction with each other because the matching performed by the memory string is performed in the context of the encoding selected for the feature storage.
[0189] Exemplary encodings that can be used to store features and / or perform searches in a three-dimensional (search) engine include unfolding encoding, one-hot encoding, low-effect encoding (principle a variation of unfolding encoding that can reduce pattern effects and body effects), and inverse unfolding encoding.
[0190] As a brief introduction, Figure 1E and Figure 1F Respectively similar to Figure 1D 2D(search)Engine_0 shows an example of the use of expansion encoding and one-hot encoding in the context of the two-dimensional (search) engine 150. Correspondingly, Figure 2 and Figure 3 Examples of unfolded encoding and one-hot encoding are shown respectively.
[0191] 3D (search) engine expansion code
[0192] Referring first to the unfolded encoding, an N-bit binary value (feature or search) is encoded in a 2^N bitfield. The encoding produces two consecutive bit fields. Each of the adjacent fields is a repetition of 1 or 0. Therefore, 2^N-1 is encoded as a single leading 0 and all others are 1. The number of 1s is equal to the value being encoded. In a special case, a 0 is encoded as all 0s. Therefore, the most-significant bit is always 0. Therefore, in some variations, the most-significant bit cannot be implemented.
[0193] Figure 2 An exemplary expanded encoding for a three-dimensional (search) engine is shown. There are two rows in the table. The value 251 in the first row is a 6-bit binary value that will be encoded according to the expanded encoding. The expanded encoding 252 in the second row is the 64-bit binary expanded encoding of the corresponding binary value in the first row. For ease of reading, every four bits of the 64-bit value are separated by an underscore symbol. For brevity, this table omits some values.
[0194] Figure 1E An example of the use of the expansion code of the two-dimensional (search) engine of the three-dimensional (search) engine is shown, specifically Figure 1D The two-dimensional (search) engine 150 of 2D (search) engine_0. Figure 1E and Figure 1D The same component numbers indicate the same components.
[0195] in accordance with Figure 2 Four exemplary features of the unfolding code are summarized in Figure 1E Specifically, the memory string 130 of MemStr0 is programmed to store feature 0, 1000E, which is expanded to 1 (a single 1 bit and 63 0 bits). Each feature bit is stored in a serial pair of memory devices. The feature expansion code is Figure 1A 1001E, expanded encoding is 4 (four 1 bits and 60 0 bits). Feature 2, 509 is programmed to store expanded encodings of various values not specifically discussed. Memory string 138 of MemStr510 is programmed to store feature 510, 1510E, expanded encoding is 5 (five 1 bits, of which 4 1 bits are shown, and 59 0 bits). Memory string 139 of MemStr511 is programmed to store feature 511, 1511E, expanded encoding is 63 (63 1 bits and a single 0 bit).
[0196] In the usage scenario shown, also according to Figure 2 The search expansion code encodes the search. The search expansion code is composed of Figure 1A126_0 and WL127_0. In response to the expanded encoding of 63 presented on the word lines of block 0, the memory string 130 of MemStr0 enables a current indicated by string current 136E. This current indicates a match sensed by the sense amplifier of SA 143 and presented on sense output 144 for further processing.
[0197] In response to the stored signature not matching any of the corresponding search inputs, the current indicating the mismatch is sensed by the sense amplifier of SA 143 and presented on sense output 144 for further processing.
[0198] The expanded code can be used to search for an approximate match between the stored signatures. A best match (e.g., high similarity) has a highest string current (e.g., lowest impedance). A worst match (e.g., low similarity) has a lowest string current (e.g., highest impedance). The number of bits that match (or mismatch) can be determined by measuring string current and / or impedance.
[0199] Exemplary search expansion encoding resolutions include 4 bits implemented using 32 3D NAND layers, 5 bits implemented using 64 3D NAND layers, 6 bits implemented using 128 3D NAND layers, 7 bits implemented using 256 3D NAND layers, and 8 bits implemented using 512 3D NAND layers. Thus, the resolution increases by 1 bit as the number of 3D NAND layers doubles.
[0200] One-hot encoding for 3D (search) engines
[0201] Next, referring to one-hot encoding, an N-bit binary value (feature or search) is encoded in a 2^N bit field. The encoding produces a single 1 bit, while all other bits are 0. The single 1 bit is located at the bit position corresponding to the encoded value.
[0202] Figure 3 An exemplary one-hot encoding for a three-dimensional (search) engine is shown in tabular form. There are two rows in the table, and the value 351 of the first row is a 6-bit binary value encoded according to the one-hot encoding. The one-hot encoding 352 of the second row is a 64-bit binary one-hot encoding of the corresponding binary value in the first row. For ease of reading, every four bits of the 64-bit value are separated by an underscore symbol. For brevity, this table omits some values.
[0203] Figure 1FAn example of the use of one-hot encoding for a two-dimensional (search) engine of a three-dimensional (search) engine is shown, specifically Figure 1D The two-dimensional (search) engine 150 of 2D (search) engine_0. Figure 1F and Figure 1D The same component numbers indicate the same components.
[0204] in accordance with Figure 3 Four exemplary features encoded by one-hot encoding are summarized in Figure 1F Specifically, the memory string 130 of MemStr0 is programmed to store feature 0, 1000F, which is one-hot encoded as 1 (a single 1 bit in bit position 1 and 63 0 bits in other positions). Each feature bit is stored in a serial pair of memory devices. The feature one-hot encoding is Figure 1A 1003 in the feature encoder 9003. The memory string 131 of MemStr1 is programmed to store feature 1, 1001F, where the one-hot encoding is 3 (a single 1 bit at bit position 3 and 63 0 bits in other positions). Feature 2, 509 is programmed to store one-hot encodings of various values not specifically discussed. The memory string 138 of MemStr510 is programmed to store feature 510, 1510F, where the one-hot encoding is 4 (a single 1 bit at bit position 4 and 63 0 bits in other positions). The memory string 139 of MemStr511 is programmed to store feature 511, 1511F, where the one-hot encoding is 63 (a single 1 bit at bit position 63, the most significant bit, and 63 0 bits in other positions).
[0205] In the usage scenario shown, also according to Figure 3 The search is encoded by the one-hot encoding of Figure 1A 1. The search for the value 1 is encoded on the word line pair as a search for a single 1 bit and 63 0 bits in bit position 1. Assuming this search is in block 0, the word line pairs used for the search are WL0_0, WL1_0, ..., WL126_0, and WL127_0. In response to the one-hot encoding of 1 presented on the word line of block 0, the memory string 130 of MemStr0 enables a current indicated by string current 136F. This current indicates a match sensed by the sense amplifier of SA 143 and presented on sense output 144 for further processing.
[0206] In response to the stored signature not matching any of the corresponding search inputs, the current indicating the mismatch is sensed by the sense amplifier of SA 143 and presented on sense output 144 for further processing.
[0207] 3D (search) engine low-efficiency coding
[0208] Next, reference is made to low-effect coding, which is a variation of spreading coding that can reduce pattern effects and body effects of floating-gate transistors in a memory string. In principle, floating-gate transistors and corresponding word lines are rearranged so that floating-gate transistors programmed with the same threshold voltage tend to be contiguous with each other in a memory string. This is achieved by changing the spreading coding rather than actually moving transistors.
[0209] Figure 4A An exemplary expanded encoding and a corresponding exemplary low-efficiency encoding are shown. For simplicity, these examples are for 16 floating-gate transistors of a memory string and corresponding 16 word lines.
[0210] The table includes three columns. The value 461 of the first column is a binary value to be encoded. The expanded code 462 of the second column is an 8-bit expanded code of the value of the first column. The low-significant code 463 of the third column is an equivalent 8-bit low-significant code of the value of the first column.
[0211] The rows of this table have three parts. The first part is the search input 404, which represents a single row of respective codes of the same search value 3. The second part is the storage feature 464, which has nine columns, each column representing a respective storage feature value according to the expansion code or the low-effect code. The third part is the memory string 430, which represents from left to right starting from the bit line "BL", passing through the string select line "SSL", passing through multiple word lines "WL...WL", passing through the ground select line "GSL" and ending at the shared source line "CSL".
[0212] The "HL" entry in the expanded code 462 corresponds to a data bit 1, while the "LH" entry corresponds to a data bit 0. Each "H" or "L" character corresponds to a threshold voltage of a floating gate transistor. The "H" entry in the storage feature 464 corresponds to a "high" threshold voltage, while the "L" entry corresponds to a "low" threshold voltage. The H2 and H1 entries in the search input 404 correspond to two corresponding word line voltages. The specific search input is represented as 3. The underscore in the expanded code 462 and the low effect code 463 is used to indicate the conceptual rearrangement of the floating gate transistors and the corresponding word lines.
[0213] Figure 4B An exemplary expanded code having a Hamming distance (HD) compared to input 1 is shown. Input 1 corresponds to a search input 1 and is encoded according to the expanded code represented as input 450. This exemplary code is Figure 2There are three rows in the table. The value 451 in the first row is a 6-bit binary value that will be encoded according to the expanded code. The expanded code 452 in the second row is the 64-bit binary expanded code of the corresponding binary value in the first row. The third row is HD 453, which is the Hamming distance between each expanded code and the expanded code of the search input 1 (input 450). When the input matches the stored value, the Hamming distance is minimum (0). When the input does not match almost all bits of the stored value, the Hamming distance is maximum (62). For ease of reading, every four bits of the 64-bit value are separated by underscores. For brevity, some values are omitted from this table.
[0214] Figure 4C Exemplary threshold voltages and word line voltages that can compare 1, 0, and don't care are shown. The left portion of this figure shows cells in memory state 1, such as memory state 1, 471. The right portion of this figure shows cells in memory state 0, such as memory state 0, 470. This figure shows two sets of relationships between threshold voltages and word line voltages for storing features in floating gate transistors and driving search information on word lines to control the gate of the floating gate transistor. Depending on the usage scenario, either set can be used.
[0215] The first group corresponds to a first word line voltage VH1 having a value of V2, 473 and a second word line voltage VH2 having a value of V3, 474. Thus, for floating gate transistors programmed to be in memory state 1, 471, applying VH1 (having a value of V2, 473) or applying VH2 (having a value of V3, 474) will result in a high gate overdrive. For floating gate transistors programmed to be in memory state 0, 470, applying VH2 (having a value of V3, 474) will result in a high gate overdrive. However, applying VH1 (having a value of V2, 473) will result in a low gate overdrive of said floating gate transistors.
[0216] The second group corresponds to a first word line voltage VL having a value of V1, 472 and a second word line voltage VH having a value of V3, 474. Thus, for a floating gate transistor programmed to be in memory state 1, 471, applying VL (having a value of V1, 472) will result in a high-side gate overdrive. Additionally, for a floating gate transistor programmed to be in memory state 0, 470, applying VL (having a value of V1, 472) will result in no gate overdrive. For a floating gate transistor programmed to be in memory state 1, 471 or memory state 0, 470, applying VH (having a value of V3, 474) will result in a high-side gate overdrive.
[0217] Figure 4DAn icon associating high / low-side gate overvoltage drive with low / high floating gate transistors and cell impedance is shown, which in principle corresponds to high / low cell current capabilities, respectively. High-side gate overvoltage drive 483 uses a relatively thick arrow to indicate high-side gate overvoltage drive. Low-side gate overvoltage drive 484 uses a relatively thin arrow to indicate low-side gate overvoltage drive. High cell current 485 uses two relatively thick arrows to indicate relatively high cell current capability (e.g., relatively low impedance). Low cell current 486 and low cell current 487 use a relatively thick arrow and a relatively thin arrow to indicate relatively low cell current capability (e.g., relatively high impedance).
[0218] Figure 4E The current capability (e.g., impedance) of a floating gate transistor as a function of search information for stored data is shown. Stored data value 499 has four rows, store 0, 490, store 1, 491, random item 492, and store invalid (StoreInvalid) 493. Input search information 489 has three columns, search 0, 480, search 1, 481, and wildcard (Wildcard), 482. Each cell in the table is based on Figure 4D The icons shown in the figure represent the impedance of each floating gate transistor, and the cells in the table are also based on Figure 4D The diagram shown in shows the effective impedance of a pair of floating gate transistors. Cells 4900, 4902, 4911, 4912, 4920, 4921, 4922, and 4932 represent gate overdrives of both of a pair of floating gate transistors, resulting in a relatively low impedance. Cells 4901, 4910, 4930, and 4931 represent gate overdrives of only one of a pair of floating gate transistors, resulting in a relatively high impedance. A relatively low impedance indicates a match between the search and data, while a relatively high impedance indicates the opposite.
[0219] As a specific example, consider cell 4900 in a usage scenario of a first set of relationships between threshold voltage and word line voltages VH1 (having value V2, 473) and VH2 (having value V3, 474), where Vy equals VH2 and Vx equals VH1. Thus, the upper floating-gate transistor controlled by Vy (equal to VH2) has a high-side gate overvoltage drive and thus has a high current capability corresponding to a low impedance. The lower floating-gate transistor controlled by Vx (equal to VH1) also has a high-side gate overvoltage drive and thus has a high current capability corresponding to a low impedance. Thus, overall, cell 4900 has a high current capability corresponding to a low impedance.
[0220] As another specific example, consider the cell 4900 in a use scenario of a second set of relationships between the threshold voltage and the word line voltage VL (having a value of V1, 472) and the second word line voltage VH (having a value of V3, 474), Vy equals VH, and Vx equals VL. Therefore, the upper floating gate transistor controlled by Vy (equal to VH) has a high-side gate overvoltage drive, and thus has a high current capability corresponding to a low impedance. The lower floating gate transistor controlled by Vx (equal to VL) also has a high-side gate overvoltage drive, and thus has a high current capability corresponding to a low impedance. Therefore, in general, the cell 4900 has a high current capability corresponding to a low impedance.
[0221] 3D (search) engine reverse expansion coding
[0222] Next, we refer to the inverse expansion coding, which is in principle the logical inversion of the expansion coding, but is applied only to the search input rather than the features. The inverse expansion coding is suitable for approximate matching.
[0223] Figure 5A An example inverse expansion encoding for a search input to a three-dimensional (search) engine and an exemplary (non-inverse) expansion encoding for storing features are shown in table form.
[0224] The table is organized into three sections. The top section shows, for example, applying the search value to generate an inverse expanded encoding for a search input applied to a word line. The middle section shows an example input as an inverse expanded encoding. The bottom section shows the stored value expanded encoding and the Hamming distance for the example input.
[0225] The search value 555 of the first row in the upper portion is a 6-bit binary search value that will be encoded according to the inverse expansion encoding. The inverse expansion encoding 554 of the second row is the 64-bit binary inverse expansion encoding of the corresponding binary value in the first row. It should be noted that similar to the expansion encoding, when the 64-bit binary value is inversely expanded, the most significant bit is always 1. Therefore, in some variations, the most significant bit is not implemented.
[0226] The middle portion is the reverse expanded encoding of the search value 1, such as the search value example 550.
[0227] The first row of the lower portion is the stored value 551, which is the 6-bit binary feature to be encoded according to the expansion code to be stored in the memory string. The expansion code 552 of the second row is the 64-bit binary expansion code of the corresponding binary value in the first row. The HD 553 of the third row is the Hamming distance between the stored value with the respective expansion code and the search input 1 (search value example 550) in the reverse expansion code. As described above, in the 64-bit binary expansion code, the most significant bit is always 0, while in the 64-bit binary reverse expansion code, the most significant bit is always 1. Therefore, comparing a specific value encoded according to the 64-bit binary expansion code with a specific value encoded according to the 64-bit binary reverse expansion code, there is always a mismatch in the most significant bit. Therefore, HD 533 ignores the most significant bit. When the stored value is closest to the search input in the reverse expansion code (when the most significant bit is ignored), the Hamming distance is the smallest (1). When the stored value is farthest from the search input in the reverse expansion code (when the most significant bit is ignored), the Hamming distance is the largest (64).
[0228] For readability, 64-bit values are separated by underscores for every four bits. This table omits some values for brevity.
[0229] Figure 5B Exemplary Hamming distances (HDs 599) as a function of stored data values (stored data values 590) and search input values (search input values 506) for an inverse unfolded code are graphically illustrated.
[0230] Figure 6 An exemplary Hamming distance (HDs 699) of a function of stored data values (stored data values 690) and search input values (search input values 606) for an expanded code is graphically shown. Figure 5B , Figure 6 Shows how the inverse expansion coding is "reverse-reversed" relative to the expansion coding.
[0231] about Figure 5B and Figure 6 , relatively small Hamming distances are shown in relatively light shades, and relatively large Hamming distances are shown in relatively dark shades. The lowest Hamming distance is 0 (indicating equality) and is shown in the lightest shade. The highest Hamming distance is 63 (indicating a 63-bit mismatch) and is shown in the darkest shade.
[0232] exist Figure 5B In the reverse unfolding code, the Hamming distance along the diagonal is 0; conversely, Figure 6 (Expanded Code), the Hamming distance along the diagonal is 63. In addition, Figure 5B (In the reverse expansion code, the Hamming distance between the lower left corner and the upper right corner is 63; on the contrary, Figure 6 (Expanded Code), the Hamming distance between the lower left corner and the upper right corner is 0. Therefore, Figure 5B The Hamming distance of Figure 6 The Hamming distance of is "reverse".
[0233] The inverse unwinding code can be used to search for an approximate match between the stored signatures. A best match (e.g., high similarity) has a highest memory string current (e.g., lowest impedance). A worst match (e.g., low similarity) has a lowest memory string current (e.g., highest impedance). The number of bits that match (or mismatch) can be determined by measuring the memory string current and / or impedance.
[0234] 3D (search) engine one-hot encoding storage and input range
[0235] Fig. 7A An example of a one-hot encoding of a storage range is shown in a multi-part table. The table is divided into two parts, the upper part shows an example input (value 2) as a one-hot encoding of input 750. The lower part shows the one-hot encoding of the storage range and the matching information related to the example input.
[0236] The first row of the lower portion is the value range 751, which is the entry for each range of 6-bit features that will be encoded and stored as features to match the search input. For example, 0-2 specifies a stored feature to match a search input from 0-2 (0, 1, or 2). The second row is the one-hot encoding of the storage range 752, which is a 64-bit binary one-hot encoding for the range in the first row. The character X indicates that the corresponding floating gate device pair is configured to match the search input 1 or 0, as is an optional item. The third row is Match?, 753, which is whether the example input 2 (the one-hot encoding of the storage range 752) matches the corresponding storage range feature. Example input 2 matches the first three one-hot encoding ranges in the table. Note that one-hot encodings without ranges (e.g. Figure 3 The expanded encoding 352) only matches the example input 2 with stored feature value 2.
[0237] For readability, 64-bit values are separated by underscores for every four bits. This table omits some values for brevity.
[0238] Figure 7B An example of a one-hot encoding of an example input range is shown in a multi-part table. This table is divided into two parts, the upper part shows an example input range (0-4) as a one-hot encoding 760 of the example input range. The character X indicates that the corresponding search input is a wildcard to match any corresponding stored feature bit. The lower part describes the one-hot encoded storage value and matching information about the example input range. The lower part shows the one-hot encoded storage value and matching information related to the exemplary input range.
[0239] The first line of the lower portion is value 761, which is a 6-bit binary value that will be encoded according to the one-hot encoding. The second line is the one-hot encoding of the stored range, 762752, which is a 64-bit binary one-hot encoding for the range in the first line. The third line is Match?, 763, which is whether the example input range 0-4 matches the corresponding stored feature.
[0240] For readability, 64-bit values are separated by underscores for every four bits. This table omits some values for brevity.
[0241] One-hot encoding can be used to search for an exact match between a stored signature. One-hot encoding with a stored range can be used for approximate match searches. Programming a stored signature with more random terms can increase the range of search inputs that the stored signature matches. One-hot encoding of input ranges can be used for approximate match searches. Changing which bits are wildcards will shift the range of search inputs that are matched. For example, a stored signature XX1XX000...00 matches search inputs 0-4, while a stored signature 0XX1XX000...00 matches search inputs 1-5.
[0242] 3D (search) engine operation
[0243] Figure 8 For 3D (search) engines (such as Figure 1A An exemplary operation flow of a three-dimensional engine 9000) is indicated by engine flow 800. The process begins with step 801 of selecting feature and search codes. In some usage scenarios, the same code is used for feature and search. In some usage scenarios, such as reverse expansion code, different codes are used for feature and search. Next, the process proceeds to step 802 of programming features, programming the features to be searched according to the selected feature code. Next, the process proceeds to step 803 of code search, encoding the search according to the selected search code. In some applications, the search is performed based on one or more parameters (optionally including mask information and / or range information). Next, the process proceeds to step 804 of driving word lines, driving word lines according to the search code. Next, the process proceeds to step 805 of sensing bit lines. The bit lines are sensed by sensor amplification. Next, the process proceeds to step 806 of priority code matching, processing the match indication provided by the sense amplifier to determine the priority match information (for example, by Figure 1A The page buffer, cache area, and priority encoding area 9008 are executed). Then, the process proceeds to step 807 to provide the result, and the result (for example, Figure 1A The output matching result 9009 is provided to, for example, a host agent. Then, the process selectively returns to encode the next search (the encoding search of step 803).
[0244] 3D (search) engine devices and equipment
[0245] Fig. 9 An example 3D engine apparatus including an example 3D engine device is shown, denoted as 3D engine apparatus 940 and 3D engine device 950 .
[0246] The 3D engine device 940 includes one or more instances of a 3D engine apparatus. For clarity, only a single apparatus (ie, the 3D engine apparatus 950 ) and the apparatus interface 941 are shown.
[0247] The 3D engine device 950 includes a 3D engine 9000 (such as Figures 1A to 1C 9000). The 3D engine device 950 also includes a controller 951, an engine interface 953, and an analog signal source 954. In principle, the components of the 3D engine device 950 and / or the 3D engine apparatus 940 (except the 3D engine 9000) serve to provide features 9001 and searches 9002 from a host to the 3D engine 9000, and return results 9099 to the host.
[0248] The controller 951 is enabled to control the operation of the 3D engine device 950 as a whole and the operation between its various hardware circuit blocks. For example, the controller 951 coordinates the operation of the engine interface 953 and the 3D engine 9000. For another example, the controller 951 coordinates the operation of the configuration information of the analog signal source 954. For another example, the controller 951 selectively coordinates the feature and / or search encoding and feature programming as indicated via the engine interface 953 and applicable to the 3D engine 9000. In other examples, the controller 951 is enabled to perform the operation related to the 3D engine 9000 in various configurations. Figure 8 The operation of any one or more of steps 801-807.
[0249] The engine interface 953 enables an agent external to the 3D engine device 950 (e.g., a host, via the device interface 941 and the host / bus coupling 942) to connect to use the 3D engine 9000. Exemplary uses are receiving commands and data, and returning status related to feature and / or search codes, feature programming, and / or search requests.
[0250] The analog signal source 954 is enabled to generate and / or provide one or more analog outputs to, for example, a memory array, a bit line circuit, a sense amplifier circuit, and / or a word line circuit of a three-dimensional (search) engine. The analog outputs include, for example, zero or more voltage and / or current sources (such as reference sources) generated by one or more bias circuits.
[0251] In operation, a host provides commands such as feature selection and / or search codes via the host / bus coupling 942 and device interface 941. The device interface 941 then provides the commands to the engine interface 953, and the commands are processed by the 3D engine 9000 as directed by the controller 951. Other similar processing commands include, for example, feature programming according to the selected feature code and searching according to the selected search code. The status or processing commands and search results are returned to the host by the controller 951 and / or the 3D engine 9000 via the host / bus coupling 942 and device interface 941 provided by the engine interface 953. In some variations of the 3D engine device 950, the controller 95 enables one or more 2D (search) engines of the 3D engine 9000 to operate in parallel.
[0252] In operation, a host external to the 3D engine device 940 provides feature information to the 3D engine apparatus 950 to be programmed into the 3D engine 9000. The host then provides one or more search parameters to the 3D engine device 940 via the device interface 941.
[0253] Although Fig. 9 The 3D engine 9000 is shown to be used in a 3D engine device 950 and then in a 3D engine apparatus 940, but other usage scenarios are possible. For example, various components (e.g., one or more dies of an integrated circuit, a system-on-a-chip, and / or a packaged die) include one or more 3D engines that are the same as, substantially similar to, or based on the 3D engine component 9000.
[0254] Other partitioning of the elements or additional elements shown in the drawings, couplings between the elements or additional elements, and capabilities and / or capacities of the elements or additional elements may be considered depending on usage requirements.
[0255] 3D (search) engine system
[0256] Fig.10 An exemplary hardware system with a 3D (search) engine is shown, such as a 3D engine system 1000. This system includes hardware blocks CPU 1010, GPU 1020, RAM 1030, 3D engine device 940 (shown in FIG. 1 ), and 3D engine device 940 (shown in FIG. 1 ). Fig. 9 ), SSD 1050, HDD 1060 and I / O 1070.
[0257] CPU 1010 includes one or more processing units, such as any combination of hardware units, microprocessors, signal processors, artificial intelligence processors, etc., enabled to execute programmed instructions.
[0258] GPU 1020 includes one or more processing units, such as any combination of units that enable accelerated processing that tends to be relatively highly parallel, such as graphics processing, signal processing, and / or artificial intelligence processing.
[0259] RAM 1030 includes one or more storage elements for storing instructions and / or data that are larger than the internal storage of CPU 1010 and / or GPU 1020. RAM 1030 may be implemented by volatile storage elements such as DRAMs and / or SRAMs.
[0260] SSD 1050 includes one or more storage elements, such as flash technology-based storage elements, for storing instructions and / or data, which may be selectively accessed with less latency than HDD 1060 .
[0261] HDD 1060 includes one or more storage elements, such as rotational magnetic and / or optical non-volatile storage elements (eg, disk platters) for storing larger amounts of instructions and / or data than SDD 1050 is capable of storing.
[0262] I / O 1070 includes interface elements that connect any combination of CPU 1010, GPU 1020, RAM 1030, 3D engine device 940, SSD 1050, and / or HDD 1060 to external components of 3D engine system 1000. Exemplary external components include mass storage devices, local and wide area networks (such as the Internet), human interface components (such as keyboards, mice, and / or displays), and other components that provide expansion functions and / or enhanced functions not otherwise provided by 3D engine 1000. I / O 1070 variously includes one or more serial and / or parallel channels and selective protocol conversion and / or adaptation capabilities to facilitate communication between components coupled thereto.
[0263] The bus / interface 1080 enables communication between components coupled thereto, such as the CPU 1010, GPU 1020, RAM 1030, 3D engine device 940, SSD 1050, HDD 1060, and / or I / O 1070. The bus / interface 1080 variously has one or more serial and / or parallel channels and optional protocol conversion and / or adaptation capabilities to facilitate communication between components coupled thereto.
[0264] Other partitions of the elements or additional elements shown in the drawings, couplings between the elements or additional elements, and capabilities and / or capacities of the elements or additional elements may be considered depending on usage requirements.
[0265] In some cases, the 3D engine device 940 may be enabled to perform processing locally to reduce the processing load (eg, on the CPU 1010 and / or GPU 1020 ) and / or reduce the amount of traffic on the bus / interface 1080 .
[0266] For example, using one-hot encoding (e.g., for storing features and / or searching input ranges) to enable range matching, once the features are programmed, searching between large databases is completely done in the three-dimensional engine device 940. For another example, using inverse unfolding encoding to enable approximate matching, once the features are programmed, comparing between large artificial intelligence data sets is completely done in the three-dimensional engine device 940. For another example, using unfolding encoding and / or one-hot encoding, once the features are programmed, sorting of large data sets is completely done in the three-dimensional engine device 940.
[0267] Additional information about 3D (search) engines
[0268] Exemplary memory technologies suitable for the memory array of the three-dimensional (search) engine disclosed herein include floating-gate, split-gate, SONOS, suspended point, DRAM, DRAM-like (e.g., 2TOC), FeFET, and any memory technology compatible with searching via word lines and bit lines. Exemplary SONOS memory technology (sometimes referred to as charge trapping memory) uses a trap insulating layer (e.g., silicon nitride) to capture and retain charge injected from the channel. Exemplary suspended point storage technology in principle uses floating silicon nanodots to replace the floating gate or embeds the floating silicon nanodots in a polysilicon gate. Exemplary 2TOC memory technology uses the parasitic capacitance of a read transistor rather than an explicit storage capacitor to store charge. Exemplary FeFET memory technology uses permanent electric field polarization of a ferroelectric material embedded between a gate and a source-gate conduction region to store information. Exemplary memory structures suitable for the memory array of the three-dimensional (search) engine include two-dimensional structures (e.g., two-dimensional flash structures) and three-dimensional structures (e.g., three-dimensional flash structures). Exemplary array architectures of memory arrays suitable for use in three-dimensional (search) engines include NOR / OR array architectures and AND / NAND array architectures.
[0269] It should be understood that the above disclosure is intended to illustrate rather than limit the embodiments, variations, embodiments and examples of the present invention. Modifications and combinations within the spirit of the present disclosure and the scope of the attached patent applications are identifiable.
Claims
1. A computing system, comprising: A plurality of memories, each of the memories comprising: a respective matching line coupled to a respective sense amplifier; a respective source line; and a plurality of respective memory strings, each coupled to the respective match line and the respective source line, and each of the memory strings comprising a plurality of floating gate transistors connected in series, each of the floating gate transistors having and responsive to a respective configuration state and a respective control input; a plurality of word lines, each of the word lines being coupled to a respective one of the respective control inputs of each of the floating gate transistors; as well as a search encoder enabled to receive a search and drive the word lines according to a search code; wherein each of the memory strings is enabled to couple the respective match line and the respective source line, the respective memory string is coupled to the respective match line and the respective source line by a discovery matching impedance that fully matches the control inputs of the series-connected floating-gate transistors of the respective memory string in response to the configuration state of the respective memory string, and each of the sense amplifiers is enabled to generate a respective indication for indicating whether the match line to which it is coupled is coupled to the source line via the discovery matching impedance; and wherein the respective memory string is coupled to the respective match line and the respective source line by an unfound matching impedance that does not completely match the control inputs of the respective memory string in response to a configuration state of the respective memory string, and each of the sense amplifiers is enabled to generate a respective indication of whether the match line to which it is coupled is coupled to the source line via the unfound matching impedance. 2 . The operating system of claim 1 , wherein the word lines are operated in pairs according to the search code, and the configuration states are managed in pairs according to a feature code. 3 . The computing system of claim 1 , wherein each of the configuration states corresponds to a respective threshold voltage. 4 . The computing system of claim 3 , wherein each of the respective control inputs is coupled to a respective gate of one or more of the series-connected floating-gate transistors. 5 . The operating system of claim 3 , wherein the threshold voltages are settable by one or more operations.
6. The computing system of claim 1, wherein each of the memory strings comprises a respective memory string enable device, the respective memory string enable device being responsive to an enable control input of a respective memory string to selectively enable or disable the respective memory string independently of the word lines.
7. A computing method, comprising: receiving one or more search parameters; encoding the search parameters to generate a plurality of pairs of word line voltages according to a specific one of a plurality of search codes, the specific search code searching for a match of a characteristic match stored in a pair of floating gate transistors according to a specific one of a plurality of characteristic codes, and wherein the specific search code and the specific characteristic code are enabled to represent 1, 0, and 1 or 0; driving word line pairs of a plurality of serially connected instances of the paired floating gate transistors according to the generated word line voltages, the serially connected instances being organized into respective memory strings operable in parallel to affect a plurality of shared match lines, each of the shared match lines being coupled to two or more of the memory strings, and each word line pair being coupled to at least two of the instances of the paired floating gate transistors; sensing voltage changes on the shared match lines to determine which of the shared match lines has a voltage change greater than a predetermined sensing threshold; and The determined results are provided as an indication of a relative match between the search parameters and the stored features. 8 . The operation method according to claim 7 , wherein at least one of the search codes or at least one of the feature codes is based on at least one of an expansion code and a one-hot code.
9. The operation method according to claim 7, wherein at least one of the search codes or at least one of the feature codes is encoded according to a reverse expansion.
10. The computing method of claim 7, wherein at least one of the stored features corresponds to a range stored using a one-hot encoding.
11. The calculation method of claim 7, wherein at least one of the paired word line voltages is according to a range encoded as a one-hot encoding. 12 . The operation method according to claim 7 , wherein at least one of the search codes or at least one of the feature codes is based on a low-effect code.
13. A computing system comprising: an interface circuit for enabling a host agent to provide a search and receive one or more results of the search; a block of a plurality of memory strings, each of the blocks comprising a plurality of individual memory strings, each of the memory strings receiving a same set of word lines having voltages determinable in response to the search; a plurality of sense amplifiers, each of the sense amplifiers coupled to a corresponding one of the memory strings from each of the blocks; a priority encoder circuit enabled to receive match indications from the sense amplifiers; wherein the sense amplifiers are enabled to determine a match, a mismatch, and at least a match between characteristic information stored in the memory strings coupled to the sense amplifiers and word lines received by the memory strings coupled to the sense amplifiers; and The priority encoder circuit enables decisions based on the sense amplifiers to indicate at least a portion of the results as a highest priority match of any matches identified by the sense amplifiers.
14. The operating system according to claim 13, further comprising the host agent.
15. The computing system of claim 13, wherein based on the determination made by the sense amplifier, the priority encoder circuit enables a high priority match to indicate at least a portion of the results as any match identified by the sense amplifier. 16 . The computing system of claim 13 , wherein the characteristic information is stored in the memory strings by configuring threshold voltages of a plurality of floating gate transistors of the memory strings.
17. The computing system of claim 13, wherein two or more of the blocks are operable in parallel.
18. The operating system of claim 13, wherein a feature code is used to encode at least some portion of the feature information, and a search code is used to encode at least a portion of the search. 19 . The computing system of claim 18 , wherein the feature encoding is at least one of an unfolding encoding, an inverse unfolding encoding, a one-hot encoding, and a low-efficiency encoding.
20. The computing system of claim 13, wherein the computing system is part of a single chip system.
