Zeyu Han
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Zeyu Han has a total of 8 co-authors. Here are the top co-authors by number of shared publications:
- Yuki Yamanashi
Yokohama National University
9 shared publications - Nobuyuki Yoshikawa
Yokohama National University
8 shared publications - Hongxiang Shen
Department of Electrical and Computer Engineering, Yokohama National University, Yokohama 240-8501, Japan
2 shared publications Nobuyuki Yoshikawa
Department of Electrical and Computer Engineering, Yokohama National University, Yokohama 240-8501, Japan
2 shared publications- Zongyuan Li
Department of Electrical and Computer Engineering, Yokohama National University, Yokohama 240-8501, Japan
2 shared publications - Yuxing He
School of Information Science and Technology (SIST), Southwest Jiaotong University, Chengdu, China
1 shared publication
Latest Publications
- Open Access
- 1 October 2026
- Published by IOP Publishing in Superconductor Science and Technology
- Vol. 39 (10) , 105006
Exploiting stable double switching in single-flux-quantum circuits for compact scaled unary addition
Single-flux-quantum (SFQ) circuits conventionally avoid simultaneous or competing switching events in Josephson junction (JJ) comparator because they can induce pulse errors and nondeterministic behavior. In this work, we demonstrate that such double switching can instead be intentionally exploited as a useful computational mechanism. By introducing unshunted JJs before the comparator structure, stable double switching of two junctions is realized, enabling two input SFQ pulses to be merged into a single output pulse. The proposed mechanism operates correctly regardless of the switching order of the two junctions. To demonstrate the feasibility of the double switching mechanism, a double-switching confluence buffer was designed as a compact scaled unary adder. The proposed adder requires only 12 JJs, achieving approximately 84% area reduction compared with a conventional SFQ scaled unary adder. Simulations achieved bias margins of −18.6% to +20.2%, while measured bias margins of +17.9% to +28.5% and correct operation up to approximately 55 GHz were experimentally demonstrated. The results indicate that double switching, traditionally regarded as undesirable in SFQ circuit design, can instead be harnessed as a useful primitive for compact SFQ logic circuits.
- 1 October 2026
- Published by IOP Publishing in Superconductor Science and Technology
- Vol. 39 (10) , 105006
Open AccessExploiting stable double switching in single-flux-quantum circuits for compact scaled unary addition
Single-flux-quantum (SFQ) circuits conventionally avoid simultaneous or competing switching events in Josephson junction (JJ) comparator because they can induce pulse errors and nondeterministic behavior. In this work, we demonstrate that such double switching can instead be intentionally exploited as a useful computational mechanism. By introducing unshunted JJs before the comparator structure, stable double switching of two junctions is realized, enabling two input SFQ pulses to be merged into a single output pulse. The proposed mechanism operates correctly regardless of the switching order of the two junctions. To demonstrate the feasibility of the double switching mechanism, a double-switching confluence buffer was designed as a compact scaled unary adder. The proposed adder requires only 12 JJs, achieving approximately 84% area reduction compared with a conventional SFQ scaled unary adder. Simulations achieved bias margins of −18.6% to +20.2%, while measured bias margins of +17.9% to +28.5% and correct operation up to approximately 55 GHz were experimentally demonstrated. The results indicate that double switching, traditionally regarded as undesirable in SFQ circuit design, can instead be harnessed as a useful primitive for compact SFQ logic circuits.
- Open Access
- 4 June 2026
- Published by IOP Publishing in Superconductor Science and Technology
- Vol. 39 (7)
A Superconducting Unary Arithmetic Logic Unit with Ultra-Low Hardware Cost
Single-flux-quantum (SFQ) circuits are promising candidates for post-Moore integrated circuits owing to their ultra-high-speed operation and low power consumption. However, limitations in current superconducting fabrication processes pose significant challenges to achieving area-efficient SFQ-based microprocessors, particularly in the design of compact arithmetic logic units (ALUs). Unary computing enables arithmetic operations using extremely simple logic, but typically suffers from long computation latency. This paper focuses on unary computing for superconducting circuits that exploits the complementary characteristics of SFQ logic and unary computing. The ultra-high-speed nature of SFQ circuits compensates for the inherent latency of unary computation, while unary computing enables an ultra-low hardware cost by substantially reducing circuit complexity. Based on this concept, we introduce a unary computing methodology tailored for SFQ circuits and design a superconducting unary ALU with ultra-low hardware cost. The proposed ALU requires only 183 Josephson junctions (JJs), achieving approximately 88.2% reduction in JJ count and 80.4% reduction in area compared with a conventional bit-serial SFQ ALU. The proposed design, including binary-to-unary and unary-to-binary converters, was fabricated using a 10 kA/cm² Nb four-layer superconducting circuit fabrication process and experimentally verified to operate correctly at approximately 72 GHz. The total power consumption at this frequency is approximately 58.9 µW. This work quantitatively demonstrates the effectiveness of combining unary computing with SFQ circuits for ultra-low-cost processing systems and provides superconducting circuits as a new design option for alternative computation paradigms, such as unary computing.
- 4 June 2026
- Published by IOP Publishing in Superconductor Science and Technology
- Vol. 39 (7)
Open AccessA Superconducting Unary Arithmetic Logic Unit with Ultra-Low Hardware Cost
Single-flux-quantum (SFQ) circuits are promising candidates for post-Moore integrated circuits owing to their ultra-high-speed operation and low power consumption. However, limitations in current superconducting fabrication processes pose significant challenges to achieving area-efficient SFQ-based microprocessors, particularly in the design of compact arithmetic logic units (ALUs). Unary computing enables arithmetic operations using extremely simple logic, but typically suffers from long computation latency. This paper focuses on unary computing for superconducting circuits that exploits the complementary characteristics of SFQ logic and unary computing. The ultra-high-speed nature of SFQ circuits compensates for the inherent latency of unary computation, while unary computing enables an ultra-low hardware cost by substantially reducing circuit complexity. Based on this concept, we introduce a unary computing methodology tailored for SFQ circuits and design a superconducting unary ALU with ultra-low hardware cost. The proposed ALU requires only 183 Josephson junctions (JJs), achieving approximately 88.2% reduction in JJ count and 80.4% reduction in area compared with a conventional bit-serial SFQ ALU. The proposed design, including binary-to-unary and unary-to-binary converters, was fabricated using a 10 kA/cm² Nb four-layer superconducting circuit fabrication process and experimentally verified to operate correctly at approximately 72 GHz. The total power consumption at this frequency is approximately 58.9 µW. This work quantitatively demonstrates the effectiveness of combining unary computing with SFQ circuits for ultra-low-cost processing systems and provides superconducting circuits as a new design option for alternative computation paradigms, such as unary computing.
- 26 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 36 (5) , 1-6
A Finite State Automaton for the Tsetlin Machine Using High-Capacity Resettable D Flip-Flop Cells
Single Flux Quantum (SFQ) circuits provide ultrafast operation and low power consumption, making them attractive candidates for machine learning accelerators. The Tsetlin Machine (TM) is a low-complexity and highly interpretable architecture, organized as a voting system of propositional clauses. In the TM, an ensemble of finite-state automata (FSAs), called Tsetlin automata, independently decides whether to include or exclude specific literals. The chosen literals are assembled into propositional clauses that collectively form the voting system, while each FSA updates its state based on rewards or penalties determined by the prediction outcome. However, existing FSA designs for TM based on SFQ circuits still suffer from either a narrow margin or excessive circuit area. To achieve compact and stable state transitions using SFQ circuits, we designed High-Capacity Resettable D Flip-Flop (HC-RDFF) cells by cascading storage loops. The proposed FSA was implemented using two HC-RDFFs to store the include and exclude states, respectively. The FSA with 16 states requires only 395 Josephson junctions with a static power of approximately 128 μW, exhibiting a bias margin of −21% to +25% at low speed and a maximum operating frequency of 20 GHz in simulation. We fabricated the HC-RDFF and the FSA using a 10 kA/cm² Nb four-layer superconducting process and verified their correct operation through low-speed measurements. The number of FSA states can be increased simply by extending the cascaded storage loops in the HC-RDFFs. We expect the proposed FSA to be suitable for constructing stable TM systems based on SFQ circuits with reduced hardware costs.
- 26 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 36 (5) , 1-6
A Finite State Automaton for the Tsetlin Machine Using High-Capacity Resettable D Flip-Flop Cells
Single Flux Quantum (SFQ) circuits provide ultrafast operation and low power consumption, making them attractive candidates for machine learning accelerators. The Tsetlin Machine (TM) is a low-complexity and highly interpretable architecture, organized as a voting system of propositional clauses. In the TM, an ensemble of finite-state automata (FSAs), called Tsetlin automata, independently decides whether to include or exclude specific literals. The chosen literals are assembled into propositional clauses that collectively form the voting system, while each FSA updates its state based on rewards or penalties determined by the prediction outcome. However, existing FSA designs for TM based on SFQ circuits still suffer from either a narrow margin or excessive circuit area. To achieve compact and stable state transitions using SFQ circuits, we designed High-Capacity Resettable D Flip-Flop (HC-RDFF) cells by cascading storage loops. The proposed FSA was implemented using two HC-RDFFs to store the include and exclude states, respectively. The FSA with 16 states requires only 395 Josephson junctions with a static power of approximately 128 μW, exhibiting a bias margin of −21% to +25% at low speed and a maximum operating frequency of 20 GHz in simulation. We fabricated the HC-RDFF and the FSA using a 10 kA/cm² Nb four-layer superconducting process and verified their correct operation through low-speed measurements. The number of FSA states can be increased simply by extending the cascaded storage loops in the HC-RDFFs. We expect the proposed FSA to be suitable for constructing stable TM systems based on SFQ circuits with reduced hardware costs.
- Open Access
- 1 January 2026
- Published by IOP Publishing in Superconductor Science and Technology
- Vol. 39 (1) , 015003
Small-area and high-precision deterministic stochastic adder using single flux quantum circuit
Superconducting single flux quantum (SFQ) circuits can operate at ultrafast speeds, and their superconducting random number generators (SRNGs) produce high-quality stochastic bit sequences at several tens GHz, with an arbitrary probability of outputting ‘1’, making them ideal for stochastic computing. The conventional superconducting stochastic adder uses a multiplexer (MUX) controlled by a random bit sequence generated by the SRNG. However, the SFQ-based MUX occupies a large circuit area, and any deviation in the probability of the control signal directly degrades the accuracy of stochastic addition. To overcome these limitations, we propose a deterministic stochastic adder comprising only an AND gate, an XOR gate, and a toggle flip-flop. Our design outputs one ‘1’ for every two input ‘1’s, achieving stochastic addition without the random control signal. Compared to the MUX-based stochastic adder, the proposed adder reduces the root-mean-square error by a factor of 32 when adding 1024-bit random sequences, reduces circuit area by 49.0%, and Josephson junction count by 28.1%, while supporting operating frequencies of about 120 GHz in simulation. We fabricated the proposed adder with two SRNGs for generating input sequences using a 10 kA cm −2 Nb four-layer superconducting process. High-speed operation was verified up to approximately 35.4 GHz with a constant bias margin by measuring the average voltage at the input and outputs. To the best of our knowledge, this is the first demonstration of ultrafast stochastic arithmetic operations above 30 GHz.
- 1 January 2026
- Published by IOP Publishing in Superconductor Science and Technology
- Vol. 39 (1) , 015003
Open AccessSmall-area and high-precision deterministic stochastic adder using single flux quantum circuit
Superconducting single flux quantum (SFQ) circuits can operate at ultrafast speeds, and their superconducting random number generators (SRNGs) produce high-quality stochastic bit sequences at several tens GHz, with an arbitrary probability of outputting ‘1’, making them ideal for stochastic computing. The conventional superconducting stochastic adder uses a multiplexer (MUX) controlled by a random bit sequence generated by the SRNG. However, the SFQ-based MUX occupies a large circuit area, and any deviation in the probability of the control signal directly degrades the accuracy of stochastic addition. To overcome these limitations, we propose a deterministic stochastic adder comprising only an AND gate, an XOR gate, and a toggle flip-flop. Our design outputs one ‘1’ for every two input ‘1’s, achieving stochastic addition without the random control signal. Compared to the MUX-based stochastic adder, the proposed adder reduces the root-mean-square error by a factor of 32 when adding 1024-bit random sequences, reduces circuit area by 49.0%, and Josephson junction count by 28.1%, while supporting operating frequencies of about 120 GHz in simulation. We fabricated the proposed adder with two SRNGs for generating input sequences using a 10 kA cm −2 Nb four-layer superconducting process. High-speed operation was verified up to approximately 35.4 GHz with a constant bias margin by measuring the average voltage at the input and outputs. To the best of our knowledge, this is the first demonstration of ultrafast stochastic arithmetic operations above 30 GHz.
- Open Access
- 1 December 2025
- Published by IOP Publishing in Superconductor Science and Technology
- Vol. 38 (12) , 125027
Reconfigurable in-memory XNOR processing element with external DC control currents for scalable superconducting binary neural networks
Binary neural networks (BNNs) reduce memory demands and computational complexity by replacing multiply–accumulate operations with bitwise exclusive-NOR (XNOR) and popcount. On the other hand, in-memory computing improves efficiency by integrating storage and computation within a single element to minimize data transfer, reduce energy consumption, and enhance area utilization. This paper proposes a reconfigurable in-memory XNOR processing element (PE) for BNNs using single-flux-quantum (SFQ) circuits. Each PE incorporates a memory cell reconfigured by two-dimensional DC control currents to store weights before inference and perform XNOR operations with input features during inference. The proposed PE was fabricated using a 10 kA cm −2 Nb four-layer superconducting process and was experimentally verified to operate correctly, achieving a bias margin of +23.7%–+34.6%. The static power consumption of a single PE is approximately 4 μ W, and the designed array achieves an energy efficiency of about 1.5 × 10⁵ TOPS W −1 . Compared with CMOS–SFQ hybrid PEs controlled by DACs, the proposed design reduces power consumption by up to three orders of magnitude and decreases control-line complexity from quadratic to linear, enabling more efficient scaling. These results demonstrate that the proposed in-memory PE provides an approach toward energy-efficient and scalable superconducting neural network accelerators with on-chip training capability.
- 1 December 2025
- Published by IOP Publishing in Superconductor Science and Technology
- Vol. 38 (12) , 125027
Open AccessReconfigurable in-memory XNOR processing element with external DC control currents for scalable superconducting binary neural networks
Binary neural networks (BNNs) reduce memory demands and computational complexity by replacing multiply–accumulate operations with bitwise exclusive-NOR (XNOR) and popcount. On the other hand, in-memory computing improves efficiency by integrating storage and computation within a single element to minimize data transfer, reduce energy consumption, and enhance area utilization. This paper proposes a reconfigurable in-memory XNOR processing element (PE) for BNNs using single-flux-quantum (SFQ) circuits. Each PE incorporates a memory cell reconfigured by two-dimensional DC control currents to store weights before inference and perform XNOR operations with input features during inference. The proposed PE was fabricated using a 10 kA cm −2 Nb four-layer superconducting process and was experimentally verified to operate correctly, achieving a bias margin of +23.7%–+34.6%. The static power consumption of a single PE is approximately 4 μ W, and the designed array achieves an energy efficiency of about 1.5 × 10⁵ TOPS W −1 . Compared with CMOS–SFQ hybrid PEs controlled by DACs, the proposed design reduces power consumption by up to three orders of magnitude and decreases control-line complexity from quadratic to linear, enabling more efficient scaling. These results demonstrate that the proposed in-memory PE provides an approach toward energy-efficient and scalable superconducting neural network accelerators with on-chip training capability.
- Open Access
- 25 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 35 (8) , 1-6
High-Capacity DFF Cells for Unary Computing Using Single-Flux Quantum Circuits
Due to their high-speed operation and low power consumption, single-flux quantum (SFQ) circuits are promising technologies for integrated circuits. However, current superconducting fabrication limitations impose strict area constraints, challenging the construction of large-scale systems. Unary computing (UC) encodes numerical values by concentrating all “1”s at the beginning of a bit sequence. This approach enables arithmetic operations with fewer logic gates, thereby improving the area efficiency. In this article, we propose a novel high-capacity D flip-flop (HC-DFF) cell for UC-based SFQ circuits that achieves high-density data registration while preserving UC encoding. The HC-DFF employs a cascaded storage loop structure in which each loop sequentially stores one flux quantum, avoiding the margin deterioration associated with storing multiple flux quanta in a single loop. Simulation shows that the proposed HC-DFF achieves a bias margin of –20.8% to +32.7% and supports read/write operations above 100 GHz. We fabricated the HC-DFF using a 10 kA/cm2 Nb four-layer superconducting process and verified its operation through low-speed measurements. The HC-DFF offers enhanced scalability and stability while reducing the number of required Josephson junctions by 83.3% per bit compared to conventional DFF-based shift registers for saving UC signals. For short unary sequences of less than 31 bits, it retains an advantage over storing binary data with equivalent precision using conventional DFF-based shift registers.
- 25 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 35 (8) , 1-6
Open AccessHigh-Capacity DFF Cells for Unary Computing Using Single-Flux Quantum Circuits
Due to their high-speed operation and low power consumption, single-flux quantum (SFQ) circuits are promising technologies for integrated circuits. However, current superconducting fabrication limitations impose strict area constraints, challenging the construction of large-scale systems. Unary computing (UC) encodes numerical values by concentrating all “1”s at the beginning of a bit sequence. This approach enables arithmetic operations with fewer logic gates, thereby improving the area efficiency. In this article, we propose a novel high-capacity D flip-flop (HC-DFF) cell for UC-based SFQ circuits that achieves high-density data registration while preserving UC encoding. The HC-DFF employs a cascaded storage loop structure in which each loop sequentially stores one flux quantum, avoiding the margin deterioration associated with storing multiple flux quanta in a single loop. Simulation shows that the proposed HC-DFF achieves a bias margin of –20.8% to +32.7% and supports read/write operations above 100 GHz. We fabricated the HC-DFF using a 10 kA/cm2 Nb four-layer superconducting process and verified its operation through low-speed measurements. The HC-DFF offers enhanced scalability and stability while reducing the number of required Josephson junctions by 83.3% per bit compared to conventional DFF-based shift registers for saving UC signals. For short unary sequences of less than 31 bits, it retains an advantage over storing binary data with equivalent precision using conventional DFF-based shift registers.
- 30 January 2025
- Published by IOP Publishing in Superconductor Science and Technology
- Vol. 38 (2) , 025017
Design and fabrication of integrator using adiabatic quantum-flux-parametron circuit
We believe that the adiabatic quantum flux parametron (AQFP) circuit, as a novel low-power superconducting circuit technology, has potential application in various fields, including utilizing its high precision to determine the phase of input signals. To improve the precision of AQFP circuits, it is crucial to reduce the gray zone of the system. In this paper, we investigate a digital integration method to reduce the gray zone and enhance the precision of the AQFP circuits. Theoretical analysis demonstrates that the gray zone narrows, and the precision of the system improves as more data points are integrated. Additionally, we establish the quantitative relationship between the reduction in gray zone and the number of integrated data points. To validate these findings, we fabricate AQFP integrator circuits and perform experimental verification. Furthermore, we estimate the relationship between the scale of the integration circuit and the chip area, emphasizing possibilities for various future application scenarios.
- 30 January 2025
- Published by IOP Publishing in Superconductor Science and Technology
- Vol. 38 (2) , 025017
Design and fabrication of integrator using adiabatic quantum-flux-parametron circuit
We believe that the adiabatic quantum flux parametron (AQFP) circuit, as a novel low-power superconducting circuit technology, has potential application in various fields, including utilizing its high precision to determine the phase of input signals. To improve the precision of AQFP circuits, it is crucial to reduce the gray zone of the system. In this paper, we investigate a digital integration method to reduce the gray zone and enhance the precision of the AQFP circuits. Theoretical analysis demonstrates that the gray zone narrows, and the precision of the system improves as more data points are integrated. Additionally, we establish the quantitative relationship between the reduction in gray zone and the number of integrated data points. To validate these findings, we fabricate AQFP integrator circuits and perform experimental verification. Furthermore, we estimate the relationship between the scale of the integration circuit and the chip area, emphasizing possibilities for various future application scenarios.
- 7 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 35 (5) , 1-5
Small-Area Sorting Network Based on Unary Coding Using Single Flux Quantum Circuits
The sorting network is crucial in fields like communication switching, data mining, image processing, neural networks, and others, where the bitonic sorting network (BSN) is an efficient comparison sorting method. It uses the compare-and-swap (CAS) units to compare two input data and decide whether to swap them. The single flux quantum (SFQ) circuits are well known for their high-speed operation and low power consumption, but they are less effective at performing numerical comparisons in CAS units. To address this issue, the unary coding (UC) method, an entropy encoding method that represents the natural number n with n ones followed by zeros, is introduced to implement a smaller SFQ-based BSN. The UC requires only AND and OR gates for CAS unit implementation. Consequently, the 8-input 32-bit UC BSN is two orders of magnitude smaller than the conventional SFQ binary coding BSN with the same precision and achieves approximately 3.3 times better energy efficiency. Additionally, compared to the CMOS-based UC BSN circuit, the energy efficiency improves by 5.8 times, even considering the cooling cost (×400). We also implemented a 4-input BSN circuit, using the 1.0 μm 10-layer Nb SFQ circuit fabrication process, and verified its operation at low speed.
- 7 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 35 (5) , 1-5
Small-Area Sorting Network Based on Unary Coding Using Single Flux Quantum Circuits
The sorting network is crucial in fields like communication switching, data mining, image processing, neural networks, and others, where the bitonic sorting network (BSN) is an efficient comparison sorting method. It uses the compare-and-swap (CAS) units to compare two input data and decide whether to swap them. The single flux quantum (SFQ) circuits are well known for their high-speed operation and low power consumption, but they are less effective at performing numerical comparisons in CAS units. To address this issue, the unary coding (UC) method, an entropy encoding method that represents the natural number n with n ones followed by zeros, is introduced to implement a smaller SFQ-based BSN. The UC requires only AND and OR gates for CAS unit implementation. Consequently, the 8-input 32-bit UC BSN is two orders of magnitude smaller than the conventional SFQ binary coding BSN with the same precision and achieves approximately 3.3 times better energy efficiency. Additionally, compared to the CMOS-based UC BSN circuit, the energy efficiency improves by 5.8 times, even considering the cooling cost (×400). We also implemented a 4-input BSN circuit, using the 1.0 μm 10-layer Nb SFQ circuit fabrication process, and verified its operation at low speed.
- 14 March 2024
- Published by IOP Publishing in Superconductor Science and Technology
- Vol. 37 (4) , 045014
High area efficiency binary neural processing elements with time-domain signals using SFQ circuits
With the increasing processor performance requirements of convolutional neural networks, we believe that research on superconducting neural networks based on single flux quantum (SFQ) circuits holds significant potential. However, the typical multiply and accumulate operations used in neural networks often lead to a significant increase in circuit area due to the limitations of the integration density of the current SFQ circuit fabrication process. In this work, we propose a binary neural processing element (PE) that utilizes time-domain signal expression. Time-domain signals express multi-bit binary data in time information, which reduces the number of data lines to decrease the circuit area. In the proposed binary neural PE, multiplication is achieved through exclusive-NOR gates and addition is realized by controlling the delay of time-domain signals. We designed and simulated a 3 × 3 convolution circuit using 10 Nb layer process with a critical current density of 10 kA cm−2. Compared to the convolution circuit implemented using conventional SFQ logic, the proposed circuit reduces the number of Josephson junctions by approximately 56% and decreases the area by about 60%.
- 14 March 2024
- Published by IOP Publishing in Superconductor Science and Technology
- Vol. 37 (4) , 045014
High area efficiency binary neural processing elements with time-domain signals using SFQ circuits
With the increasing processor performance requirements of convolutional neural networks, we believe that research on superconducting neural networks based on single flux quantum (SFQ) circuits holds significant potential. However, the typical multiply and accumulate operations used in neural networks often lead to a significant increase in circuit area due to the limitations of the integration density of the current SFQ circuit fabrication process. In this work, we propose a binary neural processing element (PE) that utilizes time-domain signal expression. Time-domain signals express multi-bit binary data in time information, which reduces the number of data lines to decrease the circuit area. In the proposed binary neural PE, multiplication is achieved through exclusive-NOR gates and addition is realized by controlling the delay of time-domain signals. We designed and simulated a 3 × 3 convolution circuit using 10 Nb layer process with a critical current density of 10 kA cm−2. Compared to the convolution circuit implemented using conventional SFQ logic, the proposed circuit reduces the number of Josephson junctions by approximately 56% and decreases the area by about 60%.
- 31 January 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 33 (5) , 1-5
Design of Max Pooling Operation Circuit for Binarized Neural Networks Using Single-Flux-Quantum Circuit
For the binary neural network (BNN), We design a max pooling operation circuit (MPOC) using single-flux-quantum circuits based on the multiple data comparators. To continuously compare multiple data and find the maximum value, the multiple data comparator uses the internal state of non-destructive read-out (NDRO) flip-flop and NDRO with complementary outputs to store the maximum value. The MPOC is designed with a gate-level pipeline architecture and uses the internal state to store the maximum value, avoiding the use of feedback loops or tree structures when comparing multiple data. Therefore, the area of the MPOC is reduced by approximately 74% compared to the conventional MPOC using many shift registers with tree structures. The throughput is approximately 21.2 times higher and the area is reduced by 22.1% when compared with the conventional MPOC using feedback loops. We designed an MPOC with 2878 Josephson junctions using a 10 kA/cm $^{2}$ Nb process. The proposed circuit can be used for pooling operations of any size after a 1-channel 3x3 convolutional layer, which accelerates the forward propagation process of binary neural networks (BNNs). Simulation results show that the MPOC can perform pooling operations continuously at the frequency of 50 GHz.
- 31 January 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 33 (5) , 1-5
Design of Max Pooling Operation Circuit for Binarized Neural Networks Using Single-Flux-Quantum Circuit
For the binary neural network (BNN), We design a max pooling operation circuit (MPOC) using single-flux-quantum circuits based on the multiple data comparators. To continuously compare multiple data and find the maximum value, the multiple data comparator uses the internal state of non-destructive read-out (NDRO) flip-flop and NDRO with complementary outputs to store the maximum value. The MPOC is designed with a gate-level pipeline architecture and uses the internal state to store the maximum value, avoiding the use of feedback loops or tree structures when comparing multiple data. Therefore, the area of the MPOC is reduced by approximately 74% compared to the conventional MPOC using many shift registers with tree structures. The throughput is approximately 21.2 times higher and the area is reduced by 22.1% when compared with the conventional MPOC using feedback loops. We designed an MPOC with 2878 Josephson junctions using a 10 kA/cm $^{2}$ Nb process. The proposed circuit can be used for pooling operations of any size after a 1-channel 3x3 convolutional layer, which accelerates the forward propagation process of binary neural networks (BNNs). Simulation results show that the MPOC can perform pooling operations continuously at the frequency of 50 GHz.