Jonathan Ku
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Jonathan Ku has a total of 32 co-authors. Here are the top co-authors by number of shared publications:
- Qilin Zheng
Department of Electrical and Computer Engineering, Duke University, Durham, NC, USA
5 shared publications - Yiran Chen
Duke University, Durham, North Carolina, USA
5 shared publications - Hai Li
Duke University, Durham, NC, USA
4 shared publications - Ziru Li
Department of Electrical and Computer Engineering, Duke University, Durham, NC, USA
4 shared publications - Brady Taylor
Sandia National Laboratories (SNL-NM), Albuquerque, NM (United States)
3 shared publications - Junyao Zhang
Duke University, Durham, NC, USA
3 shared publications
Latest Publications
- Open Access
- 29 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 791-797
AutoRAC: Automated Processing-in-Memory Accelerator Design for Recommender Systems
- 29 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 791-797
Open AccessAutoRAC: Automated Processing-in-Memory Accelerator Design for Recommender Systems
- 13 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Circuits and Systems
- Vol. 19 (4) , 756-766
MulPi : A Mul ti-class and P atient- I ndependent Epileptic Seizure Classifier With Co-Designed Input-stationary Computing-in-SRAM
Unprovoked seizures have threatened epilepsy patients over 70 million. Automated classification to detect and predict seizures could bring seizure-free lives to epilepsy patients, delivering them from fatal danger and increasing the quality of life. Authentic detection and prediction of seizures require 1) multi-class (Mul) and 2) patient-independent (Pi) classification. Recent implementable chips for seizure classification rarely satisfy the two requirements due to restricted resources in small chips; therefore, high efficiency is imperative along with accuracy. This paper introduces an efficient MulPi chip, fabricated for the first time to simultaneously fulfill multi-class and patient independence, based on a co-design approach. We develop a 5-layer convolutional neural network (CNN), MulPiCNN, with advanced training techniques for lightness and accuracy. At the hardware level, our SRAM-based chip leverages computingin- memory (CIM) for efficiency. The fabricated MulPi chip is distinguished from prior CIMs in two folds, namely ISRW-CIM: a) input-stationary (IS) CIM for resource-saving, and b) rowwise (RW) computing to address a challenge of SRAM CIM, empowered by our novel 2T-Hadamard product unit (HPU). MulPi outperforms state-of-the-art chips with 98.5% sensitivity and 99.2% specificity, classifying in 0.12s and 0.348mm2.
- 13 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Circuits and Systems
- Vol. 19 (4) , 756-766
MulPi : A Mul ti-class and P atient- I ndependent Epileptic Seizure Classifier With Co-Designed Input-stationary Computing-in-SRAM
Unprovoked seizures have threatened epilepsy patients over 70 million. Automated classification to detect and predict seizures could bring seizure-free lives to epilepsy patients, delivering them from fatal danger and increasing the quality of life. Authentic detection and prediction of seizures require 1) multi-class (Mul) and 2) patient-independent (Pi) classification. Recent implementable chips for seizure classification rarely satisfy the two requirements due to restricted resources in small chips; therefore, high efficiency is imperative along with accuracy. This paper introduces an efficient MulPi chip, fabricated for the first time to simultaneously fulfill multi-class and patient independence, based on a co-design approach. We develop a 5-layer convolutional neural network (CNN), MulPiCNN, with advanced training techniques for lightness and accuracy. At the hardware level, our SRAM-based chip leverages computingin- memory (CIM) for efficiency. The fabricated MulPi chip is distinguished from prior CIMs in two folds, namely ISRW-CIM: a) input-stationary (IS) CIM for resource-saving, and b) rowwise (RW) computing to address a challenge of SRAM CIM, empowered by our novel 2T-Hadamard product unit (HPU). MulPi outperforms state-of-the-art chips with 98.5% sensitivity and 99.2% specificity, classifying in 0.12s and 0.348mm2.
- Top Cited Papers
- 6 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Circuits and Systems Magazine
- Vol. 25 (1) , 35-57
A Survey: Collaborative Hardware and Software Design in the Era of Large Language Models
The rapid development of large language models (LLMs) has significantly transformed the field of artificial intelligence, demonstrating remarkable capabilities in natural language processing and moving towards multi-modal functionality. These models are increasingly integrated into diverse applications, impacting both research and industry. However, their development and deployment present substantial challenges, including the need for extensive computational resources, high energy consumption, and complex software optimizations. Unlike traditional deep learning systems, LLMs require unique optimization strategies for training and inference, focusing on system-level efficiency. This paper surveys hardware and software co-design approaches specifically tailored to address the unique characteristics and constraints of large language models. This survey analyzes the challenges and impacts of LLMs on hardware and algorithm research, exploring algorithm optimization, hardware design, and system-level innovations. It aims to provide a comprehensive understanding of the trade-offs and considerations in LLM-centric computing systems, guiding future advancements in AI. Finally, we summarize the existing efforts in this space and outline future directions toward realizing production-grade co-design methodologies for the next generation of large language models and AI systems.
- 6 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Circuits and Systems Magazine
- Vol. 25 (1) , 35-57
Top Cited PapersA Survey: Collaborative Hardware and Software Design in the Era of Large Language Models
The rapid development of large language models (LLMs) has significantly transformed the field of artificial intelligence, demonstrating remarkable capabilities in natural language processing and moving towards multi-modal functionality. These models are increasingly integrated into diverse applications, impacting both research and industry. However, their development and deployment present substantial challenges, including the need for extensive computational resources, high energy consumption, and complex software optimizations. Unlike traditional deep learning systems, LLMs require unique optimization strategies for training and inference, focusing on system-level efficiency. This paper surveys hardware and software co-design approaches specifically tailored to address the unique characteristics and constraints of large language models. This survey analyzes the challenges and impacts of LLMs on hardware and algorithm research, exploring algorithm optimization, hardware design, and system-level innovations. It aims to provide a comprehensive understanding of the trade-offs and considerations in LLM-centric computing systems, guiding future advancements in AI. Finally, we summarize the existing efforts in this space and outline future directions toward realizing production-grade co-design methodologies for the next generation of large language models and AI systems.
- 4 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Artificial Intelligence
- Vol. 1 (1) , 26-36
TFSRAM: A 249.8TOPS/W Timing-to-First-Spike Compute-in-Memory Neuromorphic Processing Engine With Twin-Column SRAM Synapses
Spiking neural networks (SNNs) have shown high efficiency in information processing. Time-to-first-spike (TTFS) coding, which encodes the information to the times of first spikes, has been introduced in SNNs due to its simple data representation. However, there is a lack of studies in synapse mapping and neuron circuits for compute-in-memory (CIM) neuromorphic design for low-power TTFS-based SNN inference. This work presents a CIM neuromorphic processing engine for TTFS-based SNNs, which comprises a 64x64 8T-SRAM array for synapse storage and current-based post-neuron circuits. We propose to map the positive and negative synaptic weights to adjacent columns in the SRAM array and accumulate the membrane potential in the dedicated post-neuron circuits. Two techniques are presented through hardware-software co-design: ”Multi-Level Firing Threshold Adjustment” to mitigate the impact of process variations, and “Timing Threshold Adjustment” to further speed up the computation. The proposed processing engine achieves the energy efficiency 249.8 TOPS/W under 8-bit inputs and signed 4-bit weights.
- 4 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Artificial Intelligence
- Vol. 1 (1) , 26-36
TFSRAM: A 249.8TOPS/W Timing-to-First-Spike Compute-in-Memory Neuromorphic Processing Engine With Twin-Column SRAM Synapses
Spiking neural networks (SNNs) have shown high efficiency in information processing. Time-to-first-spike (TTFS) coding, which encodes the information to the times of first spikes, has been introduced in SNNs due to its simple data representation. However, there is a lack of studies in synapse mapping and neuron circuits for compute-in-memory (CIM) neuromorphic design for low-power TTFS-based SNN inference. This work presents a CIM neuromorphic processing engine for TTFS-based SNNs, which comprises a 64x64 8T-SRAM array for synapse storage and current-based post-neuron circuits. We propose to map the positive and negative synaptic weights to adjacent columns in the SRAM array and accumulate the membrane potential in the dedicated post-neuron circuits. Two techniques are presented through hardware-software co-design: ”Multi-Level Firing Threshold Adjustment” to mitigate the impact of process variations, and “Timing Threshold Adjustment” to further speed up the computation. The proposed processing engine achieves the energy efficiency 249.8 TOPS/W under 8-bit inputs and signed 4-bit weights.
- Open Access
- 23 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-6
ModSRAM: Algorithm-Hardware Co-Design for Large Number Modular Multiplication in SRAM
Elliptic curve cryptography (ECC) is widely used in security applications such as public key cryptography (PKC) and zero-knowledge proofs (ZKP). ECC is composed of modular arithmetic, where modular multiplication takes most of the processing time. Computational complexity and memory constraints of ECC limit the performance. Therefore, hardware acceleration on ECC is an active field of research. Processing-in-memory (PIM) is a promising approach to tackle this problem. In this work, we design ModSRAM, the first 8T SRAM PIM architecture to compute large-number modular multiplication efficiently. In addition, we propose R4CSA-LUT, a new algorithm that reduces the cycles for an interleaved algorithm and eliminates carry propagation for addition based on look-up tables (LUT). ModSRAM is co-designed with R4CSA-LUT to support modular multiplication and data reuse in memory with 52% cycle reduction compared to prior works with only 32% area overhead.
- 23 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Open AccessModSRAM: Algorithm-Hardware Co-Design for Large Number Modular Multiplication in SRAM
Elliptic curve cryptography (ECC) is widely used in security applications such as public key cryptography (PKC) and zero-knowledge proofs (ZKP). ECC is composed of modular arithmetic, where modular multiplication takes most of the processing time. Computational complexity and memory constraints of ECC limit the performance. Therefore, hardware acceleration on ECC is an active field of research. Processing-in-memory (PIM) is a promising approach to tackle this problem. In this work, we design ModSRAM, the first 8T SRAM PIM architecture to compute large-number modular multiplication efficiently. In addition, we propose R4CSA-LUT, a new algorithm that reduces the cycles for an interleaved algorithm and eliminates carry propagation for addition based on look-up tables (LUT). ModSRAM is co-designed with R4CSA-LUT to support modular multiplication and data reuse in memory with 52% cycle reduction compared to prior works with only 32% area overhead.
- Open Access
- 23 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Improving the Efficiency of In-Memory-Computing Macro with a Hybrid Analog-Digital Computing Mode for Lossless Neural Network Inference
Analog in-memory-computing (IMC) is an attractive technique with a higher energy efficiency to process machine learning workloads. However, the analog computing scheme suffers from large interface circuit overhead. In this work, we propose a macro with a hybrid analog-digital mode computation to reduce the precision requirement of the interface circuit. Considering the distribution of the multiplication and accumulation (MAC) value, we propose a nonlinear transfer function of the computing circuits by only accurately computing low MAC value in the analog domain with a digital mode to deal with the high MAC value with smaller possibility. Silicon measurement results show that the proposed macro could achieve 160 GOPS/mm2 area efficiency and 25.5 TOPS/W for 8b/8b matrix computation. The architectural-level evaluation for real workloads shows that the proposed macro can achieve up to 2.92× higher energy efficiency than conventional analog IMC designs.
- 23 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Open AccessImproving the Efficiency of In-Memory-Computing Macro with a Hybrid Analog-Digital Computing Mode for Lossless Neural Network Inference
Analog in-memory-computing (IMC) is an attractive technique with a higher energy efficiency to process machine learning workloads. However, the analog computing scheme suffers from large interface circuit overhead. In this work, we propose a macro with a hybrid analog-digital mode computation to reduce the precision requirement of the interface circuit. Considering the distribution of the multiplication and accumulation (MAC) value, we propose a nonlinear transfer function of the computing circuits by only accurately computing low MAC value in the analog domain with a digital mode to deal with the high MAC value with smaller possibility. Silicon measurement results show that the proposed macro could achieve 160 GOPS/mm2 area efficiency and 25.5 TOPS/W for 8b/8b matrix computation. The architectural-level evaluation for real workloads shows that the proposed macro can achieve up to 2.92× higher energy efficiency than conventional analog IMC designs.