Overveiw of this Dissertation

Chapter 1 Introduction

1.1 Overveiw of this Dissertation

With ceaseless advances in semiconductor technology shrinkage, the main contributing factors to the increasingly more challenging routing problem include the high number of metal layers, wide range of metal thickness, and complex design rules, thus routability has become a critical issue in VLSI physical design flow. To address the routability issue, global routing plays an important role since global routing bridges the gap between placement and detailed routing. In traditional physical design flow, the global routing stage follows the placement stage to yield a rough routing result for most nets, and then the detailed routing stage based on the rough routing result completes physical routes for every net and finally realizes a detailed routing result. For example, Figs. 1.1(a), (b) and (c) respectively show a placement solution, a global routing result and a detailed routing result for a design, in which the gray rectangles denote macros, the small rectangles denote cells, and red circles denote pins of a net. Figure 1.1 illustrates that global routing identifies a set of routing regions that the net should pass through, and detailed routing finds a physical routing path in these regions.

In this dissertation, a routability-driven placement and routing (P&R) flow (Fig. 1.2) is presented based on the proposed global routing engines and post-placement routability optimizer. The proposed global routing engines can not only cooperate with placers to obtain better routability placement Fig. 1.1 (a) A placement solution; (b) a global routing result; (c) a detailed routing result of a design.

(a) (b) (c)

2

solutions, but can also yield high-quality global routing results so that detailed routers can apply the global routing results to generate good detailed routing results. Compared to the traditional placement and routing flow, the flow shown in Fig. 1.2 pays more attention to the interaction between placement and routing and optimizes the main factors to influence routability such like global congestion, local congestion, wirelength and via count, which can detect and fix the congestion problems in the early stages and thus contributes to faster design closure. The red boxes in Fig. 1.2 highlight the contributions of this dissertation to deal with routability issue, introduced in the following four paragraphs.

To avoid wasting time on routing unroutable designs, a routing congestion estimator (RCE) can help designers to fast judge whether a design is routable in the early stages to speed up the design closure.

Also, a RCE can cooperate with placers to optimize routability. Thus, this dissertation presents a global-routing-based RCE cooperating with placers to improve routability. The proposed global-routing-based RCE can offer more accurate routing congestion estimation than probabilistic-based RCE [35, 36] since global-routing-based RCE can better capture the actual routing behaviors. However, global-routing-based RCE is typically slower than probabilistic-based RCEs.

Because a RCE may be frequently launched in the placement stage, the placement stage would slow down if the launched RCE is not fast enough. Accordingly, the objective of a global-routing-based RCE is to identify an accurate congestion map as fast as possible. This dissertation presents an accurate and fast global-routing-based RCE called Grace. In addition to testing Grace on academic benchmarks, we

Fig. 1.2 Routability-driven placement and routing (P&R) flow.

3

also enhance Grace to fulfill industrial requirements and then apply the enhanced Grace in the industrial design flow.

In the flow of Fig.1.2, RCE reports the congestion information of a placement solution to placers, and then placers can move cells based on the congestion map to improve the placement solution's routability. However, as cells move, the congestion map also changes, thereby degrading the effectiveness to improve the routability of a placement. To resolve this problem, we develop a routability optimizer Ropt that takes a placement solution and optimizes its routability by incremental place-and-route. Ropt always maintains a global routing instance based on the current placement solution. The global routing instance is built on a local-routability-aware model. Therefore, the global routing instance provides both global and local congestion information to guide the placement algorithms. Also, the placement algorithms in Ropt invoke a global routing engine to decide the placed locations for movable cells.

After a placement solution is optimized by Ropt, the proposed global router NCTU-GR 2.0 is used to identify a high-quality global routing result of the placement solution to provide a good blueprint for detailed routing. Generally, the runtime of the detailed routing stage is hundred or thousand times of that of the global routing stage. Good global routing results can diminish the time of detailed routing and promote the final interconnection quality significantly. Because the routing quality of a global routing result can be measured by the total overflow and total wirelength, minimizing overflows and wirelength is the major task for global routing researches [3-21], where overflow means that a region's routing demand exceeds its routing capacity. Compared to other state-of-the-art global routers [6, 11, 16, 17], the proposed NCTU-GR 2.0 can get global routing results with fewer overflows and shorter wirelength. Note that, although Grace can be treated as a light global router, the algorithms used in Grace and NCTU-GR 2.0 are largely different since their purposes are different. Chapter 5 will detail the algorithmic differences between designing a global-routing-based RCE and a global router.

With semiconductor technology shrinkage, the number of metal layers ceaselessly increases. Thus, the problem of planning routing wires on which metal layers becomes more challenging. The layer

4

planning for routing wires impacts the amount of vias, timing, and many manufacturing issues such like antenna effect and double-vias. However, the typical global routing stage ignores these issues and leaves these issues to detailed routing, which may make detailed routers struggle for these issues.

Accordingly, we develop a post-3D-global-routing tool Post3DGR between global routing and detailed routing to refine a given 3D global routing result, which can ease the effort of detailed routers to speed up the design closure. The proposed Post3DGR can reduce the vias, congestion, and wirelength of a given 3D global routing result by re-routing nets and re-planning wires' layers. With some modifications, Post3DGR also can take antenna effect into account.

The rest of this dissertation is organized as follows. Chapter 1 introduces the problem formulation and background of global routing. Chapter 2 presents a fast global-routing-based RCE called Grace whose goal is to identify a satisfactory global routing result to predict routing congestion as fast as possible. Chapter 3 presents an enhanced Grace applied in the industrial flow to consider timing and local congestion, and target congestion ratio. Chapters 4 introduces the proposed incremental place-and-route tool Ropt that can optimize the routability of a given placement solution. Chapter 5 presents a global router called NCTU-GR 2.0 whose objective is to obtain high-quality global routing results in a reasonable runtime to guide detailed routers. A post-3D-global-routing tool Post3DGR is detailed in Chapter 5. Finally, Chapter 7 draws conclusions.