Architecture for computational memory and memory system
Patent Information
- Application Number
- JP2023147714
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-05
- Filing Date
- 2023-09-12
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2043-09-12
Smart Images

Figure 2024160913000001_ABST
Abstract
Description
[Technical field]
[0001] This disclosure relates to memory capable of performing computations, including memory that enables computational solid-state drives (SSDs). [Background technology]
[0002] Increasing amounts of data increase the processing, storage, and bandwidth requirements for managing, organizing, selecting, and analyzing the data in real time. Techniques are needed to address these increasing requirements.
[0003] In an exemplary conventional system, data is stored in a storage device accessible to a processor through a bus. Before processing, the data is transferred from the storage device to the processor (and / or memory directly accessible to the processor) through the bus. The data is then manipulated by the processor and, optionally, the results are returned to the storage device through the bus. Thus, as the amount of data increases, processor utilization and / or bus bandwidth utilization increases. In some usage scenarios, throughput is limited by the bandwidth and / or latency between the processor and the storage device. Techniques are needed to reduce, for example, processor utilization and / or bus bandwidth utilization. Summary of the Invention
[0004] One or more computer systems may be configurable to perform certain operations and / or actions by having software, firmware, hardware, or a combination thereof installed on the system that, during operation, causes the system to perform and / or control operations and / or actions. One or more computer programs may be configurable to perform certain operations and / or actions by containing instructions that, when executed by a data processing device, cause the device to perform operations and / or actions.
[0005] A first aspect of a method for operating a memory system comprises first sensing a first portion of memory strings of a memory array according to (i) a first voltage driven on a word line of the memory array and (ii) a predetermined operating condition of the first portion of the memory strings; second sensing a second portion of the memory strings of the memory array according to (i) a second voltage driven on the word line of the memory array and (ii) a dynamic operating condition of the second portion of the memory strings; and providing a result of the second sensing, where the second voltage is based on a result of the first sensing. Other embodiments of this aspect comprise corresponding computer systems, apparatus, and computer programs recorded on one or more computer storage devices, each configured to perform the actions of the method.
[0006] The variants optionally include one or more of the following features: In a first aspect, the dynamic operating condition is generated by manipulating a bit line circuit terminal of the second portion of the memory string according to the dynamic value. The second sensing step optionally includes counting approximate matches between the value stored in the second portion of the memory string and the dynamic value. The second sensing step optionally includes summing according to a current in the second portion of the memory string. The dynamic value is an analog value, the value stored in the second portion of the memory string is an analog value, and the summing is according to an analog technique. The second sensing step optionally includes scaling and then summing the current in the second portion of the memory string. The dynamic operating condition is based on the dynamic value and a mask value that allows the dynamic value to match any value stored anywhere in the second portion of the memory string. The dynamic operating condition is based on the dynamic value and a mask value that disables the dynamic value from matching any value stored anywhere in the second portion of the memory string. A result of the second sensing step indicates a multiply-accumulate operation, a first operand of the multiply-accumulate operation being a dynamic value, a dynamic operating condition being determined according to the dynamic value, and a second operand of the multiply-accumulate operation corresponding to a value to be programmed into the second portion of the memory string. The dynamic operating condition is determined according to the dynamic value, and a result of the second sensing indicates a search operation for the dynamic value of the values programmed into the second portion of the memory string. Implementations of the described techniques optionally include hardware, methods or processes, or computer software on a computer-accessible medium.
[0007] A second aspect of the memory system optionally comprises a memory array including first and second portions of memory strings, where one or more of the memory strings of the first portion of the memory strings and one or more of the memory strings of the second portion of the memory strings each include a plurality of serially connected memory devices; word line circuitry having terminals coupled to respective control inputs of respective rows in which the memory strings are organized; bit line circuitry having terminals coupled to respective bit line terminals of the memory strings; sense amplifier circuitry having terminals coupled to respective sense amplifier terminals of the memory strings; where the bit line terminals of the first portion of the memory strings The terminals of the line circuit are capable of providing identical predetermined operating conditions to the bit line terminals of the first portion of the memory strings; wherein the terminals of the bit line circuit coupled to the bit line terminals of the second portion of the memory strings are capable of providing respective dynamic values to the bit line terminals of the second portion of the memory strings; wherein the word line circuit is capable of selectively driving the control inputs in accordance with one of first and second word line control values, the first word line control value corresponding to the provided value and the second word line control value corresponding to a result produced by the sense amplifier circuit based on information provided via the terminals coupled to the bit line terminals of the first portion of the memory strings. Other embodiments of this aspect include corresponding computer systems, apparatus, and computer programs stored on one or more computer storage devices, each configured to perform the actions of the method.
[0008] Variations optionally include one or more of the following features. The second aspect optionally comprises a counting circuit coupled to the sense amplifier circuit and capable of counting exact matches between the values stored in the second portion of the memory string and the respective dynamic values. The system optionally comprises a summing circuit coupled to the sense amplifier circuit and capable of summing according to currents in the second portion of the memory string. The respective dynamic values are analog values and the values stored in the second portion of the memory string are analog values, and the summing circuit is operable according to an analog technique. The summing circuit is further capable of scaling the currents before summing. The system optionally includes one or more processors capable of accessing the memory array. The terminals of the bitline circuits coupled to the bitline terminals of the second portion of the memory string further optionally include a page cache circuit coupled to the sense amplifier circuit and capable of providing the same predetermined operating conditions to the bitline terminals of the second portion of the memory string and capable of caching the resulting pages read from the values stored in the second portion of the memory string. Implementations of the described techniques optionally include hardware, methods or processes, or computer software on a computer-accessible medium.
[0009] A third aspect of the system optionally comprises a memory array including first and second portions of memory strings, where one or more of the memory strings of the first portion of the memory strings and one or more of the memory strings of the second portion of the memory strings each include a plurality of serially connected memory devices; a word line circuit having a terminal coupled to a respective control input of a respective row in which the memory strings are organized; a bit line circuit having a terminal coupled to each bit line terminal of the memory strings and operable to provide identical predetermined operating conditions to the bit line terminals; and a sense amplifier circuit having a terminal coupled to each sense amplifier terminal of the memory strings, the word line circuit operable to selectively drive the control input according to one of first and second word line control values, the first word line control value corresponding to a provided value and the second word line control value corresponding to a result generated by the sense amplifier circuit based on information provided via the terminal coupled to the bit line terminal of the first portion of the memory strings. Other embodiments of this aspect comprise corresponding computer systems, apparatus, and computer programs recorded on one or more computer storage devices, each configured to perform the actions of the method.
[0010] Variations optionally include one or more of the following features. In a third aspect, the provided values optionally include a key and a mask, and selectively driving the control inputs optionally includes selectively driving the control inputs in respective pairs according to a plurality of encodings respectively determined from respective bits of the key and corresponding respective bits of the mask, where (i) a first one of the encodings enables selectively matching a stored zero value in response to the provided value being zero, (ii) a second one of the encodings enables selectively matching a stored one value in response to the provided value being one, (iii) a third one of the encodings enables selectively matching a stored zero value or a stored one value in response to the provided value being zero or one, and (iv) a quarter of the encodings disable selectively matching any stored value to the provided value. The system optionally includes one or more processors that enable access to a computational solid-state drive (SSD) optionally including at least one memory device optionally including a memory array. Implementations of the described techniques optionally include hardware, a method or process, or computer software on a computer-accessible medium.
[0011] Variations of the above aspects optionally include hardware, methods or processes, or computer software on a computer-accessible medium.
[0012] Other aspects and features of the disclosure will be apparent from the drawings, the detailed description, and the claims. [Brief description of the drawings]
[0013] [Figure 1] 1 illustrates an exemplary hardware system based on one or more computational solid-state drives (SSDs).
[0014] [Diagram 2]1 illustrates an exemplary computational memory for performing a coarse search followed by a fine search responsive to results of the coarse search.
[0015] [Diagram 3] 1 illustrates an exemplary computational memory for performing a coarse search followed by calculations dependent on the results of the coarse search.
[0016] [Figure 4] 1 illustrates an exemplary computational memory for performing a coarse search followed by an access operation responsive to the results of the coarse search.
[0017] [Diagram 5] 1 illustrates an example flow diagram of a technique for operating a computational memory.
[0018] [Figure 6A] 13 summarizes the search through information provided on a word line. [Figure 6B] 13 summarizes the search through information provided on a word line.
[0019] [Figure 6C] 1 summarizes a search through the information provided on the bit lines. [Figure 6D] 1 summarizes a search through the information provided on the bit lines.
[0020] [Figure 7] 1 illustrates an exemplary computational SSD.
[0021] [Figure 8] 1 illustrates an exemplary hardware system having a computational SSD. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0022] A detailed description of an architectural technique for a computational solid-state drive (SSD) is provided with reference to FIGS.
[0023] One or more flow diagrams are described herein. The processes described by the flow diagrams can be implemented and / or instructed using a processor programmed using a computer program stored in a memory accessible to a computer system, can be executed by a processor using dedicated logic hardware (including field programmable integrated circuits), and various combinations thereof. Various actions can be combined, performed in parallel, and / or performed in different sequences without affecting the processing performed. In some cases, rearrangement of actions achieves the same result only if certain other changes are also made. In other cases, rearrangement of actions achieves the same result only if certain conditions are met. Moreover, for clarity, some of the flow diagrams herein omit some actions that are not necessary for understanding the disclosed techniques. Various additional actions can be performed before, after, and / or between the actions shown.
[0024] Examples of selected acronyms, mnemonics, and abbreviations used in the description are as follows: [Table 1] JPEG2024160913000003.jpg70128
[0025] An exemplary key-value store (e.g., a key-value SSD) is an object-based storage that receives an object description (e.g., a key) and, in response, accesses an object identifiable by the object description (e.g., a value corresponding to the key). In some key-value stores, a translation layer translates between the object description and a corresponding reference to the underlying data block in non-volatile storage. The Concept of Computational SSD
[0026] A system based on computational SSD technology described herein reduces processor utilization and / or bus bandwidth utilization. The system can use computational SSD resources to perform computational techniques (e.g., search, calculate, and / or access) rather than processor and / or bus resources, thus reducing or minimizing information movement between processing elements and storage devices. Computational SSD technology allows ever-increasing amounts of data to be managed, organized, selected, and analyzed in real time. Computational SSDs (sometimes referred to as "smart SSDs") can, for example, use computational SSD resources to store and manipulate data locally, run CIM, IMC, and / or IMS, etc. Thus, by using computational SSDs, the processing, storage, and bandwidth requirements of the system are reduced.
[0027] Additionally, the use of computational SSDs allows for the use of CPU and / or GPU resources for other tasks, enabling faster results. In some usage scenarios, the use of computational SSDs reduces the need to purchase additional processors and / or servers. In some usage scenarios, the use of computational SSDs reduces the need to purchase additional power and / or cooling resources. In some usage scenarios, the use of computational SSDs enables new insights and opportunities to use self-processing technology to accelerate various applications. Exemplary applications where computational SSD technology is applicable include video processing, database management, and artificial intelligence.
[0028] The exemplary computational SSD allows for performing a coarse search function followed by a second function responsive to the coarse search function. The exemplary computational SSD is based on a single memory array available in two parts: (1) a part dedicated to the coarse search function and (2) a part dedicated to the second function. The exemplary second function includes a fine search (e.g., IMS) function, a calculation (e.g., IMC) function, and an access (e.g., value for a key / value pair) function.
[0029] As one specific example, a key-value SSD can be implemented using the computational SSD described herein.
[0030] A computational SSD is comprised of one or more computational memories. A computational SSD may perform multiple functions using the same computational memories. For example, a first function may be performed on a particular computational memory, and then a second function (e.g., based on the result of the first function) may be performed on the particular computational memory. In some cases, multiple functions may be performed on the same memory array of the same computational memory.
[0031] For an example of a search operation, the first function includes a coarse filter and / or search function and the second function includes a fine filter and / or search function. For an example of a calculation operation, the first function includes a coarse filter and / or search function and the second function includes an in-memory calculation function. For an example of an access operation, the first function includes a coarse filter and / or search function and the second function includes a data access function.
[0032] In some scenarios, a coarse filter and / or search function can be used to reduce or eliminate processing on uninteresting or unnecessary portions of the database. For example, in response to a coarse filter function, word lines determined by the coarse filter and / or search function are activated for use in a subsequent action. In the example of search, computation, and access operations, the subsequent actions include a fine filter and / or search function, an in-memory computation function, and a data access function, respectively.
[0033] In some variations, the coarse filter and / or search function uses search information provided on the word lines, e.g., similar to a ternary CAM. In some variations, the fine search function uses search information provided on the bit lines, e.g., for hyper-dimensional searches. In some variations, the computation function uses computation information provided on the bit lines. In some variations, the access function is usable to provide a value corresponding to a key of a key-value pair, e.g., for use by a CPU and / or GPU. Exemplary Computational SSD System
[0034] In some computational SSD systems, one or more computational SSDs are capable of performing a first function followed by a second function dependent on the results of the first function. In a first example, the first function is a coarse search function and the second function is a fine search function. In a second example, the first function is a coarse search function and the second function is a calculation function. In a third example, the first function is a coarse search function and the second function is an access function.
[0035] In some configurable computational SSD systems, one or more computational SSDs are capable of performing a first function followed by a second configurable function depending on the results of the first function. In various examples, the granularity of the configuration of the second function may vary, for example, from the entire SSD to a single memory array of the configurable computational SSD. For example, an SSD may be configured as a specialized search computational SSD with the entire SSD configured with the first function as a coarse search function and the second function as a fine search function. For example, another SSD may be configured with a first portion of the memory array configured for a computation function and a second portion of the memory array configured with an access function.
[0036] 1 illustrates an exemplary hardware system based on one or more computational SSDs as a computational SSD system 100. The system includes hardware blocks CPU 110, GPU 120, DRAM 130, and one or more instances of computational SSD 140 coupled by hardware block PCIe bus 180. Each instance of computational SSD 140 includes one or more computational SSDs, such as one or more instances of computational SSD 700 of FIG. 7. The instances of computational SSD 140 are capable of performing computational SSD-related functions as indicated by the functions: course, fine, compute, access 126.
[0037] CPU 110 includes one or more processing units, such as any combination of hardware units capable of executing programmed instructions, such as microprocessors, signal processors, AI processors, etc. GPU 120 includes one or more processing units, such as any combination of units capable of accelerating processes that are subject to relatively high parallelism, such as graphics processing, signal processing, and / or AI processing. DRAM 130 includes one or more DRAM devices for storing instructions and / or data in a larger amount than the internal storage of CPU 110 and / or GPU 120. PCIe bus 180 enables communication between elements coupled thereto (e.g., CPU 110, GPU 120, and computational SSD 140 instances) using one or more PCIe buses 180. Computational SSD 140 instances can variously perform coarse search, fine search, calculation, and access functions on data stored on the respective computational SSDs without transferring the data to CPU 110, DRAM 130, and / or GPU 120.
[0038] Computational SSD system 100 allows for the reduction or minimization of data movement between conventional processing elements (e.g., CPU 110 and / or GPU 120) and computational storage elements (e.g., one or more instances of computational SSD 140). Various computational operations are performed on the computational SSD using locally stored data available on the SSD.
[0039] Other compartmentalizations of elements, connections between elements, and capabilities and / or capacities of the elements shown in the figures, as well as additional elements, are contemplated according to usage requirements. Computational Memory
[0040] Three examples of computational memories are described. A first example is a coarse search followed by two successive IMS operations, e.g., a fine search depending on the result of the coarse search. A second example is a coarse search followed by an IMS operation, e.g., a calculation operation, depending on the result of the coarse search. A third example is a coarse search followed by an access operation and a successive IMS operation, e.g., an access operation depending on the result of the coarse search. Computational Memory - IMS=>IMS
[0041] Returning to the first example of computational memory (two consecutive IMS operations), a single computational memory array has two portions (one for each of the IMS operations) each capable of performing a coarse search (e.g., operating as a TCAM) followed by a fine search depending on the results of the coarse search. Conceptually, the coarse search portion implements a filter (e.g., via the word line inputs) to reduce the calculations performed by the fine search portion (e.g., a hyper-dimensional search via the bit line inputs).
[0042] As an example of operation, selected coarse search information (e.g., all or any part of the key, optionally together with key mask information) is used to generate (coarse search) values for word lines of a single computational memory array ("array"). Bit lines of the coarse search portion are manipulated to allow matching of word line values to the coarse search information stored in the memory devices of the coarse search portion. The memory devices are arranged in contiguous groups of devices called memory strings. Each memory string allows matching of word line values to the coarse search information stored in the memory string. The memory string indicates a match, for example, by changing impedance. The change in impedance is then reflected by a change in voltage on the bit line to which the memory string is coupled. For example, the memory string indicates a match by discharging the bit line. A sense amplifier of the coarse search portion detects the bit line voltage change (if any) and generates a match result in response. An exact match and / or an approximate match corresponds to a high degree of similarity between the word line values and the coarse search information stored in the memory devices of the coarse search portion. For example, in one-hot encoding, a match is coded as a "1" and a mismatch is coded as a "0."
[0043] The matches are then used to generate values for the word lines of the array (fine search). In an exemplary one-hot encoding (single match), a single word line is driven to "read" and the remaining word lines are driven to "pass" to enable a fine search according to the coarse match results. At the same time, the bit lines of the fine search section are manipulated according to the matching selected fine search parameters (e.g., all or any portion of one or more hyperdimension search parameters, optionally in conjunction with hyperdimension mask information), and the fine search information is stored in the memory device of the fine search section. The sense amplifiers of the fine search section detect the bit line changes and generate match results indicating a high degree of similarity between the fine search parameters and the fine search information stored in the memory device, and the match results are then combined (e.g., counted) across one or more bit lines to determine one or more overall results. For example, the highest of the overall results corresponds to the best match.
[0044] The coarse search portion accommodates "match 1" and "match 0" match operations, as well as, optionally, "match anything" (e.g., "don't care" or alternatively, "wildcard") and "match none" (e.g., "invalid value" or alternatively, "invalid search") match operations through pairs of word lines and corresponding pairs of memory devices in the first portion. Similarly, the fine search portion accommodates similar encodings through pairs of bit lines and corresponding pairs of memory devices in the second portion.
[0045] FIG. 2 illustrates an exemplary computational memory for performing a coarse search followed by a fine search depending on the results of the coarse search as computational memory 200.
[0046] Computational memory 200 includes two portions: coarse 210 for performing a coarse search, and fine 220 for performing a fine search. Coarse portion 212 and fine portion 222 are portions of the same memory array, array 250. Word line circuitry 252 allows the results of the coarse search to be provided to the fine search.
[0047] The course 210 includes a bit line circuit 211, a course portion 212, and a sense amplifier circuit 214. The course portion 212 includes a plurality of memory strings, each including a plurality of serially coupled memory devices. The memory strings are conceptually coupled to the bit line circuit 211 and the sense amplifier circuit 214, and are thus serially coupled in the vertical direction of the figure. The memory strings are conceptually replicated in the horizontal direction of the figure, with the word lines coupling the corresponding devices of each of the memory strings in the horizontal dimension. For each memory string, one end of the memory string (e.g., a bit line terminal) is coupled to the bit line circuit 211, and the other end of the memory string (e.g., a sense line terminal) is coupled to the sense amplifier circuit 214.
[0048] The bitline circuitry 211 includes bitline circuitry capable of manipulating each of the bitline terminals according to a predetermined value. For example, the bitline circuitry 211 includes a precharger for each of the bitline terminals of the course portion 212. The precharger operates (in conjunction with other circuits, such as the wordline circuitry 252 and / or the sense amplifier circuitry 214) to precharge the memory strings of the course portion 212.
[0049] The sense amplifier circuit 214 includes a sense circuit capable of determining a current and / or voltage change of each of the sense amplifier terminals of the course portion 212 according to one or more predetermined thresholds. For example, the sense amplifier circuit 214 includes a sense amplifier for each of the sense amplifier terminals. The sense amplifier operates (in conjunction with other circuits, such as the word line circuit 252 and / or the bit line circuit 211) to determine a current and / or voltage change of the memory string of the course portion 212. The sense amplifier circuit 214 provides the result of the change to the word line circuit 252 via the course search result 215.
[0050] The current and / or voltage changes in the memory strings in course portion 212 result from, for example, selective discharging of the memory strings based on the stored values of the serially coupled memory devices of the memory strings and the values provided to the word lines by word line circuitry 252.
[0051] Fine 220 includes bit line circuitry 221, fine portion 222, sense amplifier circuitry 224, and post-processing (count) 225. Fine portion 222 (similar to coarse portion 212) includes multiple memory strings, each including multiple serially coupled memory devices. The memory strings are coupled to bit line circuitry 221 and sense amplifier circuitry 224, and are thus conceptually coupled in series in the vertical direction of the diagram. The memory strings are conceptually replicated in the horizontal direction of the diagram, with word lines coupling corresponding devices in each of the memory strings in the horizontal dimension. For each memory string, one end of the memory string (e.g., a bit line terminal) is coupled to bit line circuitry 221, and the other end of the memory string (e.g., a sense line terminal) is coupled to sense amplifier circuitry 224.
[0052] Bitline circuitry 221 (similar in some respects to bitline circuitry 211) includes bitline circuitry that is capable of operating each of the bitline terminals according to dynamic values provided by fine search parameters 229 (rather than predetermined values in bitline circuitry 211). For example, bitline circuitry 221 includes a conditional precharger for each of the bitline terminals in fine portion 222. Bitline circuitry 221 operates pairs of bitline terminals as conceptually represented by IMS cell (fine) 227. The conditional precharger operates (in conjunction with other circuitry, e.g., wordline circuitry 252 and / or sense amplifier circuitry 224) to conditionally precharge memory strings in fine portion 222.
[0053] Sense amplifier circuitry 224 includes sense circuitry capable of determining a current and / or voltage change in each of the sense amplifier terminals of fine portion 222 according to one or more predetermined thresholds. For example, sense amplifier circuitry 224 includes a sense amplifier for each of the sense amplifier terminals. The sense amplifier operates (in conjunction with other circuits, e.g., word line circuitry 252 and / or bit line circuitry 221) to determine a current and / or voltage change in a memory string in fine portion 222. Sense amplifier circuitry 224 provides results of the change to post-processing (counting) 225, such as a single binary digit for each of the sense amplifiers. Each single binary digit indicates one of two mutually exclusive conditions: (1) the change is less than or equal to the threshold, and (2) the change is greater than the threshold.
[0054] The current and / or voltage changes in the memory strings in the fine portion 222 result from selective discharging of the conditionally pre-charged memory strings based, for example, on the stored values of the serially coupled memory devices of the memory strings and the values provided to the word lines by the word line circuitry 252.
[0055] Post-processing (counting) 225 includes a counting circuit capable of counting the number of sense amplifiers of sense amplifier circuit 224 that indicate a change greater than (or alternatively less than or equal to) a threshold value. The counting indicates a relative match of the coarse search parameters 219 and / or the fine search parameters 229 compared to various values stored in the coarse portion 212 and / or the fine portion 222, respectively, that are provided (e.g., to a host) via fine search results 226. In another example, the outputs of a first portion of the sense amplifiers of sense amplifier circuit 224 are counted together as a first result provided via fine search results 226, and the outputs of the remaining portions of the sense amplifiers of sense amplifier circuit 224 are counted together as a second result provided via fine search results 226. Various configurations of sense amplifiers and counts can be used to enable comparison of various portions of the fine search parameters 229, such as one or more hyperdimensions.
[0056] Thus, the fine search results 226 include course search parameter 219 information stored in the course portion 212, information based on the fine search parameters 229, information stored in the fine portion 222, and information based on counting as performed by post processing (counting) 225.
[0057] The word line circuit 252 includes word line conversion and drive circuitry that can conditionally operate in two modes, e.g., a first mode during a first phase of operation and then a second mode during a second phase of operation. In the first mode, one or more search parameters, coarse search parameters 219 (e.g., a key and optionally a mask), are converted to voltages suitable for performing a coarse search function using memory strings in the coarse portion 212 (e.g., for parallel comparison against each instance of the key information 218). In the first mode, the word line circuit 252 converts paired voltages, conceptually represented by IMS cells (coarse) 217. In the second mode, one or more results from operation in the first mode, coarse search results 215, are converted to voltages suitable for performing a fine search function using memory strings in the fine portion 222. In some variations, in the second mode, circuits not used in the second mode are not activated (e.g., bit lines in the coarse search portion are not activated in the second mode).
[0058] In some variations and / or usage scenarios, during a first phase of operation, the course 210 functions as a TCAM that enables an exact match, and at most one sense amplifier of the sense amplifier circuit 214 detects a change corresponding to the exact match. In a second phase of operation (responsive to detecting the exact match), the fine 220 is operated once. The word line circuit 252 drives a single one of the word lines of the array 250 (corresponding to the exact match represented on the course search results 215) for reading (e.g., provides a voltage Vread). The word line circuit 252 activates the remaining word lines of the array 250 for passing (e.g., provides a voltage Vpass).
[0059] The coarse 210 and / or the fine 220 can be independently implemented according to various features. The features include the match type (e.g., exact and / or approximate), the number of matches (zero, one, and / or multiple), and the order of the coarse matches (parallel or sequential). Exemplary match types include exact and approximate matches. An exact match corresponds to a full bit match, e.g., all bits of the search parameter match all bits of the value stored in the memory string. An approximate match corresponds to less than all bit matches, e.g., all bits except one bit match, all bits except two bits match, or alternatively a percentage of bit matches, e.g., 90% bit match (e.g., 9 out of 10 bits). For example, the coarse 210 can perform an exact match and the fine 220 can perform an approximate match. In another example, both the coarse 210 and the fine 220 can perform an exact match (or can perform an approximate match). Exemplary numbers of matches include at most one match and at most multiple matches. For example, the coarse 210 can detect at most one match (either exact or approximate) and the fine 220 can detect at most one match (either exact or approximate). In another example, the coarse 210 can detect at most multiple matches (either exact or approximate) and the fine 220 can detect at most multiple matches (either exact or approximate). An example of a coarse match order includes parallel processing of all matches (exact or approximate), such as a single use of the fine portion 222. Another example of a coarse match order includes serial processing of each match (exact or approximate) in a respective single use of the fine portion 222. For example, if there are two coarse matches, then in response the fine portion 222 is used twice in succession, once for each of the coarse matches. Below is a description of some exemplary combinations of match types, numbers of matches, and coarse match orders.
[0060] In some similar variations and / or usage scenarios, the multiple sense amplifiers of the sense amplifier circuit 214 detect a corresponding multiple exact matches. According to one technique, in response to the multiple matches, the second phase of the operation is repeated successively for each exact match, and the counting as performed by the post-processing (counting) 225 is performed over the successive iterations. For example, each of the first, second, and third sense amplifiers of the sense amplifier circuit 214 detects an exact match. In response, the second phase of the operation is repeated three times (i.e., three iterations including the first iteration, the second iteration, and the third iteration), and the counting is performed over the three iterations. The first iteration starts counting. The second phase of the operation is performed a first time using the match detected by the first sense amplifier that detects the match in the first phase of the operation. The post-processing (counting) 225 starts counting using the result of the second phase of the operation as performed in the first iteration to start counting. The second iteration continues counting. The second phase of the operation is performed a second time using the match detected by the second sense amplifier that detects the match in the first phase of the operation. The post-processing (counting) 225 continues the counting using the results of the second phase of the operation as performed in the second iteration to continue the counting by counting the results of the second first iteration using the results counted before the results of the first iteration. The third iteration completes the counting. The second phase of the operation is performed a third time using the match detected by the third sense amplifier that detects the match in the first phase of the operation. The post-processing (counting) 225 completes the counting using the results of the second phase of the operation as performed in the third iteration to complete the counting by counting the results of the third iteration using the results counted before the first and second iterations. According to another technique, in response to multiple matches, the second phase of the operation is performed once and multiple word lines of the array 250 are activated (corresponding to multiple exact matches). For example, each of the first, second, and third sense amplifiers of the sense amplifier circuit 214 detects an exact match.In response, the second phase of the operation is performed once. Post-processing (counting) 225 performs the counting, using the results of the second phase of the operation to perform the counting.
[0061] In some variations and / or usage scenarios, the course 210 can operate as an exact and approximate match TCAM. During a first phase of operation, the course 210 operates as a TCAM that allows exact and approximate matches. Thus, zero or more of the sense amplifiers of the sense amplifier circuit 214 detect a match. In response, the second phase of operation is repeated successively for each match, and counting as performed by post-processing (counting) 225 is performed over successive iterations. For one approximate match technique, the sense amplifiers of the sense amplifier circuit 214 can detect changes according to more than one threshold (e.g., in series through different sensing periods or in parallel through parallel operation of multiple detection mechanisms). For example, one threshold corresponds to an exact match and another threshold corresponds to an approximate match (e.g., all matches except one or two bits). Each amount of approximate matching can be implemented through a respective threshold (e.g., through a respective sensing period). For example, a first amount of approximate matching (corresponding to a mismatch of 2 or fewer bits) is implemented via a first threshold (e.g., via a first sensing period), and a second amount of approximate matching (corresponding to a mismatch of 4 or fewer bits) is implemented via a second threshold (e.g., via a second sensing period that is greater than the first sensing period).
[0062] In some variations and / or usage scenarios, the fine 220 can operate to match the stored information exactly and approximately against the search parameters according to the word line value. During the second phase of operation, the fine 220 operates to detect exact and approximate matches, so that zero or more of the sense amplifiers of the sense amplifier circuit 224 detect a match. Each detection indicates whether the respective match is exact or approximate. Furthermore, techniques such as those described with respect to the approximate match for the course 210 can be applied to the fine 220. The post-processing (count) 225 can count exact and approximate matches accordingly. For example, exact matches count a value greater than approximate matches.
[0063] The dimensions of the array 250 and the allocation between the coarse portion 212 and the fine portion 222 vary according to the usage. For example, the coarse portion 212 has N memory strings and the fine portion 222 has M memory strings, so that the array 250 has a total of N+M memory strings. Each of the memory strings has L memory devices (consecutive) in each of the memory strings for a total of L×(N+M) memory devices in the array 250. There are L word lines of the word line circuit 252 coupled to the array 250. Each of the word lines conceptually couples (in parallel) N+M memory devices across the entire “width” of the “array 250”. Each of the N memory strings of the coarse portion 212 is coupled to a respective bit line circuit of the bit line circuit 211 and a respective sense amplifier of the sense amplifier circuit 214. Thus, there are N bit line circuits in the bit line circuit 211 and N sense amplifiers in the sense amplifier circuit 214. Similarly, each of the M memory strings in fine portion 222 is coupled to a respective bit line circuit in bit line circuit 221 and a respective sense amplifier in sense amplifier circuit 224. Thus, there are M bit line circuits in bit line circuit 221 and M sense amplifiers in sense amplifier circuit 224. As a specific example, N=48, M=128K-48, and L=48. As another specific example, N=232, M=128K-232, and L=232.
[0064] In some variations, the course search results 215 are selectively provided elsewhere, for example, via an interface circuit to a host.
[0065] The division of functionality between sense amplifier circuitry 224 and post-processing (counting) 225 is arbitrary. For example, variations are envisioned that have some or all of the functionality of post-processing (counting) 225 incorporated into sense amplifier circuitry 224, and vice versa.
[0066] An example implementation of the IMS cell (course) 217 is described with respect to Figures 6A and 6B. An example implementation of the IMS cell (fine) 227 is described with respect to Figures 6C and 6D. In some variations, the pair of devices in the IMS cell (course) 217 and / or the IMS cell (fine) 227 are implemented using a pair of SONOS devices. Computational Memory - IMS => IMC
[0067] Continuing with the second example of a computational memory (an IMS operation followed by an IMC operation), similar to the first example, a single computational memory array has two portions (one for an IMS operation and one for an IMC operation) each capable of performing a coarse search followed by a computation depending on the result of the coarse search (e.g., operating as a TCAM). Conceptually, the coarse search portion implements a filter (e.g., via the word line inputs) to reduce the computation performed by the computation portion (e.g., via the bit line inputs).
[0068] As an example of operation, the coarse search portion is similar to the first example and is used in a similar manner. Thus, the coarse search portion is operated as a TCAM and the example results are one-hot encoded. Furthermore, matches from the coarse search portion are then used to generate (calculate) values for word lines of the array. At the same time, the bit lines of the computation portion are manipulated according to the computation, such as being manipulated according to each term / bit of the first vector input of the dot product. The stored information in the memory cells of the computation portion corresponds to each term / bit of the second input vector of the dot product.
[0069] As a specific example, the TCAM result corresponds to a matched category, the first vector input corresponds to a neuron value, and the second vector input corresponds to a weight to apply to the neuron value. As another specific example, the bitlines of the computation portion are driven to (e.g., analog) voltages representing the first vector input (each bitline voltage corresponds to an element of the first vector). The threshold voltages of the memory cells are programmed to voltages representing the second input vector (each memory cell threshold voltage corresponds to an element of the second vector). Summing (optionally scaled) the currents across the bitlines serves to accumulate individual products (bitline voltages multiplied by memory cell currents determined in part by their threshold voltages) into the dot product result of the first and second input vectors. One form of scaling is according to a power of two, such as to effectively encode a binary value. Another form of scaling is according to relative importance, such as scaling according to the significance of the features. Some variations produce a single sum across all bitlines of the computation portion. Another variation generates multiple sums corresponding to different groups of bit lines in the computation portion.
[0070] The sensing in the computational portion is multi-valued and encoded via either digital or analog techniques. For example, in analog techniques, the sense amplifiers in the computational portion respond to bit line changes and generate analog outputs, which are then optionally scaled via analog techniques. The scaled analog outputs are then combined, for example, via analog summation, to generate a computation result. In the context of the specific example, the computation result corresponds to a dot product result. The computation result is optionally converted to a binary representation.
[0071] FIG. 3 illustrates an exemplary computational memory as computational memory 300 for performing a course search and subsequent calculations responsive to the results of the course search.
[0072] 2, computational memory 300 includes two portions: a course 310 for performing a coarse search, and a computation 320 for performing a calculation. The course portion 312 and the computation portion 322 are portions of the same memory array, array 350. A word line circuit 352 can provide the results of the coarse search to the computation.
[0073] Some elements in FIG. 3 are identical, substantially similar, or similar to corresponding elements in FIG. 2, and reference should be made to the corresponding elements for further description. For example, the sense amplifier circuit 314 includes a sense circuit capable of determining the current and / or voltage change of each of the sense amplifier terminals of the course portion 312, like the sense amplifier circuit 214. In another example, the IMS cell (course) 317 and the key information 318 correspond to the IMS cell (course) 217 and the key information 218, respectively. As a final example, the bit line circuit 311 includes a bit line circuit capable of manipulating each of the bit line terminals according to a predetermined value, like the bit line circuit 211. Furthermore, the array 350 corresponds to the array 250, the course 310 corresponds to the course 210, and the calculation 320 generally corresponds to the fine 220 (post-processing (scaling and summing) 325, which has a different function than the post-processing (counting) 225, etc.). The elements in FIG. 3 correspond to the elements in FIG. 2 identified by the last two digits of the element identifier. Therefore, the following description focuses on elements of FIG. 3 that differ in operation from, or are not present in, the corresponding elements of FIG.
[0074] The computational memory 200 is intended for a coarse search and then a fine search. In contrast, the computational memory 300 is intended for a coarse search and then a computation. Considering the coarse search, in some variations, the coarse 310 operates identically to the coarse 210. Considering the computation for the fine search, the computation 320 operates differently than the fine 220. The bit line circuit 321 operates the memory strings of the computation portion 322 individually, rather than in pairs as illustrated by the bit line circuit 221 (IMS cell (fine) 227) for weights 328. The bit line circuit 321 can optionally drive the memory strings according to analog values. The computation portion 322 can optionally store analog values. The sense amplifier circuit 324 and post-processing (scaling and summing) 325 can optionally operate with analog values, etc.
[0075] More specifically, the bit line circuits 321 can operate the bit lines independently of each other for their respective voltages and / or currents, such as via conditional precharge techniques or continuously provided techniques. The sense amplifiers of the sense amplifier circuits 324 can represent more than two values, either via digital (e.g., multiple bits on respective multiple terminals) or analog (e.g., different currents on a single terminal) techniques. Post-processing (scaling and summing) 325 can optionally scale the sense amplifier results (voltages or currents) according to one or more scale factors. For example, a respective scale factor is applied to a respective group of the sense amplifier outputs of the sense amplifier circuits 324. One form of scaling is via analog techniques. Another form of scaling is via digital techniques. Post-processing (scaling and summing) 325 can further combine the optionally scaled sense amplifier outputs, for example, via analog and / or digital summing. One or more outputs of the post-processing (scaling and summing) 325 are presented as one or more results via the computation results 326. In some variations, the operations of sense amplifier circuitry 324 and post processing (scaling and summing) 325 are combined, with the combining occurring conceptually before some or all of the sensing (e.g., via analog summation), and a reduced number of sense amplifiers are used. In some variations, the output of computation result 326 is analog, while in other variations the output is digital.
[0076] Like fine portion 222, the current and / or voltage changes in the memory strings in computation portion 322 result from selective discharging of conditionally pre-charged memory strings based, for example, on the stored values of the serially coupled memory devices of the memory strings and the values provided to the word lines by word line circuitry 352.
[0077] Thus, the calculation result 326 includes information based on the course search parameters 319, information stored in the course portion 312, information based on the calculation parameters 329, information stored in the calculation portion 322, and summation performed by post processing (scaling and summation) 325.
[0078] Like the word line circuit 252, the word line circuit 352 can conditionally operate in two modes during respective first and second phases of operation. The first mode of operation is identical to the word line circuit 252. In summary, one or more search parameters (coarse search parameters 319) are pairwise converted (as conceptually represented by IMS cells (coarse) 317) to voltages suitable for performing a coarse search function using memory strings of the coarse portion 312 (as conceptually represented by key information 318). In the second mode, one or more results from the operation in the first mode, the coarse search results 315, are converted to voltages suitable for performing a computation function using memory strings of the computation portion 322 (rather than a fine search function as in the fine portion 222). For example, the word line circuit 352 drives zero or more of the word lines of the array 350 represented by the coarse search results 315 (corresponding to an exact match) for reading (e.g., providing a voltage Vread). Wordline circuitry 352 drives the remaining wordlines of array 350 to pass (eg, provides a voltage Vpass).
[0079] Examples of memory cells used as weights are shown as weights 328. In some variations, each memory cell used as a weight is implemented with a respective SONOS device and each weight value is implemented as a respective threshold voltage. In some variations, each memory cell used as a weight is implemented using SLC, MLC, TLC, QLC, or analog techniques.
[0080] An example of a row of memory cells used as a category is shown as category 327 .
[0081] In some variations and / or usage scenarios, a portion of the computation portion 322 is used according to each feature, and summation in post-processing (scaling and summing) 325 across the computation portion 322 produces an overall score across the features. The overall score can be used as a classification result, such as to rank the highest scoring instances. In some variations, the computation 320 functions to perform a MAC process to calculate a likelihood score between the query and a selected instance in the computation portion 322. The selected instance is determined by the course 310 and functions to improve efficiency. Computational Memory - IMS => Access
[0082] Continuing with the third example of a computational memory (an IMS operation followed by an access operation), similar to the first example, a single computational memory array has two portions (one for an IMS operation and one for an access operation) each capable of performing a coarse search (e.g., operating as a TCAM) followed by an access operation depending on the result of the coarse search. Conceptually, the coarse search portion implements a filter (e.g., via a word line input) to reduce the access cost performed by the access portion (e.g., providing one or more result values).
[0083] As an example of operation, the coarse search portion is similar to the first example and is used in a similar manner. Thus, the coarse search portion is operated as a TCAM and the example results are one-hot encoded. Furthermore, matches from the coarse search portion are then used to generate values for word lines of the array (access operations). At the same time, the bit lines of the access operation portion are operated to enable reading stored values from memory cells of the access operation portion as selected by the word lines. Sense amplifiers of the access operation portion detect the stored values from the selected memory cells, providing, for example, a page of data.
[0084] As a specific example, the stored values in the coarse search portion correspond to key values, and the (optionally masked) input key is matched against the stored values in the access operation portion corresponding to each value. The coarse search portion is used such that at most one match is found, and only one word line is activated during the access operation. Thus, the computational memory is operable as a key-value store.
[0085] FIG. 4 illustrates an exemplary computational memory as computational memory 400 for performing a coarse search followed by an access operation responsive to the results of the coarse search.
[0086] 2, the computational memory 400 includes two portions: a coarse portion 410 for performing a coarse search, and an access portion 420 for performing an access operation. The coarse portion 412 and the access portion 422 are portions of the same memory array 450. A word line circuit 452 is capable of providing the results of the coarse search to the access operation.
[0087] Some elements in FIG. 4 are identical, substantially similar, or similar to corresponding elements in FIG. 2, and reference should be made to the corresponding elements for further description. For example, IMS Cells (Course) 417 and Key Information 418 correspond to IMS Cells (Course) 217 and Key Information 218, respectively. Furthermore, Array 450 corresponds to Array 250, Course 410 corresponds to Course 210, Access 420 generally corresponds to Fine 220 (Bitline Circuit 421 has no input corresponding to Fine Search Parameters 229, Post Processing (Count) 225 has no corresponding element in Fine 220), and so on. As in FIG. 3, elements in FIG. 4 correspond to elements in FIG. 2 identified by the last two digits of the element identifier. Thus, the following description focuses on elements in FIG. 4 that operate differently than their corresponding elements in FIG. 2.
[0088] Similar to wordline circuitry 252, wordline circuitry 452 drives (e.g., provides voltage Vread) a single one of the wordlines of array 450 for a read (corresponding to the exact match represented in coarse search results 415). Wordline circuitry 452 activates (e.g., provides voltage Vpass) the remaining wordlines of array 450 for passing.
[0089] Bitline circuitry 421 corresponds in configuration to bitline circuitry 221, but differs in that it operates the bitline terminals of access portion 422 according to predetermined values, e.g., includes a precharger for each of the bitline terminals (similar to bitline circuitry 211). The prechargers operate (in conjunction with other circuits, e.g., wordline circuitry 452 and / or sense amplifier circuitry 424) to precharge the memory strings of access portion 422.
[0090] Sense amplifier circuit 424 corresponds to sense amplifier circuit 224. However, rather than providing the results to another element for further processing, the results are provided directly (e.g., in a binary encoding) as access results 426. The results correspond to a selected page of data, an example of which is shown as page data 427.
[0091] Thus, the access results 426 include information based on the course search parameters 419 as well as information stored in the course portion 412 and the access portion 422 .
[0092] In some variations of the computational memory described above, the word lines are continuous across the two portions of the array. For example, the word lines (e.g., WL1...WL48) are continuous across the coarse portion 212 and the fine portion 222 (FIG. 2). In other examples, the word lines are continuous across the coarse portion 312 and the computation portion 322 (FIG. 3), and across the coarse portion 412 and the access portion 422 (FIG. 4). In some variations, the word lines of the two portions are driven separately. For example, the word lines are not continuous between the two portions.
[0093] In some variations of the computational memory, the bit line circuitry and / or the sense amplifier circuitry are partitioned between two portions of the array. For example, bit line circuitry 211 and bit line circuitry 221 are two segments of the bit line circuitry that span the coarse portion 212 and the fine portion 222 (FIG. 2). In another example, bit line circuitry 311 and bit line circuitry 321 are two segments of the bit line circuitry that span the coarse portion 312 and the computation portion 322 (FIG. 3), and bit line circuitry 411 and bit line circuitry 421 are two segments of the bit line circuitry that span the coarse portion 412 and the access portion 422. In a similar example, sense amplifier circuitry 214 and sense amplifier circuitry 224 are two segments of the sense amplifier circuitry that span the coarse portion 212 and the fine portion 222 (FIG. 2). In another similar example, the sense amplifier circuit 314 and the sense amplifier circuit 324 are two segments of the sense amplifier circuit that span the course portion 312 and the calculation portion 322 (FIG. 3), and the sense amplifier circuit 414 and the sense amplifier circuit 424 are two segments of the sense amplifier circuit that span the course portion 412 and the access portion 422. In some examples of the partitioned sense amplifier circuit, the sense amplifier post-processing is included in the segment of the sense amplifier circuit. As an example, the sense amplifier circuit 214 is included in the first segment of the sense amplifier circuit, and the sense amplifier circuit 224 is included in the second segment of the sense amplifier circuit (FIG. 2) together with the post-processing (counting) 225. As another example, the sense amplifier circuit 314 is included in the first segment of the sense amplifier circuit, and the sense amplifier circuit 324 is included in the second segment of the sense amplifier circuit (FIG. 3) together with the post-processing (scaling and summing) 325.
[0094] In some variations of the above computational memories, the bit lines are precharged, for example, according to a predetermined value or according to a dynamic value depending on whether the bit lines are being used as inputs. In other variations, the bit lines are not precharged but are driven according to a predetermined or dynamic value.
[0095] In some variations of the above computational memories, each memory cell can store one or more symbols, for example, via binary or analog techniques. For example, a memory cell can store a single symbol as an analog value, such as a device threshold voltage. As a specific example, programming the threshold voltage of a SONOS cell can be used for (e.g., analog) data storage. In another example, a memory cell can store a single symbol, interpreted as, for example, a single bit of information, using SLC techniques. In other examples, a memory cell can store multiple symbols using MLC, TLC, or QLC techniques.
[0096] In some variations of the computational memory described above, a page buffer (e.g., implemented as a page buffer circuit) and / or a page cache (e.g., implemented as a page cache circuit) are used to store the results. For example, the fine search results 226 are associated with the fine portion 222 (FIG. 2) and stored in a page buffer. In another example, the calculation results 326 are associated with the calculation portion 322 (FIG. 3) and stored in a page buffer, and the access results 426 are associated with the access portion 422 (FIG. 4) and stored in a page buffer and / or a page cache. In another example, the access results 426 are stored in a page cache circuit coupled to the sense amplifier circuit 424. The page cache circuit can cache pages of results read from the access portion 422. The pages of results correspond to information stored in the page data 427, for example. Computational Memory Devices
[0097] Some elements (e.g., integrated circuits, one or more dies of a system-on-chip, and / or a packaging die) include one or more computational memories (such as computational memory 200 of FIG. 2, computational memory 300 of FIG. 3, and / or computational memory 400 of FIG. 4) along with one or more optional hardware circuit blocks for implementing memory components that can be used as standalone components and / or as components in a system. For example, the hardware circuitry enables the use of the computational memory in a computational SSD (such as computational SSD 700 of FIG. 7 and / or computational SSD 840 of FIG. 8 as a configurable computational memory device 765). In another example, the optional hardware circuit blocks enable the use of the computational memory as a component of a system-on-chip.
[0098] Exemplary optional hardware circuit blocks are an I / O interface, a controller, and an analog source.
[0099] The I / O interface allows for interfacing to an agent external to the computational memory (such as a host) to use the computational memory. Exemplary uses are to configure an array of the computational memory (such as configure array 501 of FIG. 5), to program an array of the computational memory (such as program array 502 of FIG. 5), to instruct the computational memory to perform a first function and optionally receive a result of the first function (such as first function 503 of FIG. 5), and to instruct the computational memory to perform a second function and optionally receive a result of the second function (such as second function 504 of FIG. 5).
[0100] The controller may provide overall control of operation of and among various hardware circuit blocks of one or more computational memories. For example, the controller coordinates operation of I / O interfaces to and from portions of a memory array of a computational memory. In another example, the controller coordinates operation of analog sources with reference to configuration information. In another example, the controller optionally coordinates an order of a second function depending on a first function.
[0101] The analog sources may generate and / or provide one or more analog outputs to, for example, a memory array, bit line circuitry, sense amplifier circuitry, and / or word line circuitry of a computational memory. The analog outputs may comprise zero or more voltage and / or current sources (e.g., reference sources), such as those generated by one or more bias circuits. Computational Memory Operations
[0102] FIG. 5 illustrates an example flow diagram for a technique for operating computational memory, such as with the computational SSD techniques shown or referenced in any of the other figures.
[0103] The flow proceeds with four actions: Configure Array 501, Program Array 502, First Function 503, and Second Function 504. The flow is for one of three types of dedicated arrays, and for arrays that are configurable according to any one or more of the three types of dedicated arrays. The flow for each type of dedicated array is described, and then the flow for configurable arrays.
[0104] The flow for any of the three types of dedicated arrays begins with the array's program 502. Consider an array (such as array 250 in FIG. 2) that is dedicated to a first IMS (a coarse search) followed by a second IMS (a fine search) responsive to the first search. The details of the flow are illustrated by the actions Coarse 519 and Fine 529.
[0105] The flow begins with programming the array 502 by storing values into respective portions of the dedicated arrays. The coarse search portion of the array is programmed with values that will be searched for a match using the coarse search portion of the array (program coarse array 512). The fine search portion of the array is programmed with values to store and match search information provided via the bit lines of the fine search portion of the array (program fine array 522).
[0106] The flow then proceeds to a first function 503. The first of the two functions (corresponding to a coarse search) is performed using a coarse search portion of the array. The word lines of the array are driven and the bit lines of the coarse search portion of the array are manipulated and sensed (Drive WL, Manipulate BL, Sense 513). More specifically, the word lines of the array are driven according to coarse search information (e.g., key information and optionally a mask). The bit lines of the coarse search portion of the array are manipulated according to predetermined values that enable the coarse search portion of the array to operate as a readable memory with respect to the word lines and values programmed into the coarse search portion of the array. The bit lines of the coarse search portion of the array are sensed to determine a match between the word lines and the values programmed as a result of the first function. The results of the first function (coarse search) are provided for use by the second function.
[0107] The flow then proceeds to the second function 504. The second of the two functions (fine search) is performed using the fine search portion of the array. The word lines of the array are driven and the bit lines of the fine search portion of the array are manipulated, sensed and counted (Drive WL; Manipulate BL; Sense and Count 524). More specifically, the word lines are driven according to the results of the coarse search. The bit lines of the fine search portion of the array are manipulated according to their respective dynamic values of the stored values of the fine search portion of the array, which serve as search parameters for which the fine search portion of the array searches (in parallel). The bit lines of the fine search portion of the array are sensed to determine matches, which are then counted (e.g., in one or more groups) to determine the overall best match or matches as the result of the second function. The results of the second function (fine search) are then available for use by other agents. Thus, the flow for the array dedicated to the first IMS followed by the second IMS is completed.
[0108] Next, consider an array (such as array 350 in FIG. 3) that is specific to an IMS (course search) followed by an IMC (calculation) that depends on the IMS. The details of the flow are shown by the actions Course 519 and Calculation 539.
[0109] For arrays dedicated to the first and second IMS searches, the flow begins with programming the arrays 502 by storing values in the respective portions of the dedicated arrays. The coarse search portion of the array is programmed with values to be searched for matches using the coarse search portion of the array (program the coarse array 512). The computation portion of the array is programmed with values to store and compute with computation parameters provided via the bit lines of the computation portion of the array (program the computation array 532).
[0110] For arrays dedicated to first and second IMS searches, flow then proceeds to a first function 503 which performs the first function by driving the word lines and steering and sensing the bit lines of the coarse search portion of the array (Drive WL; Steering BL; Sensing 513). The results of the first function (the coarse search) are provided for use by the second function.
[0111] For arrays dedicated to the first and second IMS searches, the flow then proceeds to a second function 504. The second of the two functions (computation) is performed using the computational portion of the array. The word lines of the array are driven and the bit lines of the computational portion of the array are manipulated, sensed and summed (Drive WL; Manipulate BL; Sense and Add 534). More specifically, the word lines of the array are driven according to the results of the course search. The bit lines of the computational portion of the array are manipulated according to respective dynamic values provided as computational parameters that the computational portion of the array uses as computational operands (in parallel) with the stored values of the computational portion of the array. The bit lines of the computational portion of the array are sensed to determine respective intermediate results (e.g., respective analog currents or voltages). The intermediate results are then combined (e.g., summed using analog techniques) to generate one or more results of the second function. The results are optionally converted to a digital representation. In some variations, the sensing and summing are performed together. The results of the second function (computation) are then available for use by other agents. Thus, the flow for an IMS specific array followed by an IMC is complete.
[0112] Next, consider an array (such as array 450 in FIG. 4) specific to an IMS (course search) followed by an access operation according to the IMS. The details of the flow are shown by the actions Course 519 and Access 549.
[0113] For arrays dedicated to the first and second IMS searches, flow begins with programming the arrays 502 by storing values in the respective portions of the dedicated arrays. The coarse search portion of the array is programmed with values that will be searched for matches using the coarse search portion of the array (program coarse array 512). The access operation portion of the array is programmed with values to store (program access array 542) that will be available for reading by subsequent access operations.
[0114] For arrays dedicated to first and second IMS searches, flow then proceeds to a first function 503 which performs the first function by driving the word lines and steering and sensing the bit lines of the coarse search portion of the array (Drive WL; Steering BL; Sensing 513). The results of the first function (the coarse search) are provided for use by the second function.
[0115] For arrays dedicated to the first and second IMS searches, the flow then proceeds to a second function 504. The second of the two functions (the access operation) is performed using the access operation portion of the array. The word lines of the array are driven and the bit lines of the access operation portion of the array are manipulated and sensed (Drive WL; Manipulate BL; Sense 544). More specifically, the word lines are driven according to the results of the coarse search. The bit lines of the access operation portion of the array are manipulated according to predefined values that enable the access operation portion of the array to operate as a readable memory with respect to the word lines, and values programmed in the fine search portion of the array. The bit lines of the access operation portion of the array are sensed to determine respective digital results that collectively form one or more results of the second function. The results of the second function (the access operation) are then available for use by other agents. Thus, the flow is completed for arrays dedicated to IMS followed by an access operation.
[0116] The configurable array flow begins with an action of configure array 501. A configurable array (such as a configurable combination of elements from Figures 2, 3, and / or 4) is configured according to a specialized allocation for each of the respective types of operations, i.e., fine search, computation, and access operations.
[0117] Flow then proceeds individually (optionally in parallel) for each assignment according to the configuration as indicated by configurable 550. Thus, for each assignment configured as a first IMS (coarse search) followed by a second IMS (fine search) responsive to the first IMS search, flow proceeds according to course 519 and fine 529. For each assignment configured as a first IMS (coarse search) followed by an IMC (calculation) responsive to the IMS search, flow proceeds according to course 519 and calculation 539. For each assignment configured as a first IMS (coarse search) followed by an access operation responsive to the IMS, flow proceeds according to course 519 and access 549. Exemplary Wordline and Bitline Based Search
[0118] Figures 6A and 6B collectively illustrate a search through information provided on a word line, and Figures 6C and 6D collectively illustrate a search through information provided on a bit line.
[0119] Referring to Figures 6A, 6B and the search through the word lines from the beginning, the course portion 612 and the IMS cell (course) 617 correspond to the course portion 212 and the IMS cell (course) 217 of Figure 2, respectively (although Figure 6A omits some details shown in Figure 2). T1 671 and T2 672 are consecutive on BL1 693 and are coupled to the word lines WL1 691 and WL2 692, respectively. Figure 6B shows an example of threshold voltages for programming T1 671 and T2 672 and search voltages for driving WL1 691 and WL2 692, which allow matching search information to stored information.
[0120] The threshold voltages for T1 671 and T2 672 shown in FIG. 6B specify encodings that allow for matching on only zeros ("0"), only ones ("1"), anything ("X"), and nothing ("void"). The search voltages for WL1 691 and WL2 692 shown in FIG. 6B specify encodings that allow for searching on only zeros "0", only ones ("1"), anything / wildcard ("WC"), and nothing ("void"). As one specific example, for matching on zeros, T1 671 is programmed to a threshold voltage of Vt.High and T2 672 is programmed to a threshold voltage of Vt.Low. To search for zeros, WL1 691 is driven to V.High2 and WL2 692 is driven to V.High1. The last row of the table in FIG. 6B shows an exemplary relationship between two threshold voltages Vt.Low and Vt.High and two search voltages V.High1 and V.High2. Specifically, the two threshold voltages (Vt.Low and Vt.High) are lower than either of the two search voltages (V.High1 and V.High2). In other words, the two search voltages are higher than either of the two threshold voltages. Additionally, Vt.Low is a voltage lower than Vt.High, and V.High1 is a voltage lower than V.High2. As a specific example, Vt.Low (the lowest voltage) is −1V, Vt.High is 3V, V.High1 is 4V, and V.High2 (the highest voltage) is 5V.
[0121] During operation, a pair of memory devices in a course portion of a memory array (e.g., memory devices BL1, ... BL48 in course portion 212 of FIG. 2) are programmed with data to search according to FIG. 6B. Programming is performed, for example, as part of program 512 of a course array of FIG. 5. A search is then performed according to driving a pair of word lines (e.g., WL1, WL2, ... WL47, WL48 of FIG. 2) according to FIG. 6B. The search is performed as part of drive WL; operate BL; sense 513 of FIG. 5.
[0122] Referring now to Figures 6C, 6D, and the search through bit lines, fine portion 622 and IMS cell (fine) 627 correspond to fine portion 222 and IMS cell (fine) 227 of Figure 2, respectively (although Figure 6C omits some details shown in Figure 2). T1 681 and T2 682 are coupled to respective bit lines BL128K-1698 and BL128K699, and share word lines. Figure 6D shows an example of threshold voltages for programming T1 681 and T2 682, and search voltages for driving BL128K-1698 and BL128K699, that allow the search information to match against the stored information.
[0123] Similar to FIG. 6B, the threshold voltages for T1 681 and T2 682 shown in FIG. 6D specify encodings to allow matching for zeros only ("0"), ones only ("1"), anything ("X"), and nothing ("void"). The search voltages for BL128K-1 698 and BL128K 699 shown in FIG. 6D specify encodings to allow searching for zeros only "0", ones only ("1"), anything / wildcard ("WC"), and nothing ("void"). As one specific example, for matching for zeros, T1 681 is programmed to a threshold voltage of Vt.High and T2 682 is programmed to a threshold voltage of Vt.Low. To search for zeros, BL128K-1 698 is driven to V.High and BL128K 699 is driven to V.Low. The last row of the table in FIG. 6D shows an exemplary relationship between the threshold voltages Vt.Low and Vt.High and the search voltages V.Low and V.High. Specifically, two threshold voltages (Vt.Low and Vt.High) are mixed with respect to the voltage levels between the two search voltages (V.Low and V.High). V.Low is the lowest voltage, Vt.Low is the next highest voltage, V.High is the next highest voltage, and Vt.High is the highest voltage. As a specific example, V.Low is 0V, Vt.Low is 1V, V.High is 2V, and Vt.High is 3V.
[0124] During operation, pairs of memory devices in the fine portion of the memory array (e.g., memory devices BL49, BL50, BL128K-1, BL128K of FIG. 2) are programmed with data to search, according to FIG. 6D. Programming is performed, for example, as part of program fine array 522 of FIG. 5. A search is then performed, according to FIG. 6D, in accordance with the driving pairs of bit lines (e.g., BL128K-1 and BL128K of FIG. 2). The search is performed, for example, as part of drive WL; drive BL; sense and count 524 of FIG. 5.
[0125] As an example of a word line encoding (such as in FIG. 6B), the search information is a binary key value, allowing only searches for zeros and ones. As another example, the search information is a binary key value together with a binary mask value of equal length to the binary key value. Each bit in the binary mask value indicates whether the corresponding bit in the binary key value is ignored (e.g., match anything) or matches according to the corresponding binary key value bit. As yet another example, building on the previous example, another binary mask value is further used to indicate whether the corresponding bit in the binary key value is considered invalid (e.g., match nothing). For example, a 4-bit key value is 0b0111 and the corresponding mask value is 0b0001, resulting in a search for 0b011x, which matches against stored values of 0b110 or 0b0111. Similar examples of bitline encoding (e.g., as in FIG. 6D) use a search value alone, a search value in conjunction with a mask value, and / or a search value in conjunction with multiple mask values. For example, a 4-bit search (e.g., key) value might be 0b1010 and the corresponding mask value might be 0b1000, resulting in a search for 0bx010, which would result in a match against a stored value of 0b1010 or 0b0010. Computational SSDs
[0126] Computational SSD resources can be used in a variety of systems (e.g., systems that use smart SSDs), such as compute servers, database servers, IMS servers, IMC servers, and AI accelerators.
[0127] FIG. 7 illustrates an exemplary computational SSD as computational SSD 700.
[0128] Computational SSD 700 includes at least one instance of a computational memory device, such as at least one of computational memory device 762, computational memory device 763, computational memory device 764, or configurable computational memory device 765. It should be noted that each instance of coarse search 710, fine search 720, calculate 730, and access 740 may be similar, substantially similar, or identical to one another, depending on the configuration.
[0129] The computational memory device 762 is intended to implement a fine search capability and includes a memory array operable as two parts to implement a coarse search 710 that provides information to a second function 752 for implementing a fine search 720. The computational memory device 762 can be used, for example, in systems (such as database servers) intended to provide large scale search capabilities.
[0130] The computational memory device 763 includes a memory array operable as two parts to implement a course search 710 that provides information to a second function 753 for implementing a calculation 730 for the purpose of implementing a computational capability. The computational memory device 763 can be used, for example, in a server (such as an AI system) for the purpose of providing large scale computational capability.
[0131] The computational memory device 764 is intended to implement the access capability and includes a memory array operable as two parts to implement a course search 710 that provides information to a second function 754 to implement the access 740. The computational memory device 764 can be used in a server intended to provide key-value information, such as, for example, a key-value SSD based system.
[0132] The configurable computational memory device 765 is intended for configurable computational SSD capabilities, such as usable in database servers, AI systems, and / or key-value SSD based systems, according to configuration. The configurable computational memory device 765 includes a memory array operable as two parts to implement a coarse search 710 that provides information to a configurable second function 755 to implement a configuration determined to be one of a fine search 720, a calculation 730, and an access 740.
[0133] Interface 780 interfaces the computational memory device to a host and / or bus via host / bus coupling 782.
[0134] Computational SSD 700 is an exemplary implementation of computational SSD 840 of FIG. 8, where host / bus coupling 782 serves to couple to bus / interface 880 of FIG.
[0135] The course search 710 is exemplified by the course 210 of FIG. 2, the course 310 of FIG. 3, and / or the course 410 of FIG.
[0136] Fine search 720 is exemplified by fine 220 in Figure 2. Calculate 730 is exemplified by calculate 320 in Figure 3. Access 740 is exemplified by access 420 in Figure 4.
[0137] Other compartmentalizations of elements, connections between elements, and capabilities and / or capacities of the elements shown in the figures, as well as additional elements, are contemplated according to usage requirements. Computational SSD Systems
[0138] 8 illustrates an exemplary hardware system having a computational SSD as computational SSD system 800. The system includes hardware blocks CPU 810, GPU 820, RAM 830, computational SSD 840, conventional SSD 850, HDD 860, and I / O 870 coupled by a hardware block bus / interface 880.
[0139] Computational SSD 840 includes one or more instances of Computational SSD 700 of FIG.
[0140] CPU 810 comprises one or more processing units, such as any combination of hardware units capable of executing programmed instructions, such as microprocessors, signal processors, AI processors, etc. One or more of the processing units optionally include one or more internal registers (some of which are, optionally, architecturally visible), one or more cache memories, and / or one or more internal memories (such as for buffering and / or coalescing), as represented by registers, caches, and internal memory 812.
[0141] GPU 820 comprises one or more processing units, such as any combination of units capable of accelerating processing for processes subject to a relatively high degree of parallelism, such as graphics processing, signal processing, and / or AI processing. Similar to the registers, caches, and internal memory 812 of CPU 810, one or more of the processing units optionally include one or more internal registers (some of which are optionally architecturally visible, represented by registers, caches, and internal memory 822), one or more cache memories, and / or one or more internal memories (such as for buffering and / or fusing).
[0142] The RAM 830 comprises one or more memory elements for storing instructions and / or data in a larger amount than storage internal to the CPU 810 and / or GPU 820. The RAM 830 may be implemented via volatile memory elements such as, for example, DRAM and / or SRAM.
[0143] Conventional SSD 850 optionally includes one or more storage elements, such as a flash-based storage element, for storing instructions and / or data that are accessible with reduced latency as compared to HDD 860.
[0144] HDD 860 optionally includes one or more storage elements, such as rotational-based magnetic and / or non-volatile optical storage elements (e.g., disks), for storing instructions and / or data in larger quantities than a conventional SSD 850 can store.
[0145] I / O 870 includes elements for interfacing any combination of CPU 810, GPU 820, RAM 830, computational SSD 840, conventional SSD 850, and / or HDD 860 with elements external to computational SSD system 800. Exemplary external elements include mass storage devices, local and wide area networks (such as the Internet), human interface components (such as a keyboard, mouse, and / or monitor), and other elements that provide capabilities to extend and / or enhance capabilities not otherwise provided by computational SSD system 800.
[0146] Bus interface 880 enables communication between elements coupled thereto (e.g., CPU 810, GPU 820, RAM 830, computational SSD 840, conventional SSD 850, HDD 860, and / or I / O 870). I / O 870 may comprise one or more serial and / or parallel communication channels, as well as optional protocol conversion and / or adaptation capabilities to facilitate communication between the elements coupled thereto.
[0147] Conceptually, the various storage elements in the computational SSD system 800 comprise a memory hierarchy, such as to reduce overall instruction and / or data access time, reduce overall cost of instruction and / or data storage, and / or reduce power consumption related to storage capabilities. For example, the registers, cache, and internal memory 812 of the CPU 810 and the registers, cache, and internal memory 822 of the GPU 820 collectively comprise the highest level of the memory hierarchy. The memory hierarchy further comprises (from top to bottom level) RAM 830, conventional SSD 850, HDD 860, and optionally storage accessible via I / O 870 (such as website information accessible via the Internet). In an example, the higher levels of the memory hierarchy have lower latency and lower capacity than the lower levels. For example, the registers, cache, and internal memory 812 and the registers, cache, and internal memory 822 have lower access latency and / or storage capacity than the RAM 830. On the other hand, the RAM 830 has lower access latency and / or storage capacity than the conventional SSD 850.
[0148] In a first variation of the example, computational SSD 840 includes a level of memory hierarchy between RAM 830 and conventional SSD 850 and / or HDD 860 to bridge the gap in access latency and / or storage capacity. In a second variation of the example, any one or more of registers, caches, and internal memory 812, registers, caches, and internal memory 822, RAM 830, conventional SSD 850, HDD 860, and / or storage accessible via I / O 870 are augmented with elements based on the techniques of computational SSD 840 to reduce access latency and / or increase storage capacity. In other examples, various combinations of SCM, NVDIMM, CXL memory, SSD, and / or DRAM-like storage are implemented using the techniques of computational SSD 840. In yet other examples, the computational SSD 840 can reduce the processing load (e.g., on the CPU 810 and / or GPU 820) and / or reduce traffic on the bus / interface 880 by performing processing locally.
[0149] Other compartmentalizations of elements, connections between elements, and capabilities and / or capacities of the elements shown in the figures, as well as additional elements, are contemplated according to usage requirements. More information on Computational SSDs
[0150] 2, 3, and 4 show techniques that may be applied to the architecture for computational SSDs. Other techniques may also be applied according to different memory technologies, etc.
[0151] Exemplary memory technologies that are applicable to the memory array of the computational SSD disclosed herein include floating gate, split gate, SONOS, floating dot, DRAM, DRAM-like (e.g., 2T0C), FeFET, and any memory technology compatible with searching through word lines and bit lines. Exemplary SONOS memory technology (sometimes referred to as charge trapping memory) uses an insulating layer (e.g., of silicon nitride) with traps to capture and hold charge when injected from the channel. Exemplary floating dot memory technology conceptually replaces the floating gate with a floating silicon nanodot or embeds floating silicon nanodots within a polysilicon gate. Exemplary 2T0C memory technology uses the parasitic capacitance of the lead transistor to store charge rather than an explicit storage capacitor. Exemplary FeFET memory technology uses the permanent electric field polarization of a ferroelectric material embedded between the gate and source-gate conducting regions to store information. Exemplary memory structures applicable to the memory array of a computational SSD include 2D structures (e.g., 2D flash structures) and 3D structures (e.g., 3D flash structures). Exemplary array architectures applicable to the memory array of a computational SSD include NOR / OR type array architectures and AND / NAND type array architectures.
[0152] It is understood that the foregoing disclosure presents implementations, variations, embodiments, and examples in an intended illustrative sense, and not a limiting sense. Modifications and combinations within the spirit of this disclosure and the scope of the following claims are contemplated as being discernible.
Claims
1. 1. A method for operating a memory system, comprising: first sensing a first portion of memory strings of a memory array according to (i) a first voltage driven on a word line of the memory array and (ii) a predetermined operating condition of the first portion of the memory strings; second sensing a second portion of the memory strings of the memory array according to (i) a second voltage driven on a word line of the memory array and (ii) a dynamic operating condition of the second portion of the memory strings; and providing a result of the second detecting step; wherein the second voltage is based on a result of the first sensing step. method.
2. The method of claim 1 , wherein the dynamic operating condition is generated by manipulating bit line circuit terminals of a second portion of the memory strings according to a dynamic value.
3. 3. The method of claim 2, wherein the second detecting step comprises counting approximate matches between a value stored in a second portion of the memory string and the dynamic value.
4. 3. The method of claim 2, wherein the second sensing step includes summing according to a current in a second portion of the memory string.
5. 5. The method of claim 4, wherein the dynamic values are analog values, the values stored in the second portion of the memory string are analog values, and the summing is according to an analog technique.
6. 6. The method of claim 5, wherein the second sensing step includes scaling the currents in a second portion of the memory strings and then summing.
7. 7. The method of claim 1, wherein the dynamic operating condition enables matching the dynamic value to any value stored in any of the second portions of the memory strings based on a dynamic value and a mask value.
8. 7. The method of claim 1, wherein the dynamic operating condition is based on a dynamic value and a mask value and makes it impossible to match the dynamic value to any value stored in any of the second portions of the memory strings.
9. 7. The method of claim 1, wherein the result of the second sensing step indicates a multiply-accumulate operation, a first operand of the multiply-accumulate operation is a dynamic value, the dynamic operating condition is determined according to the dynamic value, and a second operand of the multiply-accumulate operation corresponds to a value to be programmed into a second portion of the memory string.
10. 7. The method of claim 1, wherein the dynamic operating condition is determined according to a dynamic value, and the result of the second sensing step indicates a search operation for the dynamic value among values programmed into the second portion of the memory string.
11. a memory array including first and second portions of memory strings, where one or more of the memory strings in the first portion of memory strings and one or more of the memory strings in the second portion of memory strings each include a plurality of contiguously connected memory devices; a word line circuit having terminals coupled to respective control inputs of respective rows in which said memory strings are organized; a bit line circuit having terminals coupled to the bit line terminals of each of the memory strings; a sense amplifier circuit having a terminal coupled to a sense amplifier terminal of each of said memory strings; Equipped with wherein the terminals of the bit line circuitry coupled to the bit line terminals of the first portion of the memory strings are capable of providing identical predetermined operating conditions to the bit line terminals of the first portion of the memory strings; wherein the terminals of the bit line circuit coupled to the bit line terminals of the second portion of the memory strings are capable of providing respective dynamic values to the bit line terminals of the second portion of the memory strings; wherein the word line circuitry is capable of selectively driving the control input in accordance with one of first and second word line control values, the first word line control value corresponding to a provided value and the second word line control value corresponding to a result generated by the sense amplifier circuitry based on information provided via the terminals coupled to the bit line terminals of the first portion of the memory strings. Memory system.
12. 12. The memory system of claim 11, further comprising: a counting circuit coupled to the sense amplifier circuit and capable of counting exact matches between values stored in a second portion of the memory string and the respective dynamic values.
13. 12. The memory system of claim 11, further comprising a summing circuit coupled to the sense amplifier circuitry to enable summing according to currents in a second portion of the memory strings.
14. 14. The memory system of claim 13, wherein each of the dynamic values is an analog value, the values stored in the second portions of the memory strings are analog values, and the summing circuitry is operable according to analog techniques.
15. 15. The memory system of claim 14, wherein the summing circuit is further capable of scaling the currents prior to summing.
16. 16. The memory system of claim 11, further comprising one or more processors capable of accessing the memory array.
17. 16. The memory system of claim 11, further comprising: a page cache circuit coupled to the sense amplifier circuit and capable of caching a resulting page read from values stored in the second portion of the memory strings, the terminals of the bit line circuits coupled to the bit line terminals of the second portion of the memory strings being capable of providing identical predetermined operating conditions to the bit line terminals of the second portion of the memory strings.
18. a memory array including first and second portions of memory strings, where one or more of the memory strings in the first portion of memory strings and one or more of the memory strings in the second portion of memory strings each include a plurality of contiguously connected memory devices; a word line circuit having terminals coupled to respective control inputs of respective rows in which said memory strings are organized; a bit line circuit having terminals coupled to respective bit line terminals of the memory strings and capable of providing identical predetermined operating conditions to the bit line terminals; a sense amplifier circuit having terminals coupled to the sense amplifier terminals of each of the memory strings; the word line circuitry is capable of selectively driving the control input in accordance with one of first and second word line control values, the first word line control value corresponding to a provided value and the second word line control value corresponding to a result produced by the sense amplifier circuitry based on information provided via the terminals coupled to the bit line terminals of the first portion of the memory strings. Memory system.
19. 20. The memory system of claim 18, wherein the provided values include a key and a mask, and the selective driving of the control inputs includes selectively driving the control inputs in respective pairs according to a plurality of encodings determined respectively from a respective bit of the key and a corresponding respective bit of the mask, wherein (i) a first one of the plurality of encodings enables selectively matching a stored zero value in response to the provided value being zero, (ii) a second one of the plurality of encodings enables selectively matching a stored one value in response to the provided value being one, (iii) a third one of the plurality of encodings enables selectively matching a stored zero value or a stored one value in response to the provided value being zero or one, and (iv) a quarter of the plurality of encodings disables selectively matching any stored value to the provided value.
20. 20. The memory system of claim 18 or 19, further comprising one or more processors capable of accessing a computational solid-state drive (SSD) including the memory array, the memory array including at least one memory device.
Citation Information
Patent Citations
Semiconductor storage device and information processing device
JP2013030254A
System and method for retrieving associative memory (CAM).
JP2015523670A
Fuzzy string search circuit
US20230073272A1
Selective match line pre-charging in a CAM device using pre-compare operations
US7050318B1
