Ryun-Han Koo
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Ryun-Han Koo has a total of 80 co-authors. Here are the top co-authors by number of shared publications:
- Jong-Ho Lee
Department of Electrical and Computer Engineering and Inter‐university Semiconductor Research Center Seoul National University Seoul Republic of Korea
14 shared publications - Jiseong Im
Seoul National University
11 shared publications Jonghyun Ko
Department of Electrical and Computer Engineering of Seoul National University and Inter-university Semiconductor Research Center
10 shared publications- Wonjun Shin
Department of Semiconductor Convergence Engineering Sungkyunkwan University Suwon Republic of Korea
9 shared publications - Sung-Ho Park
Department of Electrical and Computer Engineering and Inter‐university Semiconductor Research Center Seoul National University Seoul Republic of Korea
8 shared publications - Daewoong Kwon
Department of Electronic Engineering, Hanyang University, Seoul, Republic of Korea
6 shared publications
Latest Publications
- Open Access
- 20 August 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1) , 9963
Intelligent artificial olfactory nervous system with sensory adaptation capabilities
Accurate gas identification across diverse environments is crucial for the advancement of intelligent artificial olfaction, but remains challenging for conventional data-driven learning approaches. Here, we report an artificial olfactory nervous system capable of reliably identifying novel gas stimuli regardless of environmental variations by leveraging sensory adaptation mechanisms inspired by biological olfactory systems. Inspired by biological olfactory systems, our system effectively integrates artificial olfactory receptor neurons, cortical neurons, and intermediate olfactory neurons that facilitate signal transmission. Receptor neurons can dynamically adapt their sensitivity to olfactory stimuli, enabling rapid responses to familiar gases and selective filtering of background interference, thus generating distinctive signal patterns for novel gases. Signals from receptor neurons are modulated in olfactory neurons, then fed into the cortical neurons and translated into olfactory perception. We show that, despite being trained in limited settings, the system can accurately identify gases in various environments.
- 20 August 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1) , 9963
Open AccessIntelligent artificial olfactory nervous system with sensory adaptation capabilities
Accurate gas identification across diverse environments is crucial for the advancement of intelligent artificial olfaction, but remains challenging for conventional data-driven learning approaches. Here, we report an artificial olfactory nervous system capable of reliably identifying novel gas stimuli regardless of environmental variations by leveraging sensory adaptation mechanisms inspired by biological olfactory systems. Inspired by biological olfactory systems, our system effectively integrates artificial olfactory receptor neurons, cortical neurons, and intermediate olfactory neurons that facilitate signal transmission. Receptor neurons can dynamically adapt their sensitivity to olfactory stimuli, enabling rapid responses to familiar gases and selective filtering of background interference, thus generating distinctive signal patterns for novel gases. Signals from receptor neurons are modulated in olfactory neurons, then fed into the cortical neurons and translated into olfactory perception. We show that, despite being trained in limited settings, the system can accurately identify gases in various environments.
- Open Access
- 10 July 2026
- Published by Springer Nature in Research Square
Interpretable Kolmogorov-Arnold Network Using Vertical NAND Flash Memory
Compute-in-memory (CIM) hardware based on commercial vertical NAND (V-NAND) flash memory offers high density and low-power operation, but most neural-network implementations remain difficult to interpret because the learned input-output relationship is embedded in entangled weights. Here we demonstrate an interpretable CIM implementation of a Kolmogorov-Arnold network (KAN) using commercial triple-level-cell V-NAND flash memory with more than 100 stacked layers. In the proposed V-NAND KAN, software-trained edge-wise activation functions are transferred to hardware by approximating them as weighted sums of measured V-NAND spline responses. These spline primitives are physically generated from pairs of series-connected cells driven by complementary word-line biases, while threshold-voltage programming tunes their position and width without modifying the memory structure. We first validate this mapping by implementing multiplication and division, representative symbolic operations extracted from software KANs, and show close agreement between ideal software targets and measured V-NAND KAN outputs. We then apply the approach to multimodal gas sensing using integrated gas, temperature and pressure sensor data. The KAN identifies a compact and physically meaningful functional relationship and enables real-time, low-power gas-concentration inference under mixed-gas conditions. Compared with multilayer perceptron-based implementations, the V-NAND KAN achieves comparable prediction performance while requiring far fewer memory cells and reducing the read-operation component of inference energy by up to 94%. These results establish commercial V-NAND flash memory as a practical platform for interpretable, low-power CIM hardware that can both predict and explain.
- 10 July 2026
- Published by Springer Nature in Research Square
Open AccessInterpretable Kolmogorov-Arnold Network Using Vertical NAND Flash Memory
Compute-in-memory (CIM) hardware based on commercial vertical NAND (V-NAND) flash memory offers high density and low-power operation, but most neural-network implementations remain difficult to interpret because the learned input-output relationship is embedded in entangled weights. Here we demonstrate an interpretable CIM implementation of a Kolmogorov-Arnold network (KAN) using commercial triple-level-cell V-NAND flash memory with more than 100 stacked layers. In the proposed V-NAND KAN, software-trained edge-wise activation functions are transferred to hardware by approximating them as weighted sums of measured V-NAND spline responses. These spline primitives are physically generated from pairs of series-connected cells driven by complementary word-line biases, while threshold-voltage programming tunes their position and width without modifying the memory structure. We first validate this mapping by implementing multiplication and division, representative symbolic operations extracted from software KANs, and show close agreement between ideal software targets and measured V-NAND KAN outputs. We then apply the approach to multimodal gas sensing using integrated gas, temperature and pressure sensor data. The KAN identifies a compact and physically meaningful functional relationship and enables real-time, low-power gas-concentration inference under mixed-gas conditions. Compared with multilayer perceptron-based implementations, the V-NAND KAN achieves comparable prediction performance while requiring far fewer memory cells and reducing the read-operation component of inference energy by up to 94%. These results establish commercial V-NAND flash memory as a practical platform for interpretable, low-power CIM hardware that can both predict and explain.
- Open Access
- 29 June 2026
- Published by Elsevier in SSRN Electronic Journal
Hybrid Approximate-Accurate Neural Network Using Vertical NAND Flash Memory
We propose a hybrid approximate-accurate neural-network scheme based on commercial vertical NAND (V-NAND) flash memory, in which device-level noise characteristics are directly linked to layer-dependent inference conditions. V-NAND flash memory provides a dense and manufacturable platform for neuromorphic hardware, but its low-frequency read noise can degrade analog inference accuracy. Electrical measurements on a commercial >100-layer V-NAND device show that carrier-number fluctuation is the dominant origin of the normalized read-current noise under the measured conditions. Based on this behavior, a layer-wise read-bias (LW-VRead) scheme is used to lower the read and pass biases of noise-tolerant layers, whereas a tied-word-line (t-WL) weight representation is used to suppress normalized noise and enable finer conductance control in accuracy-critical operation. The measured device characteristics are incorporated into VGG-9/CIFAR-10, ResNet-18/CIFAR-100 and VGG-19/CIFAR-100 inference simulations. The hybrid scheme improves inference accuracy while substantially reducing synaptic read power; the VGG-9/CIFAR-10 case shows a 93.24% reduction in synaptic read power, and the ResNet-18/CIFAR-100 case shows a 47.19% reduction, relative to the corresponding conventional biasing conditions. These results place commercial V-NAND flash memory in a practical operating space between approximate and accuracy-oriented neuromorphic inference through device-circuit-algorithm co-optimization.
- 29 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessHybrid Approximate-Accurate Neural Network Using Vertical NAND Flash Memory
We propose a hybrid approximate-accurate neural-network scheme based on commercial vertical NAND (V-NAND) flash memory, in which device-level noise characteristics are directly linked to layer-dependent inference conditions. V-NAND flash memory provides a dense and manufacturable platform for neuromorphic hardware, but its low-frequency read noise can degrade analog inference accuracy. Electrical measurements on a commercial >100-layer V-NAND device show that carrier-number fluctuation is the dominant origin of the normalized read-current noise under the measured conditions. Based on this behavior, a layer-wise read-bias (LW-VRead) scheme is used to lower the read and pass biases of noise-tolerant layers, whereas a tied-word-line (t-WL) weight representation is used to suppress normalized noise and enable finer conductance control in accuracy-critical operation. The measured device characteristics are incorporated into VGG-9/CIFAR-10, ResNet-18/CIFAR-100 and VGG-19/CIFAR-100 inference simulations. The hybrid scheme improves inference accuracy while substantially reducing synaptic read power; the VGG-9/CIFAR-10 case shows a 93.24% reduction in synaptic read power, and the ResNet-18/CIFAR-100 case shows a 47.19% reduction, relative to the corresponding conventional biasing conditions. These results place commercial V-NAND flash memory in a practical operating space between approximate and accuracy-oriented neuromorphic inference through device-circuit-algorithm co-optimization.
- Open Access
- 21 May 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Hybrid ferroelectric-ionic memristive hardware for high scalability in-memory computing
In-memory computing using two-terminal memristors offers a promising route to reduce the energy demands of data-intensive computing. However, current devices scale poorly due to sneak currents and materials that are incompatible with standard complementary metal-oxide-semiconductor and very large-scale integration processes. Here we demonstrate a self-rectifying memristor that unifies resistive switching and diode-like rectification in a single device, a hybrid ferroelectric-ionic tunnel diode fabricated using complementary metal-oxide-semiconductor compatible materials and processes. We harness the collective (ferroelectric-antiferroelectric polymorphism) and defective (ionic) switching behaviors of HfO2 - ZrO2 to synergistically enhance both its electroresistance and rectifying behavior. Furthermore, conformal atomic layer deposition enables the integration of three-dimensional device structures, yielding high on/off (9.3 × 107) and rectifying (1.7 × 106) ratios with a storage capacity of 10 Gb. These results highlight the potential of this device as a hardware building block for scalable in-memory computing platforms.
- 21 May 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Open AccessHybrid ferroelectric-ionic memristive hardware for high scalability in-memory computing
In-memory computing using two-terminal memristors offers a promising route to reduce the energy demands of data-intensive computing. However, current devices scale poorly due to sneak currents and materials that are incompatible with standard complementary metal-oxide-semiconductor and very large-scale integration processes. Here we demonstrate a self-rectifying memristor that unifies resistive switching and diode-like rectification in a single device, a hybrid ferroelectric-ionic tunnel diode fabricated using complementary metal-oxide-semiconductor compatible materials and processes. We harness the collective (ferroelectric-antiferroelectric polymorphism) and defective (ionic) switching behaviors of HfO2 - ZrO2 to synergistically enhance both its electroresistance and rectifying behavior. Furthermore, conformal atomic layer deposition enables the integration of three-dimensional device structures, yielding high on/off (9.3 × 107) and rectifying (1.7 × 106) ratios with a storage capacity of 10 Gb. These results highlight the potential of this device as a hardware building block for scalable in-memory computing platforms.
- Open Access
- 13 May 2026
- Published by Springer Nature in Research Square
Unified matrix-parallel outer-product updates for pretrained deployment and on-device adaptation
Analog in-memory arrays can reduce the energy of neural inference, but efficient deployment of pretrained networks on hardware is still limited by slow, sequential weight programming and chip-specific mismatch. Weight import and fine-tuning are often handled by different update modes, which increases end-to-end latency and energy in practical workflows. Here we demonstrate a unified outer-product pulse engine in a bidirectional dual-gate charge-trap flash transistor array that executes both low-rank SVD-based weight import and gradient-descent fine-tuning using a single matrix-parallel update primitive. In fabricated arrays, parallel outer-product programming achieves 93% agreement within a 2σ target window and transfers a rank-2 pattern in two update steps. In a closed-loop on-wafer Iris classifier, low-rank import reduces the required fine-tuning iterations from 69 to 3 and cuts training latency by 23×. Scaling analysis projects > 6.4× lower transfer latency, > 64× lower training-update latency, and > 209× lower total training energy for a 128×128 array near 90% accuracy. Together, these results establish an end-to-end, array-native workflow that unifies pretrained weight import and on-device fine-tuning through a single matrix-parallel outer-product primitive.
- 13 May 2026
- Published by Springer Nature in Research Square
Open AccessUnified matrix-parallel outer-product updates for pretrained deployment and on-device adaptation
Analog in-memory arrays can reduce the energy of neural inference, but efficient deployment of pretrained networks on hardware is still limited by slow, sequential weight programming and chip-specific mismatch. Weight import and fine-tuning are often handled by different update modes, which increases end-to-end latency and energy in practical workflows. Here we demonstrate a unified outer-product pulse engine in a bidirectional dual-gate charge-trap flash transistor array that executes both low-rank SVD-based weight import and gradient-descent fine-tuning using a single matrix-parallel update primitive. In fabricated arrays, parallel outer-product programming achieves 93% agreement within a 2σ target window and transfers a rank-2 pattern in two update steps. In a closed-loop on-wafer Iris classifier, low-rank import reduces the required fine-tuning iterations from 69 to 3 and cuts training latency by 23×. Scaling analysis projects > 6.4× lower transfer latency, > 64× lower training-update latency, and > 209× lower total training energy for a 128×128 array near 90% accuracy. Together, these results establish an end-to-end, array-native workflow that unifies pretrained weight import and on-device fine-tuning through a single matrix-parallel outer-product primitive.
- Open Access
- 8 May 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
CMOS-compatible ferroelectric tunnel junctions integrate stochastic sampling and deterministic computing for image generation
Recent progress in generative modeling has intensified the need for compact, energy-efficient hardware platforms. Yet, implementing image generation directly in hardware remains challenging due to the conflicting requirements of stochastic latent space sampling and deterministic decoding. Here, we show a unified hardware framework based on hafnium-oxide ferroelectric tunnel junctions (FTJs) that intrinsically support both functionalities within a single device array. Leveraging the CMOS- and VLSI-compatible fabrication of hafnia ferroelectrics, we realize dual-mode operation: random telegraph noise generation for controllable stochastic sampling, and high-fidelity vector–matrix multiplication enabled by non-volatile multi-level conductance states. Voltage and sampling-time tuning provide fine control over randomness and reliability, enabling high-quality image generation for tasks such as handwritten digit synthesis (MNIST) and high-resolution facial image generation (CelebA). Circuit-level demonstrations confirm stable performance over 10 5 cycles, surpassing prior hardware-based approaches and illustrating a viable route toward scalable, on-chip generative AI accelerators.
- 8 May 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Open AccessCMOS-compatible ferroelectric tunnel junctions integrate stochastic sampling and deterministic computing for image generation
Recent progress in generative modeling has intensified the need for compact, energy-efficient hardware platforms. Yet, implementing image generation directly in hardware remains challenging due to the conflicting requirements of stochastic latent space sampling and deterministic decoding. Here, we show a unified hardware framework based on hafnium-oxide ferroelectric tunnel junctions (FTJs) that intrinsically support both functionalities within a single device array. Leveraging the CMOS- and VLSI-compatible fabrication of hafnia ferroelectrics, we realize dual-mode operation: random telegraph noise generation for controllable stochastic sampling, and high-fidelity vector–matrix multiplication enabled by non-volatile multi-level conductance states. Voltage and sampling-time tuning provide fine control over randomness and reliability, enabling high-quality image generation for tasks such as handwritten digit synthesis (MNIST) and high-resolution facial image generation (CelebA). Circuit-level demonstrations confirm stable performance over 10 5 cycles, surpassing prior hardware-based approaches and illustrating a viable route toward scalable, on-chip generative AI accelerators.
- Open Access
- 20 February 2026
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 12 (8) , eaea2328
Repurposing Si CMOS nonidealities for stochastic and analog image processing
Conventional semiconductor device engineering regards intrinsic device nonidealities as reliability concerns to be minimized or eliminated. Here, we demonstrate the strategic repurposing of these nonidealities as functional resources for advanced stochastic analog computing. We leverage two underutilized phenomena—deep-level channel trap-induced generation-recombination (G-R) noise and impact ionization–induced negative differential resistance (NDR) in body current—which have received limited attention compared to the extensively studied 1/ f noise and monotonic drain current behavior in logic-centric transistors. By exploiting G-R noise with controllable temporal correlation and NDR with an unprecedented peak-to-valley ratio (2.78 × 10 4 ) within fully depleted silicon-on-insulator transistors fabricated in industry silicon complementary metal-oxide semiconductor (CMOS) process, we achieve multifunctional analog computation at the single-device level. Our transistor seamlessly performs stochastic encryption, deterministic signal readout, and analog inversion simply through reconfiguration of applied bias conditions, thereby eliminating the need for peripheral random-number generators, dedicated analog inverters, or amplifiers. This approach not only reveals the previously unrecognized computational potential embedded in mature CMOS technologies but also presents a scalable and energy-efficient alternative to architecture based on exotic materials, laying the foundation for next-generation analog computing systems.
- 20 February 2026
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 12 (8) , eaea2328
Open AccessRepurposing Si CMOS nonidealities for stochastic and analog image processing
Conventional semiconductor device engineering regards intrinsic device nonidealities as reliability concerns to be minimized or eliminated. Here, we demonstrate the strategic repurposing of these nonidealities as functional resources for advanced stochastic analog computing. We leverage two underutilized phenomena—deep-level channel trap-induced generation-recombination (G-R) noise and impact ionization–induced negative differential resistance (NDR) in body current—which have received limited attention compared to the extensively studied 1/ f noise and monotonic drain current behavior in logic-centric transistors. By exploiting G-R noise with controllable temporal correlation and NDR with an unprecedented peak-to-valley ratio (2.78 × 10 4 ) within fully depleted silicon-on-insulator transistors fabricated in industry silicon complementary metal-oxide semiconductor (CMOS) process, we achieve multifunctional analog computation at the single-device level. Our transistor seamlessly performs stochastic encryption, deterministic signal readout, and analog inversion simply through reconfiguration of applied bias conditions, thereby eliminating the need for peripheral random-number generators, dedicated analog inverters, or amplifiers. This approach not only reveals the previously unrecognized computational potential embedded in mature CMOS technologies but also presents a scalable and energy-efficient alternative to architecture based on exotic materials, laying the foundation for next-generation analog computing systems.
- Open Access
- 8 February 2026
- Published by Wiley in Advanced Science
- Vol. 13 (21) , e19391
Noise Fingerprints as a Quantitative Order Parameter for Polarization‐ and Defect‐Mediated Switching in Hafnia Ferroelectrics
Noise, often regarded as an unwanted background in electronic devices, can instead serve as a sensitive probe of switching dynamics. In hafnia ferroelectrics, the coexistence of polarization‐mediated (P‐RS) and defect‐mediated (D‐RS) resistance switching has been widely debated, yet prior evidence has remained qualitative. Here, we demonstrate that low‐frequency noise (LFN) can be transformed into a quantitative order parameter that disentangles the two mechanisms across program bias ( V PGM ) and processing conditions. By varying the O 3 dose time during HfZrO atomic layer deposition (ALD), the normalized power spectral density ( S I / I 2 ) consistently exhibits a rise–peak–fall profile. We introduce a deconvolution framework, grounded in monotonic baselines and alternating projections, to extract physically consistent polarization‐ and defect‐mediated current components. The resulting process‐bias maps show that reduced O 3 dose shifts the P‐RS/D‐RS crossover to lower V PGM and sharpens the transition, directly linking oxygen stoichiometry to the competition between switching pathways. This quantitative approach resolves the long‐standing controversy over P‐RS and D‐RS coexistence and provides practical guidance for the design of hafnia‐based ferroelectric devices.
- 8 February 2026
- Published by Wiley in Advanced Science
- Vol. 13 (21) , e19391
Open AccessNoise Fingerprints as a Quantitative Order Parameter for Polarization‐ and Defect‐Mediated Switching in Hafnia Ferroelectrics
Noise, often regarded as an unwanted background in electronic devices, can instead serve as a sensitive probe of switching dynamics. In hafnia ferroelectrics, the coexistence of polarization‐mediated (P‐RS) and defect‐mediated (D‐RS) resistance switching has been widely debated, yet prior evidence has remained qualitative. Here, we demonstrate that low‐frequency noise (LFN) can be transformed into a quantitative order parameter that disentangles the two mechanisms across program bias ( V PGM ) and processing conditions. By varying the O 3 dose time during HfZrO atomic layer deposition (ALD), the normalized power spectral density ( S I / I 2 ) consistently exhibits a rise–peak–fall profile. We introduce a deconvolution framework, grounded in monotonic baselines and alternating projections, to extract physically consistent polarization‐ and defect‐mediated current components. The resulting process‐bias maps show that reduced O 3 dose shifts the P‐RS/D‐RS crossover to lower V PGM and sharpens the transition, directly linking oxygen stoichiometry to the competition between switching pathways. This quantitative approach resolves the long‐standing controversy over P‐RS and D‐RS coexistence and provides practical guidance for the design of hafnia‐based ferroelectric devices.
- Open Access
- 25 November 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1)
Harnessing chaotic bifurcation in positive feedback transistors for secure and scalable random key generation
In the era of big data, robust data security is more critical than ever. One promising way to create unpredictable random keys is to use the intrinsic physical randomness of electronic devices. However, controlling and reproducing this randomness remains a significant challenge. For the first time, we reveal a parabolic input-output relationship enabled by the band-to-band tunneling (BTBT) mechanism in a positive feedback transistor (PFT). By applying iterative operations, we observe chaotic bifurcation behavior, which allows the stable generation of analog random values across more than 1010 cycles. These random values are then used to encrypt and decrypt images, showcasing how encryption technologies can be implemented with just a single PFT. We also show that the charge-trap flash memory (CTF) array integrated with the PFTs can store encrypted information. This study highlights the potential of PFTs as random key generators and examines the device’s behavior under chaotic conditions. By leveraging the physical characteristics of the device, we provide a secure and reliable method for data encryption.
- 25 November 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1)
Open AccessHarnessing chaotic bifurcation in positive feedback transistors for secure and scalable random key generation
In the era of big data, robust data security is more critical than ever. One promising way to create unpredictable random keys is to use the intrinsic physical randomness of electronic devices. However, controlling and reproducing this randomness remains a significant challenge. For the first time, we reveal a parabolic input-output relationship enabled by the band-to-band tunneling (BTBT) mechanism in a positive feedback transistor (PFT). By applying iterative operations, we observe chaotic bifurcation behavior, which allows the stable generation of analog random values across more than 1010 cycles. These random values are then used to encrypt and decrypt images, showcasing how encryption technologies can be implemented with just a single PFT. We also show that the charge-trap flash memory (CTF) array integrated with the PFTs can store encrypted information. This study highlights the potential of PFTs as random key generators and examines the device’s behavior under chaotic conditions. By leveraging the physical characteristics of the device, we provide a secure and reliable method for data encryption.
- Open Access
- 14 November 2025
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 11 (46) , eadw7498
Hybrid functional 3D artificial synapses for convolution and reinforcement learning
Recent advancements in compute-in-memory (CIM) technology have garnered substantial interest, representing a promising shift in the computing paradigm. While conventional CIM systems using two-dimensional (2D) crossbar arrays have shown remarkable area and energy efficiency in fully connected neural networks, their applicability to convolutional neural networks remains inefficient. Herein, we introduce an architecture that enables convolution and reinforcement learning by leveraging the unique configuration of vertically stacked 3D flash memory. The proposed 3D convolution block (3D CB) significantly reduces area overhead and enhances the reliability and energy efficiency of convolution operations. The 3D CB and 3D fully connected block (3D FCB) exhibit distinct functionalities with minimal structural modifications. The compatibility between the 3D CB and the 3D FCB facilitates their seamless co-integration on a single wafer. The fully integrated system presented in this study demonstrates precise and consistent path planning capabilities for autonomous driving with high area and energy efficiency, paving the way for advanced CIM technology.
- 14 November 2025
- Published by American Association for the Advancement of Science (AAAS) in Science Advances
- Vol. 11 (46) , eadw7498
Open AccessHybrid functional 3D artificial synapses for convolution and reinforcement learning
Recent advancements in compute-in-memory (CIM) technology have garnered substantial interest, representing a promising shift in the computing paradigm. While conventional CIM systems using two-dimensional (2D) crossbar arrays have shown remarkable area and energy efficiency in fully connected neural networks, their applicability to convolutional neural networks remains inefficient. Herein, we introduce an architecture that enables convolution and reinforcement learning by leveraging the unique configuration of vertically stacked 3D flash memory. The proposed 3D convolution block (3D CB) significantly reduces area overhead and enhances the reliability and energy efficiency of convolution operations. The 3D CB and 3D fully connected block (3D FCB) exhibit distinct functionalities with minimal structural modifications. The compatibility between the 3D CB and the 3D FCB facilitates their seamless co-integration on a single wafer. The fully integrated system presented in this study demonstrates precise and consistent path planning capabilities for autonomous driving with high area and energy efficiency, paving the way for advanced CIM technology.