Chirag Sudarshan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Chirag Sudarshan has a total of 33 co-authors. Here are the top co-authors by number of shared publications:
- Norbert Wehn
University of Kaiserslautern-Landau
10 shared publications - Christian Weis
TU Kaiserslautern
8 shared publications - Matthias Jung
Embedded Systems Engineering, Fraunhofer IESE, Kaiserslautern, Germany
6 shared publications - John Paul Strachan
Peter Grünberg Institute (PGI-14), Forschungszentrum Jülich GmbH, Jülich, Germany
4 shared publications - Jan Lappas
Technische Universität Kaiserslautern, Kaiserslautern, Germany
3 shared publications - Lukas Steiner
Microelectronic Systems Design Research Group, University of Kaiserslautern-Landau, Kaiserslautern, Germany
3 shared publications
Latest Publications
- Open Access
- 1 March 2026
- Published by Wiley in Advanced Intelligent Systems
- Vol. 8 (5)
A Flexible and Energy‐Efficient Compute‐in‐Memory Accelerator for Kolmogorov–Arnold Networks
Emerging Kolmogorov–Arnold networks (KANs) replace the linear weights of neural networks with trainable nonlinear functions. This modification is particularly attractive for scientific computing, where KANs can match the accuracy of conventional multilayer perceptrons (MLPs) while reducing model size by up to 100×. However, this efficiency comes at the cost of computationally expensive nonlinear evaluations, unlike conventional MLPs dominated by linear matrix multiplications. We present a flexible and energy‐efficient compute‐in‐memory accelerator tailored for KANs, developed through cross‐layer optimization across algorithm, architecture, circuit, and device levels. The accelerator computes arbitrary nonlinear functions using a single‐read scheme and read‐optimized memory arrays with nonvolatile memristive devices. Our system achieves a lowest energy of 8.69 pJ per KAN function. In terms of energy‐delay product, it provides 1996× improvement over CPUs, 208× over standard MLP‐oriented compute‐in‐memory accelerators, and up to 71× over prior KAN accelerators. These results establish energy‐efficient hardware primitives for implementing advanced nonlinear networks in scientific computing.
- 1 March 2026
- Published by Wiley in Advanced Intelligent Systems
- Vol. 8 (5)
Open AccessA Flexible and Energy‐Efficient Compute‐in‐Memory Accelerator for Kolmogorov–Arnold Networks
Emerging Kolmogorov–Arnold networks (KANs) replace the linear weights of neural networks with trainable nonlinear functions. This modification is particularly attractive for scientific computing, where KANs can match the accuracy of conventional multilayer perceptrons (MLPs) while reducing model size by up to 100×. However, this efficiency comes at the cost of computationally expensive nonlinear evaluations, unlike conventional MLPs dominated by linear matrix multiplications. We present a flexible and energy‐efficient compute‐in‐memory accelerator tailored for KANs, developed through cross‐layer optimization across algorithm, architecture, circuit, and device levels. The accelerator computes arbitrary nonlinear functions using a single‐read scheme and read‐optimized memory arrays with nonvolatile memristive devices. Our system achieves a lowest energy of 8.69 pJ per KAN function. In terms of energy‐delay product, it provides 1996× improvement over CPUs, 208× over standard MLP‐oriented compute‐in‐memory accelerators, and up to 71× over prior KAN accelerators. These results establish energy‐efficient hardware primitives for implementing advanced nonlinear networks in scientific computing.
- Open Access
- 6 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
- Vol. 16 (2) , 483-499
Algorithm-Hardware Implications of Softmax Approximations for In-Memory Computing-Based LLM Accelerators
As large language models (LLMs) continue to scale in both size and particularly in sequence length, the softmax-attention mechanism of the underlying transformer architecture has emerged as a critical bottleneck. While in-memory computing (IMC) based architectures offer efficient acceleration for vector-matrix multiplications (VMMs), integrating softmax remains challenging due to its computational complexity. This work presents a comprehensive analysis of softmax approximations and softmax precision requirements in modern LLMs spanning model sizes from 125M to 405B parameters across several natural language processing (NLP) benchmarks. Based on these findings, we present an IMC-optimized reformulation of softmax for attention that results in negligible accuracy degradation, including (a) transformation of complexity from sequence-length-dependency to the much smaller attention head dimension and (b) merging of softmax components within the adjacent IMC’s VMM operations. We present both digital and analog implementations. Our digital approach reduces area overhead to 81.2 μm2 and energy to 0.69 pJ per input-output pair, resulting in more than 20× lower energy-delay-area product compared to state-of-the-art. Our analog implementation achieves 37.1 μm2 and 0.33 pJ by leveraging a spike-based encoding. These results highlight the importance of rethinking softmax computation for IMC to enable scalable transformer accelerators.
- 6 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
- Vol. 16 (2) , 483-499
Open AccessAlgorithm-Hardware Implications of Softmax Approximations for In-Memory Computing-Based LLM Accelerators
As large language models (LLMs) continue to scale in both size and particularly in sequence length, the softmax-attention mechanism of the underlying transformer architecture has emerged as a critical bottleneck. While in-memory computing (IMC) based architectures offer efficient acceleration for vector-matrix multiplications (VMMs), integrating softmax remains challenging due to its computational complexity. This work presents a comprehensive analysis of softmax approximations and softmax precision requirements in modern LLMs spanning model sizes from 125M to 405B parameters across several natural language processing (NLP) benchmarks. Based on these findings, we present an IMC-optimized reformulation of softmax for attention that results in negligible accuracy degradation, including (a) transformation of complexity from sequence-length-dependency to the much smaller attention head dimension and (b) merging of softmax components within the adjacent IMC’s VMM operations. We present both digital and analog implementations. Our digital approach reduces area overhead to 81.2 μm2 and energy to 0.69 pJ per input-output pair, resulting in more than 20× lower energy-delay-area product compared to state-of-the-art. Our analog implementation achieves 37.1 μm2 and 0.33 pJ by leveraging a spike-based encoding. These results highlight the importance of rethinking softmax computation for IMC to enable scalable transformer accelerators.
- 10 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Artificial Intelligence
- Vol. 2 (2) , 150-161
Efficient Deep Neural Network Training With a Novel 5.3-Bit Block Floating Point Data Format
Low-bit-width data formats offer a promising solution for enhancing the energy efficiency of Deep Neural Network (DNN) training accelerators. In this work, we introduce a novel 5.3-bit data format that groups fixed-point values sharing a common exponent and scaling factor within a block of data. We propose a two-level logarithmic mantissa scaling method, providing a wide dynamic range for the Block Floating Point based data format and mitigating data loss during the conversion of DNN parameters into the proposed format. Evaluation results demonstrate that the proposed data format enables DNN training with an average of 5.3 bits, with minimal accuracy loss and no required modifications to the training process. Additionally, implementation results of proposed data format shows 1.48× reduction in memory footprint compared to 8-bit floating point formats, while addressing the lack of sufficient dynamic range for DNN training in current low-bit-width formats such as 8-bit and below.
- 10 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Artificial Intelligence
- Vol. 2 (2) , 150-161
Efficient Deep Neural Network Training With a Novel 5.3-Bit Block Floating Point Data Format
Low-bit-width data formats offer a promising solution for enhancing the energy efficiency of Deep Neural Network (DNN) training accelerators. In this work, we introduce a novel 5.3-bit data format that groups fixed-point values sharing a common exponent and scaling factor within a block of data. We propose a two-level logarithmic mantissa scaling method, providing a wide dynamic range for the Block Floating Point based data format and mitigating data loss during the conversion of DNN parameters into the proposed format. Evaluation results demonstrate that the proposed data format enables DNN training with an average of 5.3 bits, with minimal accuracy loss and no required modifications to the training process. Additionally, implementation results of proposed data format shows 1.48× reduction in memory footprint compared to 8-bit floating point formats, while addressing the lack of sufficient dynamic range for DNN training in current low-bit-width formats such as 8-bit and below.
- Open Access
- 17 January 2025
- Published by Springer Nature in Research Square
A Kolmogorov–Arnold Compute-in-Memory (KA-CIM) Hardware Accelerator with High Energy Efficiency and Flexibility
Kolmogorov-Arnold Networks (KAN) are an emerging AI model designed for AI+Science applications, offering up to 100x fewer parameters than conventional Multilayer Perceptrons (MLPs). KAN relies on computationally expensive non-linear functions, unlike MLPs, which are dominated by matrix multiplication. This limits KAN's compatibility with energy-efficient hardware accelerators like Compute-in-Memory (CIM), restricting it to energy-inefficient general-purpose chips. To address this gap, we propose KA-CIM, a memory-centric design for energy-efficient computation of KAN inference. We leverage the Piece-Wise Linear (PWL) approximation of non-linear functions to convert complex computations into Multiply-Accumulate (MAC) operations and pre-storing segment parameters, an approach suitable for memory-centric design. A specialized CIM unit dynamically routes inputs to appropriate PWL segments, while a crossbar array retrieves segment slope and intercept to execute the MAC operations. The tile partitioning feature facilitates higher PWL segments for improved accuracy without significant energy penalties. This architectural design facilitates efficient and flexible computation of arbitrary non-linear functions of KANs. Beyond KAN inference, KA-CIM's capability extends to multi-variable equations and derivative computations. KA-CIM achieves energy-delay products 1073x lower than CPU for non-linear operations. When executing KAN, it achieves 77x lower energy-delay product than a 100 TOPS/W CIM accelerator executing MLP for the same task.
- 17 January 2025
- Published by Springer Nature in Research Square
Open AccessA Kolmogorov–Arnold Compute-in-Memory (KA-CIM) Hardware Accelerator with High Energy Efficiency and Flexibility
Kolmogorov-Arnold Networks (KAN) are an emerging AI model designed for AI+Science applications, offering up to 100x fewer parameters than conventional Multilayer Perceptrons (MLPs). KAN relies on computationally expensive non-linear functions, unlike MLPs, which are dominated by matrix multiplication. This limits KAN's compatibility with energy-efficient hardware accelerators like Compute-in-Memory (CIM), restricting it to energy-inefficient general-purpose chips. To address this gap, we propose KA-CIM, a memory-centric design for energy-efficient computation of KAN inference. We leverage the Piece-Wise Linear (PWL) approximation of non-linear functions to convert complex computations into Multiply-Accumulate (MAC) operations and pre-storing segment parameters, an approach suitable for memory-centric design. A specialized CIM unit dynamically routes inputs to appropriate PWL segments, while a crossbar array retrieves segment slope and intercept to execute the MAC operations. The tile partitioning feature facilitates higher PWL segments for improved accuracy without significant energy penalties. This architectural design facilitates efficient and flexible computation of arbitrary non-linear functions of KANs. Beyond KAN inference, KA-CIM's capability extends to multi-variable equations and derivative computations. KA-CIM achieves energy-delay products 1073x lower than CPU for non-linear operations. When executing KAN, it achieves 77x lower energy-delay product than a 100 TOPS/W CIM accelerator executing MLP for the same task.
- Open Access
- 18 December 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Non-idealities and Design Solutions for Analog Memristor-Based Content-Addressable Memories
Memristor-based analog Content Addressable Memories (aCAMs) offer robust parallel pattern look-up capabilities, significantly enhancing the scope of In-Memory Computing applications. This paper presents challenges of these analog circuits, which may occur during the inference, and proposes solutions to overcome them. Precisely, we investigate the impact of temperature-dependent behavior, CMOS process variations and memristor telegraph read noise. We demonstrate that one challenging issue affecting memristors analog computing applications, namely telegraph read noise, is not a significant problem in aCAM. We introduce a framework that accounts for these combined distortions to define variability-aware aCAM windows and estimate the bit resolution of a CAM cell. Using this framework we estimate the bit resolution to 2 bits before applying compensating measures and to 4 bits afterwards. We study how variations affect the inference accuracy of the IRIS classification dataset using our novel torchCAM model. We introduce a streamlined aCAM design featuring a memristor comparator for simplified input-to-reference comparison and a novel cell architecture with two symmetrical memristor comparator units.
- 18 December 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Open AccessNon-idealities and Design Solutions for Analog Memristor-Based Content-Addressable Memories
Memristor-based analog Content Addressable Memories (aCAMs) offer robust parallel pattern look-up capabilities, significantly enhancing the scope of In-Memory Computing applications. This paper presents challenges of these analog circuits, which may occur during the inference, and proposes solutions to overcome them. Precisely, we investigate the impact of temperature-dependent behavior, CMOS process variations and memristor telegraph read noise. We demonstrate that one challenging issue affecting memristors analog computing applications, namely telegraph read noise, is not a significant problem in aCAM. We introduce a framework that accounts for these combined distortions to define variability-aware aCAM windows and estimate the bit resolution of a CAM cell. Using this framework we estimate the bit resolution to 2 bits before applying compensating measures and to 4 bits afterwards. We study how variations affect the inference accuracy of the IRIS classification dataset using our novel torchCAM model. We introduce a streamlined aCAM design featuring a memristor comparator for simplified input-to-reference comparison and a novel cell architecture with two symmetrical memristor comparator units.
- Open Access
- 3 October 2022
- Published by Association for Computing Machinery (ACM)
- p. 1-7
A Framework for Formal Verification of DRAM Controllers
- 3 October 2022
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessA Framework for Formal Verification of DRAM Controllers
- 14 August 2022
- Published by Springer Nature
- p. 362-379
A Critical Assessment of DRAM-PIM Architectures - Trends, Challenges and Solutions
- 14 August 2022
- Published by Springer Nature
- p. 362-379
A Critical Assessment of DRAM-PIM Architectures - Trends, Challenges and Solutions
- 25 April 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
- Vol. 12 (2) , 367-380
A Weighted Current Summation Based Mixed Signal DRAM-PIM Architecture for Deep Neural Network Inference
Processing-in-Memory (PIM) is an emerging approach to bridge the memory-computation gap. One of the major challenges of PIM architectures in the scope of Deep Neural Network (DNN) inference is the implementation of area-intensive Multiply-Accumulate (MAC) units in memory technologies, especially for DRAM-based PIMs. The DRAM architecture restricts the integration of DNN computation units near the area optimized commodity DRAM Sub-Array (SA) or Primary Sense Amplifier (PSA) region, where the data parallelism is maximum and the data movement cost is minimum. In this paper, we present a novel DRAM-based PIM architecture that is based on bit-decomposed MAC operation and Weighted Current Summation (WCS) technique to implement the MAC unit with minimal additional circuitry in the PSA region by leveraging on mixed-signal design. The architecture presents a two-stage design that employs light-weight current mirror based analog units near the SAs in the PSA region, whereas all the other substantial logic is integrated near the bank peripheral region. Hence, our architecture attains a balance between the data parallelism, data movement energy and area optimization. For an 8-bit CNN inference, our novel 8Gb DRAM PIM device achieves a peak performance of 142.8GOPS while consuming a power of 756.76mW, which results in an energy efficiency of 188.8GOPS/W. The area overhead of such an 8Gb device for a 2ynm DRAM technology is 12.63% in comparison to a commodity 8Gb DRAM device.
- 25 April 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
- Vol. 12 (2) , 367-380
A Weighted Current Summation Based Mixed Signal DRAM-PIM Architecture for Deep Neural Network Inference
Processing-in-Memory (PIM) is an emerging approach to bridge the memory-computation gap. One of the major challenges of PIM architectures in the scope of Deep Neural Network (DNN) inference is the implementation of area-intensive Multiply-Accumulate (MAC) units in memory technologies, especially for DRAM-based PIMs. The DRAM architecture restricts the integration of DNN computation units near the area optimized commodity DRAM Sub-Array (SA) or Primary Sense Amplifier (PSA) region, where the data parallelism is maximum and the data movement cost is minimum. In this paper, we present a novel DRAM-based PIM architecture that is based on bit-decomposed MAC operation and Weighted Current Summation (WCS) technique to implement the MAC unit with minimal additional circuitry in the PSA region by leveraging on mixed-signal design. The architecture presents a two-stage design that employs light-weight current mirror based analog units near the SAs in the PSA region, whereas all the other substantial logic is integrated near the bank peripheral region. Hence, our architecture attains a balance between the data parallelism, data movement energy and area optimization. For an 8-bit CNN inference, our novel 8Gb DRAM PIM device achieves a peak performance of 142.8GOPS while consuming a power of 756.76mW, which results in an energy efficiency of 188.8GOPS/W. The area overhead of such an 8Gb device for a 2ynm DRAM technology is 12.63% in comparison to a commodity 8Gb DRAM device.
- Open Access
- 22 November 2021
- Published by Springer Nature in Journal of Signal Processing Systems
- Vol. 93 (12) , 1467
Correction to: Efficient Hardware Architectures for 1D- and MD-LSTM Networks
- 22 November 2021
- Published by Springer Nature in Journal of Signal Processing Systems
- Vol. 93 (12) , 1467
Open AccessCorrection to: Efficient Hardware Architectures for 1D- and MD-LSTM Networks
- 22 March 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems II: Express Briefs
- Vol. 68 (5) , 1615-1619
A Novel DRAM Architecture for Improved Bandwidth Utilization and Latency Reduction Using Dual-Page Operation
Emerging memory-intensive applications require a paradigm shift from processor-centric to memory-centric computing. The performance of state-of-the-art computing systems and accelerators designed for such applications is not limited by the processing speed but rather by the limited DRAM bandwidth and long DRAM latencies. Although, the interface frequency of new commodity DRAM memories is constantly increasing to achieve a higher peak bandwidth, this bandwidth cannot be fully utilized due to long internal latencies. Page-misses (i.e. change of active row) are one of the key contributors to long access latencies that result in a low bandwidth utilization. In this paper, we propose a novel DRAM sub-array level architecture referred to as Dual-Sense-Amplifier (DSA) that masks the page-miss latency by allowing individual sub-arrays or banks to flexibly open two rows concurrently. Additionally, the DSA architecture design aim to be fully compatible with any type of commodity DRAM architecture (e.g. HBM, GDDR, etc.) and to retain most of its circuit designs. The area overhead of DSA for an 8 Gb device is 9.6% compared to a commodity DRAM device. On average, a DSA 8 Gb device improves the bandwidth utilization by 17.53% and reduces the average response latency by 31.59 ns compared to commodity DRAM devices operated at 800 MHz. Moreover, we demonstrate that the energy consumption overhead of a DSA DRAM in comparison to a commodity DRAM is negligible.
- 22 March 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems II: Express Briefs
- Vol. 68 (5) , 1615-1619
A Novel DRAM Architecture for Improved Bandwidth Utilization and Latency Reduction Using Dual-Page Operation
Emerging memory-intensive applications require a paradigm shift from processor-centric to memory-centric computing. The performance of state-of-the-art computing systems and accelerators designed for such applications is not limited by the processing speed but rather by the limited DRAM bandwidth and long DRAM latencies. Although, the interface frequency of new commodity DRAM memories is constantly increasing to achieve a higher peak bandwidth, this bandwidth cannot be fully utilized due to long internal latencies. Page-misses (i.e. change of active row) are one of the key contributors to long access latencies that result in a low bandwidth utilization. In this paper, we propose a novel DRAM sub-array level architecture referred to as Dual-Sense-Amplifier (DSA) that masks the page-miss latency by allowing individual sub-arrays or banks to flexibly open two rows concurrently. Additionally, the DSA architecture design aim to be fully compatible with any type of commodity DRAM architecture (e.g. HBM, GDDR, etc.) and to retain most of its circuit designs. The area overhead of DSA for an 8 Gb device is 9.6% compared to a commodity DRAM device. On average, a DSA 8 Gb device improves the bandwidth utilization by 17.53% and reduces the average response latency by 31.59 ns compared to commodity DRAM devices operated at 800 MHz. Moreover, we demonstrate that the energy consumption overhead of a DSA DRAM in comparison to a commodity DRAM is negligible.