Efficient Systolic-Array Redundancy Architecture for Offline/Online Repair
Round 1
Reviewer 1 Report
The manuscript "Efficient Systolic-Array Redundancy Architecture for Offline/Online Repair" is almost a well-written work introducing the benefits of hardware overhead in systolic-arrays for CNN calculation modules. In particular, the authors explained the usefulness of circuit redundancy along both directions of the array, as well as using partition techniques to reduce the size of circuit redundancies. These illustrations will provide readers with a comprehensive knowledge in using a robust-designed systolic-arrays.
There are a few recommendations that I would like to bring to the authors in improving the manuscript:
The redundancy design seems applicable to all FPGA-like circuits, while CNN requires much simpler calculations of co-adding and max-pooling. In this regard, it is probably better to elaborate the detailed mechanisms of the R modules to show/discuss the subtle differences. Uniformity is a requirement in IC manufacturing, and if the circuit redundancy is much different from the other normal modules in the systolic array, it will require a much different manufacturing process in IC layers. Some discussion regarding the influences can be added, so that uniformity can be best insured and the design maintains the overall cost at similar levels. One should also assume that all the R modules for reliability and repairing will not have defects, correct? In the introduction, what CNN has been applied to and what the proposed designs can be applied to should be stated respectively. In particular, readers are very interested to know what type of applications that the authors' designs are fit for, other than the traditional applications in CNN. The reviewer feel that at least compact and independent CNN applications in many emerging areas can benefit from the designs. For example, applications in edge computing ["Edge computing: Vision and challenges." IEEE internet of things journal 3.5 (2016): 637-646.], ["The emergence of edge computing." Computer 50.1 (2017): 30-39.]; laser beam sensing and controls ["Machine learning for improved image-based wavefront sensing." Optics letters 43.6 (2018): 1235-1238.], ["Lossy wavefront sensing and correction of distorted laser beams." Applied Optics 59.3 (2020): 817-824.]; as well as privacy preserving ["LEP-CNN: A Lightweight Edge Device Assisted Privacy-preserving CNN Inference Solution for IoT." arXiv preprint arXiv:1901.04100 (2019).], ["A lightweight privacy-preserving CNN feature extraction framework for mobile sensing." IEEE Transactions on Dependable and Secure Computing (2019).]. The authors are encouraged to find more fitting applications that require compact CNN algorithms with high reliability.Author Response
We have answered to all the comments and uploaded as a file.
Author Response File: Author Response.pdf
Reviewer 2 Report
In this work, the authors proposed a new efficient redundancy architecture to enhance the reliability of the systolic-array architecture. This manuscript is in good shape. Please correct the spelling errors. For example: leaning at line 9. Some sentences are difficult to understanding. Professional English editing is recommended.Author Response
We have answered all the comments and uploaded as a file.
Author Response File: Author Response.pdf