High-throughput RRAM content addressable memory and genome analysis search method

Abstract

The invention discloses a high-throughput RRAM (Resistive Random Access Memory) content addressable memory and a genome analysis search method, which realize higher storage density, lower storage overhead and higher parallel search throughput rate through collaborative design of circuit-level structure optimization and a search algorithm. The method can be widely applied to gene database retrieval, pattern recognition, edge intelligent calculation and other high-parallel pattern matching processing systems.

Description

Technical Field

[0001] This invention belongs to the field of integrated circuit design and in-memory computing technology, specifically relating to high-throughput RRAM content-addressable memory and genome analysis search methods. Background Technology

[0002] Genome analysis has become a key supporting technology for applications such as personalized medicine, virus surveillance, and infectious disease control. One of its core steps is read mapping, which requires comparing a large number of sequencing reads with a large-scale reference genome database. The reference sequence can be up to billions of base pairs, so this process is usually the most computationally intensive part of the entire genome analysis pipeline.

[0003] [1] Content-addressable memory (CAM) is widely used to accelerate read mapping due to its massively parallel comparison capability, but the throughput of existing CAM architectures for genome analysis is still constrained by various factors. [2-6] The relevant technologies currently have the following three main problems: 1. Large unit area: To support approximate search functions such as Hamming distance and edit distance, traditional genome analysis CAM units require complex circuit structures, resulting in limited integration density and insufficient on-chip storage capacity, which in turn limits the overall throughput. For example, existing SHD CAMs (such as ShiftCAM) use a unit structure of about 13T4R1C, which has a large area overhead. [6] 2. High array storage redundancy: To adapt to the k-mer alignment process, the reference genome typically needs to be preprocessed into fragments of length k and stored with high overlap between adjacent fragments. The overlap between adjacent k-mers reaches k... One base can result in approximately 98% redundancy in the array, severely reducing effective storage utilization and the size of the supported reference database. [2-6] 3. Low search algorithm efficiency: High-precision genome analysis usually relies on complex comparison processes such as shifted Hamming distance (SHD), which requires repeated calculation of similarity at multiple shift positions, resulting in high search latency; at the same time, the lack of an effective early filtering mechanism for candidates with low matching potential leads to a large number of invalid comparisons and a significant decrease in overall throughput. [6] Against this backdrop, it is necessary to propose a novel CAM architecture that synergistically optimizes the unit structure, array storage organization, and search algorithm to significantly improve on-chip storage efficiency and read mapping throughput for genome analysis while ensuring alignment accuracy.

[0004] [1]A. F. Laguna, H. Gamaarachchi, X. Yin, M. Niemier, S.Parameswaran, and X. S. Hu, “Seed-and-vote based In-memory accelerator forDNA read mapping,” in 2020 IEEE / ACM International Conference on ComputerAided Design (ICCAD), 2020, pp. 1–9. [2]I. Merlin, E. Garz´on, A. Fish, and L. Yavits, “DIPER: Detectionand identification of pathogens using edit distance-tolerant resistive CAM,”IEEE Transactions on Computers, vol. 73, no. 10, pp. 2463–2473, 2024. [3]Z. Jahshan, I. Merlin, E. Garz´on, and L. Yavits, “DASH-CAM:Dynamic approximate SearcH content addressable memory for genomeclassification,” in 2023 56th IEEE / ACM International Symposium onMicroarchitecture (MICRO), 2023, pp. 1453–1465. [4]R. Hanhan, E. Garz´on, Z. Jahshan, A. Teman, M. Lanuzza, and L.Yavits, “EDAM: Edit distance tolerant approximate matching contentaddressable memory,” in Proceedings of the 49th Annual InternationalSymposium on Computer Architecture, ser. ISCA ’22. New York, NY, USA:Association for Computing Machinery, 2022, pp. 495–507. [5]H. Zhong, Z. Chen, W. Huangfu, C. Wang, Y. Xu, T. Wang, Y. Yu, Y.Liu, V. Narayanan, H. Yang, and X. Li, “ASMCap: An approximate stringmatching accelerator for genome sequence analysis based on capacitive contentaddressable memory,” in 2023 60th ACM / IEEE Design Automation Conference(DAC), 2023, pp. 1–6. [6] P. He, R. Mao, K. Shan, Y. Tong, Z. Xu, M. Peng, R. Luo, and C. Li, "ShiftCAM: A time-domain content addressable memory utilizing shiftedhamming distance for robust genome analysis," in Proceedings of the 43rd IEEE / ACM International Conference on Computer-aided Design. New York, NY, USA: Association for Computing Machinery, 2025, no. 87, pp. 1–9. Summary of the Invention The purpose of this invention is to address the problems of existing architectures based on Static Random-Access Memory (SRAM), embedded Dynamic Random-Access Memory (eDRAM), and traditional RRAM-CAM, such as large cell area, high data storage redundancy, low pattern matching efficiency, and limited throughput.

[0005] To address the aforementioned technical problems, the present invention provides a high-throughput RRAM content-addressable memory and a genome analysis search method, including: Design an RRAM CAM cell structure with time-domain comparison function, use RRAM one-hot encoding and pre-charge / discharge logic to optimize the comparison function circuit, so that a single cell can perform multiple cycle comparisons, and achieve fuzzy matching through shift input; By combining the high-throughput DSHD-CAM architecture of the CAM array as a high-throughput RRAM content-addressable memory, and using a dynamic SHD search algorithm for collaborative optimization, genome analysis search is realized.

[0006] Preferably, by storing multiple reference k-mers within a single row, overlapping segments between adjacent k-mers are compressed, thereby reducing the storage redundancy of the CAM array.

[0007] Preferably, the dynamic SHD search algorithm is performed by directly and quickly locating the input to a position with a high matching probability.

[0008] Preferably, the single unit achieves the comparison between the current reference base and the current and adjacent input bases through multi-cycle comparison by pre-charging the Mid node and maintaining the charge.

[0009] Preferably, the statistical characteristics of DNA sequences are used to set multi-level thresholds for the number of matching bases. Based on the multi-level thresholds, sequences with low matching potential and sequences with high matching potential are distinguished. The sequences with low matching potential are filtered out in advance, and the sequences with high matching potential are triggered to complete the time-consuming SHD calculation.

[0010] The technical solution of this invention provides a high-throughput RRAM content-addressable memory and a genome analysis search method. Through circuit-level structure optimization and search algorithm co-design, it achieves higher storage density, lower storage overhead and higher parallel search throughput, and can be widely used in gene database retrieval, pattern recognition, edge intelligent computing and other highly parallel pattern matching processing systems. Attached Figure Description

[0011] Figure 1 The DSHD-CAM macro-architecture provided in this embodiment of the invention; Figure 2 The diagram shows the time-domain RRAM 6T1R1C CAM cell structure and CAM array row structure in the DSHD-CAM macro-architecture provided in this embodiment of the invention, as well as a schematic diagram of nucleotide one-hot encoding. Figure 3 This invention provides a schematic diagram of the CAM unit comparison operation and the relationship between the result voltage and the number of matches. Figure 4 The waveform of DSHD-CAM in detail-search working mode provided in the embodiments of the present invention; Figure 5 The waveform of DSHD-CAM in fast-search working mode provided in the embodiments of the present invention; Figure 6 The illustrations provided for the previous CAM storage reference k-mers and the schematic diagram of DSHD-CAM storing multiple reference k-mers in a single row during storage are shown in the embodiments of the present invention. Figure 7 The flowchart of the DSHD search algorithm used in the DSHD-CAM provided in the embodiments of the present invention, as well as the working stages of fast-search and detail-search and corresponding examples therein; Figure 8 A detailed hardware architecture diagram of a gene classification accelerator based on DSHD-CAM macros provided for embodiments of the present invention. Detailed Implementation

[0012] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0013] This invention addresses the problems of large area, high array storage redundancy, and large search process latency in existing genome analysis CAM units. It provides a high-throughput RRAM content-addressable memory and genome analysis search method, featuring a novel time-domain RRAM-CAM circuit structure and an efficient SHD search mechanism. Targeting data-intensive applications such as genome sequence analysis, DNA alignment, and bioinformatics computation, the invention specifically involves a high-density time-domain content-addressable memory (CAM) macrocell based on resistive random access memory (RRAM) and its accompanying dynamic shifted Hamming distance (DSHD) search algorithm. The method includes the following steps: like Figure 1 The diagram illustrates the overall framework of this design. This design implements a high-density RRAM CAM cell structure and CAM array with time-domain search functionality. Furthermore, based on cell-level and circuit array-level design, we designed a dynamic SHD search algorithm to achieve hardware-software co-optimization, realizing a high-throughput DSHD-CAM architecture as a high-throughput RRAM content-addressable memory for gene analysis.

[0014] In terms of hardware, previous hardware designs required additional comparison circuits and interconnections between CAM units to implement the fuzzy matching function (comparing a reference base with the current position and multiple adjacent input bases), which resulted in a significant reduction in CAM unit density.

[0015] Our designed CAM unit has a time-domain comparison function, and a single unit can perform multiple cycle comparisons. Fuzzy matching is achieved through shift input. Furthermore, we use RRAM one-hot encoding to optimize the comparison function circuit, which greatly reduces hardware complexity and achieves the highest integration density CAM unit with fuzzy search capability.

[0016] We optimize the reference k-mer (highly repetitive between adjacent k-mers) storage strategy at the CAM array level. Previous designs stored one reference k-mer per row, resulting in significant repetition between rows. Our design stores multiple reference k-mers within a single row, effectively compressing overlapping segments between adjacent k-mers, significantly reducing storage redundancy, and further improving density.

[0017] Based on this, leveraging the characteristics of our designed hardware, we designed a dynamic SHD search algorithm to filter out a large amount of useless SHD calculations. By directly and quickly locating the input to a position with a high matching probability, we achieve efficient SHD calculations, thereby significantly improving on-chip read mapping throughput while ensuring matching accuracy.

[0018] like Figure 2 As shown, in terms of cell structure, unlike previous designs that used hard connections to compare the current reference base with the current and adjacent input bases, our design utilizes highly integrated RRAM technology, combined with one-hot encoding and time-domain comparison calculations, to significantly improve the CAM cell density. First, our design achieves the comparison of the current reference base with the current and adjacent input bases through multi-cycle comparisons by pre-charging the Mid node and maintaining its charge, greatly reducing the hardware overhead of implementing fuzzy comparisons. Second, by using RRAM one-hot encoding, the circuit overhead for implementing the comparison function is further simplified.

[0019] Figure 3 Examples of Mid node leakage under specific matching conditions and Mid node retention under mismatch conditions are shown, along with the relationship between the final ML voltage and the comparison results.

[0020] Figure 4-5 Examples of two operating modes of a CAM array are shown. Figure 4 The "detailed search" working mode is demonstrated. The Mid node is refreshed at the beginning of the multi-cycle comparison. After multiple searches, the results are compared, and the OR operation of the search results of multiple cycles is implemented. Finally, the ML voltage corresponds to the SHD calculation result. Figure 5 The "fast search" operating mode was demonstrated, where the Mid node is refreshed every cycle, and the voltage of ML corresponds to the comparison result of a single cycle (corresponding to HD). In summary, our designed time-domain RRAMCAM cell structure and CAM array row structure achieve the most compact CAM cell design while supporting SHD search capabilities. The cell structure is compressed to 6T4R1C (compared to the previous 42T, 6T12R, and 13T4R1C), significantly reducing cell area and circuit area. This facilitates the integration of a larger-scale reference gene database within the same chip area, thereby improving overall throughput.

[0021] like Figure 6 As shown, this invention implements a low-redundancy dense array storage strategy. On a CAM array, previous designs directly stored all reference k-mers ( Figure 6 (a) and (b) ), due to the overlapping segments between adjacent k-mers leading to significant storage redundancy, our array stores multiple reference k-mers within a single row. Figure 6 (c) and (d) fully utilize overlapping fragments between adjacent k-mers and combine them with in-row shifting alignment methods to reduce the effective array size from approximately kN in traditional schemes to approximately 2N, significantly reducing redundancy. Under typical genome length and k-value settings, array storage redundancy is reduced from approximately 98% to approximately 50%, and the required capacity for storing the same reference database can be reduced by approximately 63.75 times, significantly improving the size of reference sequences that can be accommodated on the chip and the array's parallel search capability.

[0022] like Figure 7 As shown, based on our designed CAM unit with time-domain search capability and low-redundancy dense array storage strategy, we further designed a dynamic SHD search algorithm based on "fast search + detailed search" (…). Figure 7 (a) achieves collaborative optimization of circuitry and algorithms. The redundant dense array storage strategy stores multiple reference k-mers within a single row, freeing up the possibility of skipping useless computations. Our designed dynamic SHD search algorithm fully utilizes the statistical characteristics of DNA sequences to set multi-level thresholds for the number of matching bases, performing early filtering for comparisons with low matching potential, and triggering only the full, time-consuming SHD computation for candidates with high matching potential. Its specific workflow is as follows: For an input k-mer, CAM first operates in "fast search" mode (… Figure 7 (b) Demonstrates the specific operational details of CAM. It quickly finds locations with a high probability of matching that meet a low threshold, after which CAM enters "detailed search" mode. Figure 7 (c) Demonstrating the specific operational details of CAM, a detailed SHD calculation is performed to confirm whether it is a true match. If it is a match, the current input search process ends; otherwise, CAM will continue with a "fast search" to find the next high-probability match location.

[0023] Figure 8A diagram of a gene classification accelerator based on DSHD-CAM macros is shown. For an input DNA sequence (Read), which contains multiple k-mers, a dynamic SHD search algorithm is used to search for each k-mer in the DSHD-CAM. If a match is found, the hit-counter is incremented. If the final count exceeds a set threshold, the input DNA is considered to be classified into the reference type stored in the current CAM array. Under typical operating conditions, by combining increased hardware density and algorithm optimization, this system can reduce the average search latency by approximately 8.53 times while maintaining classification accuracy comparable to existing SHD-CAMs, thereby further improving the overall system throughput by 7.58 times.

[0024] This invention is implemented using TSMC's 65nm process, with a core array specification of 256×256 bits (macro area of ​​0.22mm²), a core voltage of 1.0V, and a cell area of ​​only about 2.54 μm², which is about 4.61 times smaller than existing SOTA designs (such as ShiftCAM with a 13T4R1C structure).

[0025] Simulation results show that the actual search frequency can reach 833.33MHz, the search time is 1.2ns, and the average power consumption is 1.49μW. Under Monte Carlo simulation (including transistor and RRAM errors), the worst mismatch case has a readout accuracy of 94%. For different virus datasets, the best F1 score (a measure of matching accuracy) can match the previous state-of-the-art (SOTA) (ShiftCAM) and is twice that of other HD / ED CAM designs.

[0026] Algorithm analysis experiments show that the dynamic SHD search algorithm can filter out an average of 97.5% of mismatches without sacrificing accuracy, reducing latency by approximately 8.53 times compared to not using the algorithm. Throughput experiments show that, under the same area constraint, DSHD-CAM achieves an average throughput of 7.58× of the state-of-the-art (SOTA) for gene classification tasks, while maintaining competitive sensitivity and precision. This design is suitable for applications such as genome classification and pathogen detection, significantly improving the processing efficiency of data-intensive tasks.

Claims

1. A high-throughput RRAM content-addressable memory and a genome analysis search method, characterized in that, include: Design an RRAM CAM cell structure with time-domain comparison function, use RRAM one-hot encoding and pre-charge / discharge logic to optimize the comparison function circuit, so that a single cell can perform multiple cycle comparisons, and achieve fuzzy matching through shift input; By combining the high-throughput DSHD-CAM architecture of the CAM array as a high-throughput RRAM content-addressable memory, and using a dynamic SHD search algorithm for collaborative optimization, genome analysis search is realized.

2. The high-throughput RRAM content-addressable memory and genome analysis search method as described in claim 1, characterized in that, By storing multiple reference k-mers within a single row, overlapping segments between adjacent k-mers are compressed, reducing storage redundancy in the CAM array.

3. The high-throughput RRAM content-addressable memory and genome analysis search method as described in claim 1, characterized in that, The dynamic SHD search algorithm is performed by directly and quickly locating the input to a position with a high matching probability.

4. The high-throughput RRAM content-addressable memory and genome analysis search method as described in claim 1, characterized in that, The individual unit achieves comparison between the current reference base and the current and adjacent input bases through multi-cycle comparison by pre-charging the Mid node and maintaining its charge.

5. The high-throughput RRAM content-addressable memory and genome analysis search method as described in claim 1, characterized in that, By utilizing the statistical characteristics of DNA sequences, multi-level thresholds are set for the number of matching bases. Based on these thresholds, sequences with low matching potential and sequences with high matching potential are distinguished. Low-matching-potential sequences are filtered out early, while high-matching-potential sequences are subjected to full-time SHD calculations.

Images