Hsu-Yu Kao

Mr. Hsu-Yu Kao

Department of Electronic Engineering
+1
orcid

Research Focus & Metrics

Academic Career Length

8Years (2019-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
5 July 2022 in Electronics
Verified
Journal Article
Open access
PDF file
HKHsu-Yu Kao
Hsu-Yu Kao
SHShih-Hsu Huang
Shih-Hsu Huang
WCWei-Kai Cheng
Wei-Kai Cheng
To achieve faster design closure, there is a need to provide a design framework for the design of ReRAM-based DNN (deep neural network) accelerator at the early design stage. In this paper, we develop a high-level ReRAM-based DNN accelerator design framework. The proposed design framework has the following three features. First, we consider ReRAM’s non-linear properties, including lognormal distribution, leakage current, IR drop, and sneak path. Thus, model accuracy and circuit performance can be accurately evaluated. Second, we use SystemC with TLM modeling method to build our virtual platform. To our knowledge, the proposed design framework is the first behavior-level ReRAM deep learning accelerator simulator that can simulate real hardware behavior. Third, the proposed design framework can evaluate not only model accuracy but also hardware cost. As a result, the proposed design framework can be used for behavior-level design space exploration. In the experiments, we have deployed different DNN models on the virtual platform. Circuit performance can be easily evaluated on the proposed design framework. Furthermore, experiment results also show that the noise effects are different in different ReRAM array architectures. Based on the proposed design framework, we can easily mitigate noise effects by tuning architecture parameters.

1 December 2019
Verified
Conference Paper
Closed access
WLWei-Hung Lin
Wei-Hung Lin
HKHsu-Yu Kao
Hsu-Yu Kao
SHShih-Hsu Huang
Shih-Hsu Huang
For real-time edge AI applications, there is a need to implement deep neural networks (DNNs) in hardware for high speed. The trade-off between computing accuracy and hardware cost must be made during the implementation of hardware approximation. In this paper, we propose a design framework for DNN hardware approximation. The proposed framework provides a behavior model library of approximate logic circuits (e.g., approximate multipliers) for the designers to utilize them. Moreover, the proposed framework also supports the dynamic fixed-point arithmetic for hardware simulation. To save the required total bit width, we develop an integer length tuning method in the framework to maximize the computing accuracy under a constraint on the total bit width. Experimental results on ICNet show that the proposed framework can achieve high computing accuracy with small hardware cost.

10 November 2021 in Sensors
Verified
Journal Article
Open access
PDF file
YWYui-Kai Weng
Yui-Kai Weng
SHShih-Hsu Huang
Shih-Hsu Huang
HKHsu-Yu Kao
Hsu-Yu Kao
In a CNN (convolutional neural network) accelerator, to reduce memory traffic and power consumption, there is a need to exploit the sparsity of activation values. Therefore, some research efforts have been paid to skip ineffectual computations (i.e., multiplications by zero). Different from previous works, in this paper, we point out the similarity of activation values: (1) in the same layer of a CNN model, most feature maps are either highly dense or highly sparse; (2) in the same layer of a CNN model, feature maps in different channels are often similar. Based on the two observations, we propose a block-based compression approach, which utilizes both the sparsity and the similarity of activation values to further reduce the data volume. Moreover, we also design an encoder, a decoder and an indexing module to support the proposed approach. The encoder is used to translate output activations into the proposed block-based compression format, while both the decoder and the indexing module are used to align nonzero values for effectual computations. Compared with previous works, benchmark data consistently show that the proposed approach can greatly reduce both memory traffic and power consumption.

27 July 2021 in Sensors
Verified
Journal Article
Open access
PDF file
HKHsu-Yu Kao
Hsu-Yu Kao
XCXin-Jia Chen
Xin-Jia Chen
SHShih-Hsu Huang
Shih-Hsu Huang
Convolution operations have a significant influence on the overall performance of a convolutional neural network, especially in edge-computing hardware design. In this paper, we propose a low-power signed convolver hardware architecture that is well suited for low-power edge computing. The basic idea of the proposed convolver design is to combine all multipliers’ final additions and their corresponding adder tree to form a partial product matrix (PPM) and then to use the reduction tree algorithm to reduce this PPM. As a result, compared with the state-of-the-art approach, our convolver design not only saves a lot of carry propagation adders but also saves one clock cycle per convolution operation. Moreover, the proposed convolver design can be adapted for different dataflows (including input stationary dataflow, weight stationary dataflow, and output stationary dataflow). According to dataflows, two types of convolve-accumulate units are proposed to perform the accumulation of convolution results. The results show that, compared with the state-of-the-art approach, the proposed convolver design can save 15.6% power consumption. Furthermore, compared with the state-of-the-art approach, on average, the proposed convolve-accumulate units can reduce 15.7% power consumption.

6 July 2022
Verified
Conference Paper
Closed access
BLBo-Xi Lai
Bo-Xi Lai
SHShih-Hsu Huang
Shih-Hsu Huang
HKHsu-Yu Kao
Hsu-Yu Kao
Computing-in-memories (CIM) is recognized as a useful design technique for eliminating the Von Neumann bottleneck. However, there is a need for circuit designers to determine the configuration (i.e., design parameters) of CIM-based accelerators. Note that the configuration has influences on circuit area, throughput, and energy efficiency. In this paper, we focus on the SRAM CIM-based accelerator design. A reinforcement learning methodology is proposed to assist circuit designers to find the most suitable configuration. Experiment data show that the proposed methodology works well in practice.

8 June 2025 in Electronics
Verified
Journal Article
Open access
PDF file
HKHsu-Yu Kao
Hsu-Yu Kao
LSLiang-Ying Su
Liang-Ying Su
SHShih-Hsu Huang
Shih-Hsu Huang
WCWei-Kai Cheng
Wei-Kai Cheng
ReRAM-based DNN accelerators have emerged as a promising solution to mitigate the von Neumann bottleneck. While prior research has introduced tools for simulating the hardware behavior of ReRAM’s non-linear characteristics, there remains a notable gap in high-level design automation tools capable of efficiently deploying DNN models onto ReRAM-based accelerators with simultaneous optimization of execution time and memory usage. In this paper, we propose a neural network compiler built on the open-source TVM framework to address this challenge. The compiler incorporates both layer fusion and model partitioning techniques to enhance data storage efficiency. The core contribution of our work is an algorithm that determines the optimal mapping strategy by jointly considering layer fusion and model partitioning under hardware resource constraints. Experimental evaluations demonstrate that the proposed compiler adapts effectively to varying hardware resource limitations, enabling efficient storage optimization and supporting early-stage design space exploration.