A Network Simulator for the Estimation of Bandwidth Load and Latency Created by Heterogeneous Spiking Neural Networks on Neuromorphic Computing Communication Networks^†

Round 1

Reviewer 1 Report

This paper presents the design of a traffic modelling tool for neuromorphic platforms used to simulate biologically realistic spiking neural networks. The results are useful and interesting.

The methodology is sound in so far as it evaluates the average traffic loads on individual links, but in practice the traffic in such networks is often stochastic, having peak spike rates that greatly exceed average values. It is not possible to evaluate these peaks without carrying out a full simulation, but some discussion of peak vs average loads should be included?

The paper discusses 3 casting approaches - UC, LMC and MC - and states that MC "comes at the cost of higher router complexity and latency". As MC is, indeed, shown in the paper to reduce network load by an order of magnitude, it would be useful to quantify these costs, at least approximately?

Mapping techniques are also discussed, but only very simple algorithms (random, sequential, area grouping) are considered. Might more sophisticated algorithms such as clustering based upon degree of connectivity/local traffic density yield further improvements?

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

Comments on jlpea-1601718:

This is an interesting report, the authors report on the development of a network simulator (in detail, the simulation implementation in Python to analyze the communication bandwidth and latency between processing elements in a neuromorphic computing system) for investigations of the network load and latency of different network topologies and communication protocols in neuromorphic computing. Compared to state-of-the-art network models and simulators, they demonstrated the ability to simulate the impact of heterogeneous neural networks and the neuron mapping algorithms. Upon investigations of the bandwidth load and the level of detail reached in neural network, authors show the impact of heterogeneous connectivity, for both small-scale and large-scale tests, and the influence from different neuron mapping algorithms, i.e., impact on the latency as well as the node size. The simulation experiment is well designed, and the manuscript is well written. However, the following questions should be claimed clearly before its publication.

1. First, as the authors stated in the original manuscript, this paper is an extended version of our paper published in International Symposium of Embedded Multicore/Many-core Systems on Chip (MCSoC-2021), hence authors are required to elaborate the current manuscript’s research work extension novelty and/or complementary enrichment points that are unique to the previously published on MCSoC-2021, in particular, the correlation aiming at the development of a simulation platform capable to simulate large scale BNN, with biological connectivity levels, 100 times faster than real time; the statement of a pure simulator implementation and verification, by the analysis of a large scale, heterogeneous NN is not well clarified;
2. In the Introduction, Background and motivation sections, authors should include the detailed description of the many-core systems with an overlaying interconnect communication network, e.g., the latest advances of similar homogeneous and heterogeneous connectivity model(s);
3. How is the value of connection probability ε and the matrix in table 1 determined? Upon utilizing the different models in the text use 0.048, 0.0016, etc., will it affect the final performance?
4. Any reasons for missing 100 250 data for Randomly Mapped Multi-Area Model?
5. Latest and highly correlated references should be cited and no less than 40 typically for a more complete article report structure.
6. What is the difference in hardware usage between Homogeneous Connectivity Model and Heterogeneous Connectivity Model changes?
7. The language and presentation of this manuscript requires to be improved significantly, e.g., the typo of page 2 line 26, “∼000 synapses per neuron”; line 43：it's ->its; line 55: concepts-> concept; Line 100: a->an; Line 103: AER-adress->AER-address; line 124: represents-> represent; 136: represent->represented; 226: is -> are; 260/267: send->sent; 334: cleary->clearly; 392: is->are.

In summary, this is a well-written manuscript with mostly substantiated conclusions, which will be of interest to the readers of Journal of Low Power Electronics and Applications. The manuscript is suitable for publication in Journal of Low Power Electronics and Applications after the compulsory revisions.

Comments for author File: Comments.docx

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

This paper describes the functionalities of a network simulator dedicated to help design networks for neuromorphic hardware. The specific constraints of this application domain are well taken into account and the presentation of the results is clear and understandable. The paper brings significant novelty to the state-of-the-art. 

The part of the paper that needs further developments is the description of the simulator itself. The authors only say that it is written in Python. There is no word on the software design of the simulator, nor any reference to the source code. Thus the reader has no way to check that the described results are sound and just has to believe the authors from their own word. A discussion of the software architecture, the design trade-offs, the performance characteristics, the scalability possibilities are necessary for the reproducibility of such research.

Typos:

• line 136: are representED
• line 216: a connection existS

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 4 Report

This manuscript presents a spiking neural network simulator that has been specifically developed for the analysis of network load and latency of different network topologies and protocols in neuromorphic computing communications. Communication networks for neuromorphic computing differ significantly from traditional communication networks. In neuromorphic computing communications, there are a large number of small information packets where latency is the most critical part. To overcome the challenges posed by traditional communication networks, research such as the proposed network simulator helps researchers in designing the communication protocols for neuromorphic computing hardware. One of the unique contributions of this manuscript is that the authors have also explained how the network simulator can be used to simulate the impact of communication networks for heterogeneous neural networks. The manuscript presents a thorough evaluation of the proposed network simulator.

One minor comment regarding the paper, authors can extend the result section by providing an evaluation of their network simulator for more standard spiking neural network algorithms, such as STDP unsupervised learning on the MNIST dataset. At present, the experiments were done on a uniform randomly connected neural network. An evaluation of more standard algorithms such as STDP models will give users an idea of how the network simulator would perform when implementing popular techniques.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

No further comments, the authors have adequately taken into account my previous remarks.