HubRouter: Learning Global Routing via Hub Generation and Pin-hub Connection

Summary

This paper proposes a two-phase learning framework called HubRouter for global routing in chip design. Different from previous works that directly generate routes from chip images, which potentially cause inconnectivity, this paper proposes to generate hubs representing tiles in the first phase and then construct RMST with hubs by an actor-critic model. The experimental results show the effectiveness in terms of higher correctness, and shorter WL compared with SOTA model.

Strengths

1. the paper firstly proposes hubs for the global routing task in chip design, transferring the pin-pin problem into hub-pin problems, which can passively avoid the in-connectivity problem when regarding the routing design process as image generation. 2. When dealing with large chips, HubRouter is the best performer among SOTAs regarding both correctness and wire length. 3. The proposed model can have good scalability with GAN or VAE as their generative model for the hub generation phase while maintaining the property of making wire length short and overflow less. 4. In the hub-generation phase, besides hubs, it also generates routes and masks to avoid noises brought by the generative model itself since the noises can greatly impact the results. Especially, the stripe mask can be greatly helpful for complicated cases.

Weaknesses

1. The overflow still exists for those generated global routing even though the proposed model reduces overflow better than PRNet. 2. Even though HubRouter can get a good performance on replicating the known facts, it does not discuss the quality of the generated routes regarding congestion, and it possibly cannot generate novel routes since it has limited knowledge of routing design.

Questions

Q1. It is unclear how the stripe mask is generated for chips. Q2. Why not define pixels as hubs that only connect two neighbors on the same row or columns, i.e., $r_{(i-1)j}+r_{(i+1)j}+r_{i(j-1)}+r_{i(j+1)}=2$. Without points between hubs defined in the paper, how does the second phase generate the routes in the right direction?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The authors discuss the limitations of the lack of training data about hub generation, and the model is not an end-to-end model. Still, it also should discuss how the ground truth comes and the potential impact of adopting such methods when generating it.

Summary

The paper focuses on the generative global routing tasks and mainly ensures the connectivity of generated routes via a two-stage framework. In the first phase, the approach involves a typical generative task, which exploits multi-task learning to promote the generation quality and utilizes a trick called stripe mask to decrease some redundant noise points. In the second phase, the work is formulated as an RSMT construction problem and addresses this problem by REST. The authors show that with correctly generated hubs, the RSMT construction can be solved with less time.

Strengths

+ The structure that generates hubs first and then connects them with pins is novel and interesting, and is reasonable to guarantee the connectivity of routes in the global routing. + The motivation is clear, and the authors show in Table 2 that the so-called `unconnectivity’ caused by existing generative global routing algorithms (PRNet) is severe, but I would have preferred the authors to also refer to it in the introduction part to strengthen the motivation. + The proposed approach performs better than other generative global routing algorithms in several metrics, especially the connectness rate is 100% and the running time of HubRouter (GAN) is much less than PRNet (GAN). + The authors also give some applications other than global routing to show the generality of the proposed approach.

Weaknesses

- The approach is clearly divided into two different phases, but the running time shown in the experiment seems to be combined. The authors could show both generation time and connection time to display the time overhead in either phase. - Some possible typos: The $r_{(n+1)j}$ should be $r_{(m+1)j}$ in Definition 1.

Questions

- Whether the nets in ISPD98 (Table 3) are routed sequentially? If so, what’s the performance of HubRouter when nets are concurrently routed?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

NA

Summary

This paper investigates the issue of global routing in VLSI systems and introduces HubRouter, a method that initially generates hubs and subsequently connects them to pins. In the first phase, the authors explored different generative models. In the second phase, the authors employs an actor-critic model to generate a final routing.

Strengths

1. The paper is well-written and easy to follow. 2. The authors conducted many experiments under different baselines.

Weaknesses

1. The two phases are independent, meaning that the feedback from the second phase does not influence the hub generation, potentially leading to suboptimal results. 2. The authors tried three generative models (VAE, DPM and GAN) in the paper. But the paper lacks clarity on which model should be used in specific situations. 3. The performance improvement achieved is marginal.

Questions

1. In the first phase, the authors introduced three objectives: the hub, route, and mask, with the latter two serving as auxiliary designs. Given that the goal of the route is similar to that of the second phase, it would be worth considering why the authors did not design the two phases as a loop, utilizing the second evaluation result as feedback for the first generation. Furthermore, it would be helpful to understand the differences between the results of the route and the second phase. Besides, the design of mask is rather hand-craft. Why the threshold is 1/2 in each row and each colum? Why the authors only evaluate the row and column, not the density of a specific rectangle area? 2. In the experiments, it seems that the three generative models (VAE, DPM, and GAN) have their own advantages and disadvantages. For instance, GAN achieves the best WLs on six datasets, DPM has the best WL on the remaining two datasets, and VAE is the fastest. How should one select the model when presented with a new dataset? 3. Considering the absolute value of WL, which is the primary evaluation metric, could the authors explain the real-world benefits of improving WL by 1%?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Overall, I believe that the integration of the generative model and reinforcement learning algorithm could be more elegant, potentially leading to further performance improvements.

Summary

This paper presents a new two-phase learning approach, called HubRouter, to address the issue of unconnectivity in the generated routes of global routing (GR) tasks in VLSI systems. It has two steps. Firstly, a deep generative model generates a 'hub,' which acts as a key point in the route; then secondly, HubRouter involves an actor-critic model-based RSMT construction module to connect the hubs. This shift from a pin-pin connection to a hub-pin connection method solves the unconnectivity problem in generative approaches. The HubRouter system ensures all generated routes are connected, eliminating the need for time-consuming post-processing. Experimental results show that HubRouter outperforms other state-of-the-art generative global routing models in wirelength, overflow, and time efficiency. It also finds application in RSMT construction and interactive path replanning, demonstrating its versatility.

Strengths

The paper introduces a novel approach, HubRouter, to global routing, proposing a hub generation and hub-pin connection scheme that effectively addresses the challenge of route unconnectivity. The experimental results show HubRouter outperforming existing generative global routing models in terms of wirelength, overflow, and time efficiency. The approach is very general. The authors show that the approach can also be applied to RSMT construction and interactive path replanning, demonstrating the versatility of their method.

Weaknesses

The challenge of connectivity problem is unclear. Specifically, it introduces a novel approach to tackling the unconnectivity problem in global routing, fails to establish the significance and relevance of this problem adequately. It is helpful to elaborate more about the difficulty and significance of connectivity problem. The effectiveness of the second phase is dependent on the quality of the hubs generated in the first phase. If the generative models do not create effective hubs, the entire approach could be compromised.

Questions

Overall, this paper solves the unconnectivity problem in the generated routes of global routing (GR) tasks in VLSI systems. The experimental results are promising. The efficiency of the proposed model is verified both theoretically and experimentally. However, the reviewer has two concerns: 1. The challenge of connectivity problem is unclear. It is helpful to elaborate more about the difficulty and significance of connectivity problem. 2. The effectiveness of the second phase is dependent on the quality of the hubs generated in the first phase. If the generative models do not create effective hubs, the entire approach could be compromised.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

None

Thanks

Thank you for taking the time to read our rebuttal and maintaining your positive score. We greatly appreciate your support of our work.

Thanks for authors' response

Thanks for the clarifications. Most of my concerns are addressed. After reading the additional materials and discussions, I think this paper shows good potential to assist chip design in terms of WL, correctness rate, and efficiency. However, as a generative model, it mainly generates routes within the training set knowledge. I would like to keep my current score.