Research and Implementation of High Computational Power for Training and Inference of Convolutional Neural Networks

Round 1

Reviewer 1 Report

The paper addresses the computational advantages of the CNN model, mainly the training and inference process of CNN by HA devices.  One comparison table can be provided with accuracy against the convergence time for the state of art methods Vs the proposed methodology. There are a few typos, which can be corrected in the final version.  

Author Response

Dear reviewers, first of all, thank you very much for your recognition of our work. According to your suggestion, we also found some problems and improved them. I want to share it with you.

1.We have modified the structure of the abstract section to make the structure of the article more reasonable. We also strengthen the argument for the result.

2.The running times of the programs in this paper are measured using real hardware devices. In this study, the system function is called to obtain the system clock to calculate the program running time.It is added in the fifth subsection of this paper that the device used is "Ultra96-V2" rather than a simple simulation. The test program and this hardware device together constitute the test platform.

3.Some of the tables in this article are not comprehensive. Therefore, we redrew the data table in the conclusion section to make the conclusion more clear and concise.

4.The CPU is programmed through the C language. The GPU is programmed through the python language. This article also considers this issue, so it does not include library loading and data preprocessing when testing the runtime.It is worth mentioning that the PS of MPSoC is also programmed by python language, and the PS performance is a little worse than that of CPU.Intel i5 with 2.5GHz is not a high performance platform. But we also don't compare it with the best MPSoCs, for example, we only use "3EG" instead of "15EG".

5.We face up to our shortcomings and also add a comparison of platform costs in the paper. As we all know, the price of an item is greatly affected by the volume of sales. As the ancients said, "Scarce things are valuable".At present, high-performance MPSoC is monopolized by foreign countries, and its price is relatively high due to the small sales. In recent years, domestic FPgas have developed rapidly, and I believe that the price of MPSoC can be more civilian in the near future.

6.Following your suggestion, we also carefully polished the paper and corrected some grammatical errors. And a comparison with similar works has been added to the discussion section of the paper.

Finally, thank you again for your recognition of our work. This is a great incentive for my research life. I wish you a happy life.

Reviewer 2 Report

The paper introduces the convolutional neural network implementation.
First, the paper has many grammatical errors and writing typos. The paper has to be examined by an expert.
The abstract is too long and it should be rewritten according to the journal writing rules.
The literature review and the aim of the study with the contribution of the paper are limited.
The figures are not well drawn and the algorithms should be given in a standard.
Finally, the manuscript has many cons, due to the above reasons, it is impossible to analyze the article properly.

Author Response

Dear reviewers, thank you very much for your careful review of our paper and your detailed and valuable suggestions. We have deeply realized our shortcomings and modified the following aspects according to your suggestions.

1. As you suggested, we are aware of our shortcomings. I have to admit that English expression is really my weakness. So I followed your suggestion and consulted an English expert to polish the full text of our previous paper , and corrected some grammatical errors.
2. Following your suggestion, we have modified the structure of the abstract section to make the structure of the article more reasonable. We have also adjusted the length of the abstract and introduction as required. I hope you will be satisfied.
3. In view of the research purpose and contribution you mentioned, we have made the following corrections. We redraw the data table in the conclusion section to make the conclusion more clear and concise. A comparison with similar works has been added to the discussion section of the paper.
4. We also redrew some unsightly images.The CPU is programmed through the C language. The GPU is programmed through the python language. This article also considers this issue, so it does not include library loading and data preprocessing when testing the runtime.It is worth mentioning that the PS of MPSoC is also programmed by python language, and the PS performance is a little worse than that of CPU.Intel i5 with 2.5GHz is not a high performance platform. But we also don't compare it with the best MPSoCs, for example, we only use "3EG" instead of "15EG".

Finally, thank you very much for your very useful suggestions. This is crucial to my research life. I wish you a happy life.

Reviewer 3 Report

Some comments and suggestions pertaining to the text of the paper are as follows.

1. The structure of the paper should be improved. The abstract section is enormously long, substantially longer than both the discussion and conclusion sections. Select parts of the abstract's text can be moved to the Introduction section. The text of the last two sections (7 and 8) is a little weak and does not contain important findings.

2. How has the calculations time for the MPSoC been established? Was it determined by simulations or real hardware measurement? The same questions pertain to the power consumption values. In Section 5, the authors mentioned that they constructed a dedicated platform with the Xilinx MPSoC chip. If the platform was tangibly assembled, then the picture of the platform would be valuable. If an evaluation board was used, the name of the board should be mentioned. If only the simulations were performed, this should be honestly admitted.

3. Table 4 contains only the comparison with GPU. How about the CPU? Similar issue pertains to Table 5 - only CPU results were presented. How is the calculations time for GPU?

4. How the CPU and GPU are programmed? Whether Python language (and library) was involved in these cases? If so, the speed comparison might be a bit biased since Python implementations are usually really slow. Additionally, Intel i5 with 2.5GHz is not a high performance platform. Thus, once again, the comparison might be a bit biased.

5. Maybe the cost of all considered platforms could be discussed somewhere (e.g., in Section 7). For example the entire GTX1050 graphic card is more than 2 times cheaper than the single MPSoC XCZU3EG chip.

6. There is no direct comparison (e.g., in terms of architecture, calculations times, etc.) with similar works, e.g., [14] and [15].

7. There are misspellings in the text (e.g., "Inter" or "Intel"?).

Author Response

Dear reviewers, thank you very much for your careful review of our paper and your detailed and valuable suggestions. We have modified the following aspects according to your suggestions.

1.Following your suggestion, we have modified the structure of the abstract section to make the structure of the article more reasonable. We also strengthen the argument for the result.

2.The running times of the programs in this paper are measured using real hardware devices. In this study, the system function is called to obtain the system clock to calculate the program running time.It is added in the fifth subsection of this paper that the device used is "Ultra96-V2" rather than a simple simulation. The test program and this hardware device together constitute the test platform.

3.As you said, some of the tables in this article are not comprehensive. Therefore, we redrew the data table in the conclusion section to make the conclusion more clear and concise.

4.The CPU is programmed through the C language. The GPU is programmed through the python language. This article also considers this issue, so it does not include library loading and data preprocessing when testing the runtime.It is worth mentioning that the PS of MPSoC is also programmed by python language, and the PS performance is a little worse than that of CPU.Intel i5 with 2.5GHz is not a high performance platform. But we also don't compare it with the best MPSoCs, for example, we only use "3EG" instead of "15EG".

5.We face up to our shortcomings and also add a comparison of platform costs in the paper. As we all know, the price of an item is greatly affected by the volume of sales. As the ancients said, "Scarce things are valuable".At present, high-performance MPSoC is monopolized by foreign countries, and its price is relatively high due to the small sales. In recent years, domestic FPgas have developed rapidly, and I believe that the price of MPSoC can be more civilian in the near future.

6.Following your suggestion, a comparison with similar works has been added to the discussion section of the paper. And we also carefully polished the paper and corrected some grammatical errors.

Finally, thank you very much for your very useful suggestions. This is crucial to my research life. I wish you a happy life.

Round 2

Reviewer 2 Report

The revised version of the study is enough to accept. Now, the paper can be accepted as is.

Reviewer 3 Report

 

The paper has been sufficiently improved and can be accepted in the present form.