CONNA: Configurable Matrix Multiplication Engine for Neural Network Acceleration
Round 1
Reviewer 1 Report
The manuscript is quite comprehensive and completed their analysis and review with previous work quite well. The demonstrated work is also quite competitive and even better than state of-the-art accelerators and processors. This referee would like to recommend this manuscript as publication since the authors have done a nice job in this manuscript.
Author Response
Response to Reviewer 1 Comments
Thank you for reviewing our paper. We think it’s not enough yet, so we revised this paper. The revised paper was made easier to understand, and the contents were added to the related works and evaluation parts. Also, we corrected grammatical errors and strange expressions in this paper.
Author Response File: Author Response.docx
Reviewer 2 Report
A hardware accelerator for matrix multiplication used in CNN networks is presented. The presentation could be improved to clearly show what are the new ideas and technical contributions of the paper.
The comparison section should be improved. The authors stated that "a speed-up of up to 34.1 compared to that on existing accelerators". I am not sure that this is correct. Also, a more fair comparison with existing systolic array implementations should be done. An existing such systolic array was implemented by Google for CNN networks. There is no problems to adept the dimension of the array to the dimension of the systolic array. This is an important drawback of this paper.
Author Response
Response to Reviewer 2 Comments
Point 1: The comparison section should be improved. This study implemented and compared CNN accelerators under a fixed number of computation units. For a fairer comparison, existing systolic array implementations should be compared. In other words, consider various numbers of computation units.
Response 1: We are so really thank you for your comment. As mentioned in your comment, for a fair comparison, more CNN accelerators should be compared. In the revised manuscript, we compare the systolic array with a large number of PE arrays (Google’s TPU) and a new data flow to improve latency and throughput (Chain-NN). In addition, the comparison with the MAERI to improve the computing unit utilization in SIMD is added. In the previous manuscript, it is not intuitive to understand the comparison section. We have modified the comparison session to make it easy to understand as well as to add comparison results. To help you understand the experimental results, add the number of computation units. Also, the sentence has been modified to clearly understand the session. The revised paper, I think will help you understand.
Point 2: The authors stated that "a speed-up of up to 34.1 compared to that on existing accelerators". I am not sure that this is correct. Also, sufficient background and all relevant references should be included.
Response 2: Thank you for your comment regarding the abstract. In the process of checking this paper, we note that “a speed-up of up to 34.1 compared to that on existing accelerators” was wrong. we corrected a mistake. Also corrected grammatical errors and strange expressions in this paper. In the “Background and Related works” section, the composition of this paper has been changed to make the overall content easier to understand. In addition, it was revised with appropriate references related to the contents of the text, and the existing research to solve the issues was briefly introduced.
Author Response File: Author Response.docx
Reviewer 3 Report
This paper proposed the framework CONNA. It is very interesting, and the results are reliable. However, the language must be improved, and the related work must be better described. Also, the discussion of the results must be available in the manuscript.
Author Response
Response to Reviewer 3 Comments
Point 1: The language must be improved, and related work must be better described.
Response 1: We are so really thank you for your comment. As mentioned in your comment, there are many mistakes. We corrected grammatical errors and strange expressions in this paper. In the “Background and Related works” section, the composition of this paper has been changed to make the overall content easier to understand. In addition, it was revised with appropriate references related to the contents of the text, and the existing research to solve the issues was briefly introduced.
Point 2: The discussion of the results must be available in manuscript.
Response 2: Thank you for your comment regarding the discussion of the results. To discuss the results, in the revised manuscript, first, we add comparison results with the systolic array with a large number of PE arrays (Google’s TPU) and a new data flow to improve latency and throughput (Chain-NN). In addition, the comparison with the MAERI to improve the computing unit utilization in SIMD is added. Second, we have modified the introduction and evolution session to make it easy to understand. Finally, we add a discussion section. This section includes the meaning of the CONNA accelerator and feature work.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
The paper have been improved and can be accepted now.
Reviewer 3 Report
The manuscript is improved, and it can be accepted.