Wei-Kai Cheng

Dr. Wei-Kai Cheng

orcid

Research Focus & Metrics

Academic Career Length

32Years (1995-Present)

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

Representative Publications

View All Publications
1 May 2016 in Microprocessors and Microsystems
Verified
Journal Article
Closed access
WCWei-Kai Cheng
Wei-Kai Cheng
YCYen-Heng Ciou
Yen-Heng Ciou
PSPo-Yuan Shen
Po-Yuan Shen

8 September 2019 in Micromachines
Verified
Journal Article
Open access
PDF file
WCWei-Kai Cheng
Wei-Kai Cheng
PSPo-Yuan Shen
Po-Yuan Shen
XLXin-Lun Li
Xin-Lun Li
Dynamic random access memory (DRAM) circuits require periodic refresh operations to prevent data loss. As DRAM density increases, DRAM refresh overhead is even worse due to the increase of the refresh cycle time. However, because of few the cells in memory that have lower retention time, DRAM has to raise the refresh frequency to keep the data integrity, and hence produce unnecessary refreshes for the other normal cells, which results in a large refresh energy and performance delay of memory access. In this paper, we propose an integration scheme for DRAM refresh based on the retention-aware auto-refresh (RAAR) method and 2x granularity auto-refresh simultaneously. We also explain the corresponding modification need on memory controllers to support the proposed integration refresh scheme. With the given profile of weak cells distribution in memory banks, our integration scheme can choose the most appropriate refresh technique in each refresh time. Experimental results on different refresh cycle times show that the retention-aware refresh scheme can properly improve the system performance and have a great reduction in refresh energy. Especially when the number of weak cells increased due to the thermal effect of 3D-stacked architecture, our methodology still keeps the same performance and energy efficiency.

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.

11 August 2023 in Electronics
Verified
Journal Article
Open access
PDF file
WCWei-Kai Cheng
Wei-Kai Cheng
ZYZih-Ming Yeh
Zih-Ming Yeh
HKHsu-Yu Kao
Hsu-Yu Kao
SHShih-Hsu Huang
Shih-Hsu Huang
In the clock network design, the trade-off between power consumption and timing closure is an important and difficult issue. The clock tree architecture has a shorter wire length and better power consumption, but it is more difficult to achieve timing closure with it. On the other hand, clock mesh architecture is easier to satisfy the clock skew constraint, but it usually has much more power consumption. Therefore, a hybrid clock network architecture that combines both the clock tree and clock mesh seems to be a promising solution. In a normal hybrid mesh/tree structure, a driving buffer is placed in the intersection of mesh lines. In this paper, we propose a novel cross-mesh architecture, and we distribute the buffers to balance the overall switching capacitance, reducing the number of registers connected to a subtree, and the load capacitance of a buffer. With the average dispersion of the overall driving force, our methodology creates small non-zero skew clock trees. In addition, we integrate clock gating, register clustering, and load balancing techniques to optimize clock skew and load capacitance simultaneously. The proposed methodology has four stages: cross-mesh planning, register clustering, mesh line connecting, and load balancing. Experimental results show that our cross-mesh architecture has high tolerance for process variation, and is robust in all the operation modes. Comparing it to the uniform mesh architecture, our methodology and algorithms reduce 28.9% of load capacitance and 80.4% of clock skew on average. Compared to the non-uniform mesh architecture, we also reduce capacitance by 22.4% and skew by 76.7% on average. This illustrates that we can obtain a feasible solution effectively and improve both power consumption and clock skew simultaneously.

1 December 1995 in Proceedings of the 39th conference on Design automation - DAC '02
Verified
Conference Paper
Closed access
WCWei-Kai Cheng
Wei-Kai Cheng
High-level synthesis (HLS) has been successfully targeted towards the digital signal processing (DSP) domain. Both application-specific integrated circuits (ASICs) and application-specific instruction-set processor (ASIPs) have been frequently designed using the HLS approach. Since most ASIP and DSP processors provide multiple addressing modes, and, in addition to classical constraint on the number of function units, registers, and buses, there are many resource usage rules, special considerations need to be paid to the optimizing code generation problem. In this paper we propose three transformation techniques, data management, data ordering, and transformational retiming, for storage optimization during code generation. With these transformations, some scheduling bottlenecks are eliminated, redundant instructions removed, and multiple operations mapped onto a single one. The proposed transformations have been implemented in a software system called Theda:MS. A set of benchmark programs has been used to evaluate the effectiveness of Theda:MS. Measurement on the synthesized codes targeted towards the TI-TMS320C40 DSP processor shows that the proposed approach is indeed very effective.

5 November 2021 in Micromachines
Verified
Journal Article
Open access
PDF file
WCWei-Kai Cheng
Wei-Kai Cheng
XLXiang-Yi Liu
Xiang-Yi Liu
HWHsin-Tzu Wu
Hsin-Tzu Wu
HPHsin-Yi Pai
Hsin-Yi Pai
PCPo-Yao Chung
Po-Yao Chung
Computation of convolutional neural network (CNN) requires a significant amount of memory access, which leads to lots of energy consumption. As the increase of neural network scale, this phenomenon is further obvious, the energy consumption of memory access and data migration between on-chip buffer and off-chip DRAM is even much more than the computation energy on processing element array (PE array). In order to reduce the energy consumption of memory access, a better dataflow to maximize data reuse and minimize data migration between on-chip buffer and external DRAM is important. Especially, the dimension of input feature map (ifmap) and filter weight are much different for each layer of the neural network. Hardware resources may not be effectively utilized if the array architecture and dataflow cannot be reconfigured layer by layer according to their ifmap dimension and filter dimension, and result in a large quantity of data migration on certain layers. However, a thorough exploration of all possible configurations is time consuming and meaningless. In this paper, we propose a quick and efficient methodology to adapt the configuration of PE array architecture, buffer assignment, dataflow and reuse methodology layer by layer with the given CNN architecture and hardware resource. In addition, we make an exploration on the different combinations of configuration issues to investigate their effectiveness and can be used as a guide to speed up the thorough exploration process.