WorksComputer Science
CU-3D: A Scalable Legality-Aware RTL-to-GDS Flow for 3D-IC Physical Design
Abstract
Face-to-face (F2F) hybrid bonds shorten interdie connections, but data, clock, and power-grid (PG) connections must be realized together. CU-3D is a RTL-to-GDS physical design flow based on commercial EDA tools. It derives role-compatible hybrid-bond (HB) sites from a 17-metal ASAP7-based technology model, complete HB footprints, and constructed PG geometry. To coordinate partitioning with placement flexibility, Scaled refines an inherited 2D placement, whereas Fresh places the receiving die first and uses its pin locations and clock loads to guide placement of the other die. Data and clock HB assignments are jointly optimized while retaining alternative data sites and reserving capacity for subsequent PG assignment. The selected HB sites and stacks are preserved during 3D clock-tree synthesis, supply-aware cell insertion, and routing. Across eleven designs, including a 133-macro processor, the flow achieves zero backend DRC violations and passes signal-connectivity checks. Nine paired designs admit lower-footprint, lower-power implementations; four improve setup WNS. On JPEG, early joint planning gains 87 ps and mobility-aware partitioning gains 346 ps. Path analysis explains why cell-delay savings can outweigh net-delay increases.

Figure 1. Legality is decided early and detected late. (a) A clear center has an obstructed HB footprint. (b) A DATA request blocks CLOCK access to site although is available. (c) A free row lacks reachable VSS. Filled/open markers show decision/detection stages.
Introduction
Face-to-face (F2F) hybrid bonding connects dies through their interconnect faces [1]. Submicron hybrid bonds (HBs) replace long planar wires [2], promising lower capacitance, shorter critical paths, and smaller footprints. A short vertical connection still needs feeder wires connecting cell pins to HB sites on both dies, power for the surrounding logic, and clocks for its registers. Partitioning and placement identify potential gains; a powered, clocked, and routed layout determines whether those gains survive physical implementation.
Prior flows advance 2D-engine reuse [3, 4, 5], terminal assignment [6, 7], clocks [8], and power delivery [9]. Their integration poses a cross-stage problem: a partitioner sees crossing nets, an allocator terminals, and a router conductor stacks. Optimizing these representations separately can consume resources needed later. A
This work is licensed under a Creative Commons Attribution 4.0 International License.
power grid (PG) can remove a selected landing, and clock construction can require sites already occupied by data bonds. Once placement and interface choices are fixed, ordinary routing has limited freedom to resolve the conflict. The challenge is to coordinate decisions before they become physical commitments.
Figure 1 illustrates three cases where locally feasible decisions lead to downstream conflicts. A clear grid center does not imply that the complete HB footprint fits. A DATA assignment can consume a CLOCK group’s only reachable site even when DATA has another choice. A buffer can fit a legal row yet lack a VSS connection. These are different manifestations of the same issue: the resource accepted by one stage is insufficient for the next. The interface is a shared, role-differentiated resource whose capacity is set by implemented geometry. Legality belongs to the whole decision chain.
CU-3D organizes Define–Decide–Deliver around three principles (Figure 2). Resource realism derives usable sites from complete footprints, macro exclusions, and actual PG landings. Freedom-matched commitment aligns partition capacity with placement mobility: inherited organization is refined where useful, while native placement can expose receiving-die demand before the second die commits. DATA alternatives remain available during CLOCK selection, with capacity reserved for PG. Decision preservation then protects accepted sites, stacks, and clock ownership
Figure 2. Define–Decide–Deliver: a shared F2F model defines resources, placement and allocation decide their use, and physical realization preserves the decisions. Fresh steps 1–6 follow § 7.1.

Figure 1 (PDF p. 1). Legality is decided early and detected late. (a) A clear center has an obstructed HB footprint. (b) A DATA request blocks CLOCK access to site although is available. (c) A free row lacks reachable VSS. Filled/open markers show decision/detection stages.
Introduction
Face-to-face (F2F) hybrid bonding connects dies through their interconnect faces [1]. Submicron hybrid bonds (HBs) replace long planar wires [2], promising lower capacitance, shorter critical paths, and smaller footprints. A short vertical connection still needs feeder wires connecting cell pins to HB sites on both dies, power for the surrounding logic, and clocks for its registers. Partitioning and placement identify potential gains; a powered, clocked, and routed layout determines whether those gains survive physical implementation.
Prior flows advance 2D-engine reuse [3, 4, 5], terminal assignment [6, 7], clocks [8], and power delivery [9]. Their integration poses a cross-stage problem: a partitioner sees crossing nets, an allocator terminals, and a router conductor stacks. Optimizing these representations separately can consume resources needed later. A
This work is licensed under a Creative Commons Attribution 4.0 International License.
power grid (PG) can remove a selected landing, and clock construction can require sites already occupied by data bonds. Once placement and interface choices are fixed, ordinary routing has limited freedom to resolve the conflict. The challenge is to coordinate decisions before they become physical commitments.
Figure 1 illustrates three cases where locally feasible decisions lead to downstream conflicts. A clear grid center does not imply that the complete HB footprint fits. A DATA assignment can consume a CLOCK group’s only reachable site even when DATA has another choice. A buffer can fit a legal row yet lack a VSS connection. These are different manifestations of the same issue: the resource accepted by one stage is insufficient for the next. The interface is a shared, role-differentiated resource whose capacity is set by implemented geometry. Legality belongs to the whole decision chain.
CU-3D organizes Define–Decide–Deliver around three principles (Figure 2). Resource realism derives usable sites from complete footprints, macro exclusions, and actual PG landings. Freedom-matched commitment aligns partition capacity with placement mobility: inherited organization is refined where useful, while native placement can expose receiving-die demand before the second die commits. DATA alternatives remain available during CLOCK selection, with capacity reserved for PG. Decision preservation then protects accepted sites, stacks, and clock ownership PG: constructed power grid; –: unspecified scope. MoL/LoL: memory-/logic-on-logic. DRC counts use each work’s backend checks.
through construction. Together, these principles connect early optimization to the resources that clocks, power, and routing will actually use.
Our contributions are:
Executable F2F resources: complete footprints and dual-die PG landings define role-compatible sites (§ 5).
Freedom-matched placement: mobility-aware capacity supports Scaled inheritance and receiving-die-first Fresh placement (§ 6).
Scalable joint allocation: DATA alternatives remain movable during CLOCK selection, followed by residual PG matching (§ 8).
Legality-preserving realization: receiver-first CTS and powered insertion deliver zero backend DRC and signal-connectivity errors across eleven designs, up to 133 macros (§ 9–§ 10).
Related Work
2D-engine reuse. Shrunk-2D [3], Compact-2D [4], and Snap-3D [5] develop tier-splitting surrogates; Cascade2D [13] and Pin-3D [14] address monolithic 3D physical synthesis. Macro-3D [10] addresses heterogeneous macro-on-logic F2F designs.
Assignment and electrical optimization. BTAssign [6], AnchorGrid [7], and Kim and Park [15] improve terminal assignment; differentiable partitioning [16] penalizes terminal-congested cuts. FlexiCTS [8] coordinates clock buffers and bonds; DPO-3D [9] and Zhu et al. [17] address power delivery. Distill3D [18], timing-driven placement [19], and DiffCCD [20] improve timing prediction and clock–data optimization.
Complete flows. Hetero-3D [11] couples mixed-node floorplanning and power delivery. Open3DFlow [21, 22], Open3DBench [12], and RosettaStone [23] support reproducible evaluation. Patnaik et al. [24] report DRC-clean RDL-based F2F layouts. Open3DBench’s logic-on-logic flow uses ideal clocks and reports residual DRC. CU-3D builds a shared inventory of HB sites, dual-die PG landings, and supply-reachable insertion regions from technology rules and the evolving layout. Constructed clocks, PG, and stacks undergo joint DRC and connectivity checks (Table 1).
| Method | Interface representation | Clock construction | PG | Macros | Reported DRC |
| Shrunk-2D [3] | Monolithic 3D / MIVs | 2D CTS, tier split | – | Yes | Clean |
| Compact-2D [4] | Combined metal stack | 2D CTS, tier split | – | Yes | – |
| Snap-3D [5] | Tier rows / F2F vias | Combined CTS | – | Yes | – |
| Macro-3D [10] | Macro-on-logic | Measured clock tree | – | Yes | – |
| Hetero-3D [11] | Mixed-node F2F | CTS | Built | Yes | – |
| BTAssign [6] | Legal HB mesh | – | – | – | – |
| FlexiCTS [8] | HBT allocation | CPPR-aware 3D CTS | – | – | – |
| AnchorGrid [7] | Pad array | Root-balanced | – | – | 224 (CPU2) |
| Open3DBench [12] | Vias / virtual HBs | MoL: CTS; LoL: ideal | – | Yes | Residual (LoL) |
| DPO-3D [9] | F2F power network | – | Built | – | Residual |
| CU-3D | 6/3-via HB stacks | Receiver then owner | Built | 0–133 | 0 (11 layouts) |
Table 1. What representative 3D methods physically build and check.
Overview and Objective
Flow at a Glance
Synthesis supplies the netlist; a real-CTS 2D reference supplies timing and area growth. Partitioning assigns tiers, and floorplanning establishes macros, PG, and access reservations. Placement exposes pins for joint allocation and residual PG matching. Stacks, receiver/owner clocks, powered insertion, and per-tier routing realize the selected interfaces (Figure 2).
The shared rules become progressively more specific: partitioning budgets space, allocation tests complete footprints, and insertion checks supply access. Provisional assignments are refined using feedback before physical implementation. After construction, optimization may change feeders but must preserve the assigned HB sites and stack.
Definitions and Objective
The F2F model specifies layers, windows, grid, footprints, and PG/access rules. Roles are DATA (including control/I/O), CLOCK, and PG by supply. A role-compatible site fits the required footprint and access. Physical completion checks whether committed resources coexist (§ 9.4).
We seek tier assignments, placements, HB assignments, clocks, PG, and routes that satisfy physical completion and maximize extracted setup WNS. For a register path ,
Here is setup slack, the period, launch/capture clock latency, setup time, and uncertainty. Data delay comprises cell delay (including clock-to-output) and net delay . WNS is minimum slack; TNS sums negative setup slacks over reported endpoints. Interconnect changes affect both components through load, slew, buffering, and sizing; clock phase also affects slack.
Executable F2F Resources
F2F Technology Model
CU-3D uses an ASAP7-based technology model [25, 26] to identify sites that fit complete conductor stacks and satisfy spacing, enclosure, and access rules. Two logic dies share one XY database with distinct layer identities and explicit inter-die timing nets. Twelve ordinary and five interface metals form the stack (Figure 3). TOP cell views map local M1–M6 to M12–M7 while retaining pin XY geometry. Layer order, routing direction, PG direction, and instance orientation remain distinct.
Figure 3. F2F model: 17 metals and complete HB footprints. DATA/CLOCK uses six vias; PG uses three. Each stack crosses one bond at 0.5-m nominal pitch.
DATA/CLOCK traverses six vias from M6 to M7; PG uses the central three between B_TM2 and T_TM2. Both TM2 layers carry horizontal interface PG rails. Ordinary BOTTOM routing uses M2–M6 (three horizontal, two vertical layers); TOP uses M7–M11 (three vertical, two horizontal), with cell-pin access on M1 and M12, respectively. Vertical routing on BOTTOM and horizontal routing on TOP therefore have fewer ordinary layers, motivating direction-specific congestion analysis.
Rail width and via enclosure determine PG clearance, so site extraction uses constructed metal rather than planned stripes. HB connections preserve cross-die net connectivity, while interconnect evaluation accounts for the HB conductor stack and feeder wires on both dies.
Role-Compatible Sites
Within the common die outline, grid contains points , where is its XY origin, are pitches, and are integer indices. At site , role requires : its complete footprint must avoid the applicable obstacle set , including the boundaries, both dies’ macro-access exclusions, and PG clearances. PG stripes can remove DATA/CLOCK eligibility while supplying power landings. For supply net ,
Here is the eligible PG-site set; denote BOTTOM/TOP. The operator erodes same-net landing regions by enclosure footprints , leaving legal centers. Opposite-net clearance and complete-stack containment further restrict this intersection. Both enclosures must fit; a nearby rail alone is not a landing.
Capacity and Assignability
Let be DATA/CLOCK-compatible sites and . Unique physical capacity and guarded allocation capacity are
Here counts distinct sites, is the withheld fraction, and integer releases part of the reserved capacity, with . Thus the guard is deducted once after obstacle filtering. DATA/CLOCK demand must also fit . Role allowances limit counts, not exclusive zones. For a request subset , let be the union of its eligible sites. If , no matching exists (see Figure 4c): adequate total capacity does not imply assignability.
Figure 4. Role-compatible resources. (a) Full footprints account for BOTTOM (B) and TOP (T) obstacles. (b) PG needs same-net landings on both dies. (c) Unique-site capacity differs from local assignability. (d) Sites – serve DATA requests , CLOCK groups , and PG request : relocating admits .
Freedom-Matched Placement
Capacity Follows Mobility
Capacity constraints follow the spatial freedom available to downstream placement. For total ordinary-cell area in region and usable tier areas , the area assigned to BOTTOM, , must satisfy
Local bounds suit inherited neighborhoods. Tier-aggregated capacity lets relocatable cells cross old bins without cutting cohesive logic for local balance. Fixed macros, rows, and access exclusions still limit space. Area bounds do not replace legalization or access checks; equal area is only a placement preference. The placer must permit the migration assumed by its partition, or aggregate spare area can hide local crowding.
The partitioner ranks HB overflow, area overflow, non-CLOCK crossing demand, local balance, and weighted cost lexicographically. The final cost is
For crossing net , is the base cut weight, is the timing penalty, a congestion discount, and the per-HB weight. Driver, direction, and class determine bond count . Count priorities precede : a favorable cut cannot compensate for exceeded capacity.
Scaled: Inherit and Refine
Scaled contracts 2D coordinate seeds while retaining real cell and macro dimensions. Per-tier legalization and refinement using the complete netlist preserve useful neighborhoods. The risk is a causal chain: rigid local balance splits connected clusters, increases DATA crossings, consumes CLOCK-reachable sites, and concentrates feeders on the remaining routing tracks. Local physical capacity is necessary; a forced 50/50 split in every bin is not.
CU-3D interrupts this chain through post-CTS insertion headroom, soft balance within hard capacity, congestion weights, role allowances, and a single capacity guard. DATA alternatives and targeted expansion resolve local competition; PG reservations and tier windows preserve access. Fewer cuts help only where released sites match CLOCK reach and PG access.
Fixed-size macros impose a distinct limit on coordinate inheritance: compressed seeds can enter a macro, forcing legalization to reorganize its neighborhood. Old HB targets then become stale. Re-anchoring to current pins restores consistency; releasing inherited coordinates opens a different search space.
Fresh: Expose, Then Commit
Interface-guided Fresh releases inherited ordinary-cell coordinates within assigned tiers, allowing relocatable-capacity partitioning to retain cohesive logic. TOP goes first because its register pins and capacitances define receiving CLOCK demand; BOTTOM macros and I/O already supply fixed boundaries. Figure 5b follows six steps grouped by the information they expose.
Figure 5. Freedom-matched placement. (a) Scaled retains spatial organization while controlling unnecessary crossings. (b) Fresh reveals receiving-die demand, provisionally assigns interfaces, and lets BOTTOM respond before connected refinement. (c) Macros couple placement, interface-access, and supply-service domains.
Steps 1–2 place TOP with inactive-tier attraction isolated and PG access reserved, restore complete connectivity, and export current TOP pins/capacitances plus fixed BOTTOM boundaries. Steps 3–4 jointly plan provisional DATA/CLOCK sites with retained DATA alternatives and compatible PG capacity, then place BOTTOM against these anchors while preserving TOP.
Steps 5–6 refine the connected design BOTTOM then TOP within ownership/protection boundaries. Both feeders are reevaluated; local PG, sites, pins, and loads are refreshed before final allocation. Thus the early plan guides BOTTOM, whereas final choices use both placed dies and actual landings. Costs combine pin spans, criticality, and migration. Because moving an HB can shorten one feeder and lengthen the other, connected timing judges the combined response.
Macro Feasibility Domains
Macros retain dimensions, orientation, pins, obstructions, supply rails, and tier identity. Active BLOCK and inactive COVER views (LEF macro classes) separate placement ownership while preserving geometry. Same-die bodies and rows bound cell placement; two-die footprints bound interface access; both supplies plus signal escape bound new-cell service. Legal cross-die projection overlap can therefore block an HB at the same XY point.
Cell capacity subtracts same-tier exclusion unions; HB capacity filters both dies. Rigid per-tier macro translation can reduce projected exclusions while preserving macro organization. Connected logic and HB targets must follow the moved pins. Shorter matched-pin distances nevertheless accompanied pre-CTS WNS losses of 12.4/27.9 ps on BOTTOM/TOP: proximity omits fanout, load, and collateral nets. Refinement must preserve electrical context.
Joint Interface Allocation
Shared Occupancy
DATA and CLOCK share site occupancy while preserving compatible capacity for residual PG. A CLOCK alternative identifies clock domain , source , HB site , and sink set . Current pins/caps screen fanout, diameter, reach, latency/skew, and estimated load
Here are sink ’s capacitance and XY location, is wire capacitance per unit length, and denotes Manhattan distance. Buffered-tree candidates additionally account for buffer input loads and branch lengths. DATA request retains its legal baseline and permitted relocation sites . Binary selects CLOCK alternative ; binary assigns DATA request to . Writing for ’s site and taking outside , we require
Equation 6–8 cover each clock sink once, assign each DATA request once, and allow one occupant per physical site . Exact mode applies clock-count and regional limits, then prioritizes count deviation, CLOCK timing, critical DATA, ordinary DATA, and control/I/O costs.
Retaining DATA Alternatives
Proposition. Feasible assignment sets (fixed DATA) and (movable DATA), under identical CLOCK choices and constraints, satisfy if every legal DATA baseline is retained. Setting , where is ’s retained site, recovers each fixed solution. Solver mode determines how well the enlarged set is searched.
The converse fails in Figure 4d: moving to frees ’s sole site , lets use , keeps at , and retains for PG. No request
Algorithm 1 Joint allocation with residual PG
Input: Compatible sites, current pins/caps, demand, limits
Build sparse CLOCK/DATA choices; retain legal baselines
Select size-adaptive mode; solve coverage and occupancy
In spatial mode, expand deficient neighborhoods and resolve
If incomplete, return failure category and witness
Enforce guard/count limits; match compatible residual PG
return Accepted plan or failed condition
is dropped or site added. Retaining legal DATA alternatives allows assignment to accommodate CLOCK groups with fewer options.
Adaptive Solving and Residual PG
Algorithm 1 stores alternatives and sink incidences sparsely; model size selects exact, bounded, or sparse/grouped solving. Spatial matching expands Hall-deficient neighborhoods. Missing candidates, deficient matching, incomplete coverage, infeasibility, and timeout are distinguished: a local lack of alternatives needs candidate expansion, not more solver time.
DATA/CLOCK planning reserves a PG allowance without fixing its final landings. Residual PG matches same-supply requests to unoccupied compatible sites within regional/distance limits. Let , , count assigned DATA/CLOCK/PG interfaces and be the required PG count. Acceptance enforces and plus actual matching. Counts alone are insufficient: only reachable same-supply sites can realize the PG allowance.
Legality-Preserving Realization
Interface Invariants
Physical realization preserves each accepted site’s identity, orientation, pins, and conductor stack. DATA/CLOCK interfaces become SRC/Y, a dedicated interdie net , and DST/A; PG stacks join same-supply landings. Routing refines authorized feeders around this explicit connection (Figure 6a).
Figure 6. Decision preservation: (a) interface identity; (b) measured receiver feedback and owner CTS; (c) dual-supply insertion; (d) routing and checks on one assembled state.
Receiver-First CTS and Skew
TOP trees start at DST/Y. Group ’s measured receiver latency , skew , load, and geometry feed a bounded replan before BOTTOM owner CTS. The owner targets and for local sink . Here is the HB root, the source-to-root latency, and the common end-to-end latency target. Slower receivers leave smaller owner-side budgets; faster receivers allow longer owner branches. Balancing only HB-root arrivals ignores these unequal downstream delays. Receiver-first construction exposes them while the owner can still adapt.
Let and be launch and capture latency edits. With other terms fixed, the setup/hold slack changes are
In register chain , delaying helps incoming setup but hurts outgoing setup; a shared branch affects every descendant. Physical sizing also changes upstream load and slew. Skew optimization must therefore evaluate all affected incoming and outgoing paths, including hold, in the complete clock graph rather than optimize each receiver independently.
Supply-Reachable Insertion
For placement row , let be legal cell-origin positions and , the positions with reachable supply connections. The insertion service domain is , with signal escape additionally required. A legal origin may reach VDD but not VSS (Figure 1c). Checking both supplies prevents a timing repair from creating an unpowered cell. Access reservations protect buffers before insertion; post-insertion checks verify both connections (Figure 6c).
Routing and Completion
Routing obeys tier windows around accepted stacks and PG. Physical completion requires, on one assembled state, zero backend DRC and declared signal/PG opens, intact HB stacks, covered clock sinks, and independent database readback. Timing coverage and electrical qualification accompany these checks; GDS release additionally binds stream maps, cell GDS/CDL, and matched LVS. DRC tests whether the chain’s decisions coexist in the final geometry.
Experimental Evaluation
Setup
Synopsys Design Compiler and Cadence Innovus implement the 2D/3D designs on the ASAP7-based F2F model at TT, 0.70 V, 25 °C. Nominal lengths and areas divide the 4× model geometry by 4 and 16; nominal HB pitch is 0.5 . Each design uses one synthesized netlist and a real-CTS 2D reference. Both 2D and 3D results follow multiple rounds of placement optimization, CCOpt, post-CTS setup optimization, timing/SI-driven routing, and post-route setup
optimization, with high-effort congestion and path-group settings, then extraction.
Scaled reuses the 2D placement, while Fresh rebuilds the tier placements. We evaluate interface-guided ordering on JPEG. Backend DRC and connectivity checks are performed in Innovus on the assembled two-die layouts. Footprint is reported per die, power is reported by the tool using recorded activity, and wirelength is the total routed length. Table 2 reports nominal wirelengths in mm, while Table 3 uses model lengths.
| Design | Route | WNS (ns) | TNS (ns) | (mW) | () | WL (mm) | Cell (%) | Net (%) |
| Rocket | 2D | -0.428 | -321.400 | 42.413 | 5,971.3 | 204.56 | 64.0 | 36.0 |
| Scaled | -0.410 | -455.573 | 37.112 | 3,843.0 | 203.79 | 64.1 | 35.9 | |
| AES | 2D | -0.003 | -0.005 | 6.911 | 1,672.5 | 47.31 | 86.3 | 13.7 |
| Scaled | 0.000 | 0.000 | 6.573 | 1,060.3 | 49.16 | 87.5 | 12.5 | |
| Mini | 2D | -0.290 | -38.109 | 12.154 | 2,218.9 | 62.35 | 75.4 | 24.6 |
| Scaled | -0.268 | -51.340 | 10.294 | 1,404.5 | 55.50 | 76.9 | 23.1 | |
| s35932 | 2D | 0.023 | 0.000 | 6.783 | 1,315.5 | 17.07 | 78.2 | 21.8 |
| Fresh | 0.051 | 0.000 | 5.851 | 797.6 | 19.69 | 61.4 | 38.6 | |
| Ariane133 | 2D | -2.877 | -10,388.600 | 233.445 | 71,074.4 | 1,632.51 | 38.1 | 61.9 |
| Fresh | -2.876 | -18,240.000 | 215.131 | 42,640.6 | 1,489.83 | 37.1 | 62.9 |
Table 2. 2D/3D PPA and top-20 path composition. : tool-reported power; : per-die footprint; WL: wirelength. Each row uses one implementation. Cell/net shares pool its 20 lowest-slack data paths; cell delay includes clock-to-output.
| Variant | WNS | TNS | WL | |
| Bottom-first | -1.580 | -541.766 | 959.2 | 52.90 |
| Top-first, DATA only | -0.551 | -212.371 | 723.7 | 45.66 |
| Early joint planning | -0.464 | -167.780 | 727.6 | 45.63 |
| + Movable-capacity partition | -0.118 | -54.046 | 573.9 | 41.81 |
| Timing-weighted | -0.175 | -139.920 | 606.1 | 41.73 |
| Fixed-sink refinement | -0.132 | -140.240 | 578.0 | 41.93 |
| Independent initializations | -0.219 | -203.370 | 628.8 | 43.83 |
Table 3. JPEG variants. WNS/TNS: ns; WL: model m; raw : mW. All report DRC/connectivity 0/0.
Completed Layouts and PPA
Across eleven designs, the post-route layouts exhibit zero backend DRC violations and pass connectivity checks (Figure 7). Table 1 compares implementation scope. Table 2 reports PPA for nine paired CU-3D and 2D implementations. Rocket/Mini Scaled gain 18/22 ps, AES Scaled reaches zero, and s35932 Fresh gains 28 ps. Rocket/Mini trade better WNS for worse TNS. All nine paired designs admit smaller-footprint, lower-power implementations; four improve best-route WNS.

Figure 7. Realized layouts across macro scales. BOTTOM/TOP rows share one legend. All four have zero backend DRC and pass signal connectivity; Swerv and Ariane also pass PG connectivity.
WNS measures the worst endpoint; TNS captures the aggregate setup deficit, distinguishing bottleneck gains from gains across violating endpoints. Mini reduces routed length by 11.0%, but AES/s35932 improve WNS with 3.9%/15.4% more wire. Buffering, clock distribution, and rerouting can increase total length while improving critical paths. Table 2 ties all metrics to the same implementation.
Where the Timing Gain Lands
The 20 lowest-slack paths separate cell delay (including clock-to-output) from inter-instance net delay, excluding clock latency and input arrival. Shares use summed delays, so a larger cell share need not imply larger absolute cell delay. Across nine paired 2D designs, cells contribute 51.6–95.7% of worst-path data delay and 51.3–91.5% pooled over the top 20. Mini/AES reduce mean top-20 data delay by 56.2/19.7 ps; Ariane remains net-dominated in 2D/3D (61.9%/62.9%).
Endpoint-matched analysis aligns pins, edges, type, view, and group, selected by 2D criticality. Nine of ten representative comparisons reduce cell delay by 8.6–71.1 ps; seven increase net delay. Mini/AES save 71.1/63.8 ps in cells despite net increases of 9.2/13.5 ps. Load, slew, sizing, and changed logic chains matter beyond total wirelength. Shorter feeders can reduce driver load and improve downstream slew, benefiting cell delay as well as net delay. Additional buffering likewise changes both components.
Independent cohorts identify current bottlenecks; matched endpoints track the same connection. Rocket’s matched data delay falls 28.9 ps although its top-20 mean grows as input–register paths replace input–output paths. The former s35932 worst path drops to rank 82, illustrating critical-path migration.
Order and Mobility on JPEG
Bottom-first, top-first DATA-only, and early DATA/CLOCK planning reach −1.580, −0.551, and −0.464-ns WNS (Table 3). Early planning gains 87 ps despite 0.53% more wirelength: exposing competition improves more than geometry.
Mobility-aware low-HB partitioning reaches −0.118 ns: WNS improves 346 ps, TNS 113.734 ns, wirelength 21.12%, and raw power 8.38%. DATA (excluding control/I/O) falls 6211 to 561; CLOCK rises 293 to 471; PG is unchanged. Two 50-model-m samplings reduce peak occupancy from 313/325 to 129/129. Fewer cuts release choices without dictating clock demand.
Fixed-sink refinement and independent initialization lose 14 and 101 ps, respectively. Exposing demand and allowing placement to respond are complementary: an early HB plan needs downstream mobility. Subsequent refinement reaches −0.101 ns; 18/20 worst low-HB paths stay within one die. Once avoidable cuts are removed, cell/load and clock response become the next targets.
Congestion and Macro Scaling
Table 4 compares tier-isolated routing replay. Fresh issues 5.9–14.5× fewer non-CLOCK requests and lowers TOP horizontal overflow in all six designs. Horizontal dominates at all twelve TOP records; vertical is at least horizontal in eleven BOTTOM records, consistent with unequal directional layer counts and PG occupancy. The sum of H/V overflows falls 53 to 26 for JPEG and 178 to 3 for RISC-V32i, but Rocket BOTTOM vertical overflow rises 50 to 83: released capacity can redistribute pressure between dies. The relevant resource is therefore not just a bond count, but reachable sites together with feeder capacity in each tier’s scarce routing direction.
| Design | Requests S/F | BOTTOM S F | TOP S F |
| Rocket | 5,532/468 | 7/50 4/83 | 150/28 62/11 |
| JPEG | 6,230/578 | 0/1 0/6 | 47/5 20/0 |
| RISC-V32i | 2,257/354 | 95/62 0/0 | 19/2 3/0 |
| AES | 1,593/269 | 1/10 0/6 | 58/2 30/0 |
| Dynamic | 1,998/310 | 0/3 0/9 | 28/2 19/4 |
| Mini | 1,926/133 | 0/2 0/5 | 30/1 28/0 |
Table 4. Scaled (S) to Fresh (F): non-CLOCK interface requests and tier-isolated horizontal/vertical (H/V) routing overflow. Same-stage global-route replay; entries give H/V overflow.
Figure 7 spans 0–133 macros. Swerv has 3,544 DATA/CLOCK bundles, 6,302 PG bonds, and 56 macro PG terminals; Ariane integrates 4,379 bundles and 7,218 PG bonds. Both pass signal/PG connectivity with zero backend DRC. Swerv’s tall macro blocks leave narrow channels, whereas Ariane’s repeated macro banks distribute access boundaries across the footprint. These layouts test the same interface rules under distinct constraints on cell placement, bond access, and power delivery.
Realized Structures and Scale
Figure 8 contrasts direct five-DFF and buffered seven-DFF receivers. In panel (a), 27/96 denotes straight wire sections visible in the window out of 96 in the complete BOTTOM owner net. Panel (b) uses two parallel buffers, one per sink path: allocation selects the root, whereas receiver CTS realizes the fanout. Both receiver structures use the same interface and legal-site model. Of 144 sites, 56 support DATA/CLOCK, 48 PG, and 40 are blocked; 33 are occupied. Capability differs from occupancy. Sparse allocation spans 1,661 variables (gcd) to 3,043,683 (JPEG Scaled), versus 528,181 for JPEG Fresh. These are assignment variables, not physical bonds: candidate sites and clock groupings represent alternative uses of the same resources. Retained alternatives let assignment resolve competition without changing the interface lattice.
Figure 8. Realized JPEG structures. (a,c,d) Actual clock routes, PG, and HB sites in one 6-by-6- window; XY is measured and Z illustrates layer order. (b) A second receiver’s connection topology: two parallel buffers drive seven DFFs, one buffer per path. Open/filled site markers denote eligibility/occupancy.
Conclusion
CU-3D makes legality a property of the decision chain: resource realism ties capacity to geometry, freedom-matched commitment exposes demand before fixing choices, and decision preservation carries interfaces through powered clock construction and routing. Eleven designs, including 28- and 133-macro processors, demonstrate zero backend DRC and passing signal connectivity. JPEG links retained alternatives and placement mobility to timing gains; matched paths show cell savings outweighing net-delay increases.
Cell-dominated logic calls for load- and slew-aware reorganization, whereas Ariane’s net-dominated paths motivate coordinated costs for both feeders. Retained sites make alternatives physically available; completed timing determines which alternatives are useful. Thus neither fewer bonds nor shorter total wirelength is a universal 3D objective. Future work will feed matched-path cell, feeder, and clock-phase costs into partitioning, HB selection, and macroneighborhood optimization, and extend the common physical basis to multicorner timing, power integrity, and thermal constraints.
Acknowledgments
This work is partially supported by The Research Grants Council of Hong Kong SAR (No. CUHK14211824 and T46-415/25-R).
References
- [1]M.-F. Chen, F.-C. Chen, W.-C. Chiou, and D. C. H. Yu, “System on integrated chips (SoIC) for 3D heterogeneous integration,” in Proceedings of ECTC. Piscataway, NJ, USA: IEEE, 2019, pp. 594–599.DOI
- [2]S. A. Chew, B. Zhang, K. Vanstreels, E. Chery, J. De Messemaeker, L. Witters, K. Van Sever, S. Iacovo, S. Dewilde, M. Stucchi, J. De Vos, G. Beyer, A. Miller, and E. Beyne, “The challenges and solutions of Cu/SiCN wafer-to-wafer hybrid bonding scaling down to 400nm pitch,” in Proceedings of IEDM. Piscataway, NJ, USA: IEEE, 2023, pp. 1–4.
- [3]S. Panth, K. Samadi, Y. Du, and S. K. Lim, “Shrunk-2-D: A physical design methodology to build commercial-quality monolithic 3-D ICs,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 36, no. 10, pp. 1716–1724, 2017.DOI
- [4]B. W. Ku, K. Chang, and S. K. Lim, “Compact-2D: A physical design methodology to build commercial-quality face-to-face-bonded 3D ICs,” in Proceedings of ISPD. New York, NY, USA: Association for Computing Machinery, 2018, pp. 90–97.DOI
- [5]P. Vanna-Iampikul, C. Shao, Y.-C. Lu, S. Pentapati, and S. K. Lim, “Snap-3D: A constrained placement-driven physical design methodology for face-to-face-bonded 3D ICs,” in Proceedings of ISPD. New York, NY, USA: Association for Computing Machinery, 2021, pp. 39–46.
- [6]S. Liu, J. Jiang, Z. He, Z. Wang, Y. Lin, B. Yu, and M. Wong, “Routing-aware legal hybrid bonding terminal assignment for 3D face-to-face stacked ICs,” in Proceedings of ISPD. New York, NY, USA: Association for Computing Machinery, 2024, pp. 75–82.
- [7]P. Vanna-iampikul, J. Yoon, C. Park, G. Yeap, and S. K. Lim, “Placement-aware 3D net-to-pad assignment for array-style hybrid bonding 3D ICs,” in Proceedings of ISPD. New York, NY, USA: Association for Computing Machinery, 2025, pp. 200–208.
- [8]S. Li, L. Jin, S. Liang, Z. Zhuang, R. Chen, B. Yu, and T.-Y. Ho, “FlexiCTS: CPPR-aware 3D clock tree synthesis for face-to-face bonded ICs,” in Proceedings of DAC. New York, NY, USA: Association for Computing Machinery, 2026.DOI
- [9]Z. Zhuang, Z. Yang, Y. Zhao, J. Hu, B. Yu, S. K. Lim, and T.-Y. Ho, “DPO-3D: Differentiable power delivery network optimization via flexible modeling for routability and IR-drop tradeoff in face-to-face 3D ICs,” in Proceedings of ASP-DAC. Piscataway, NJ, USA: IEEE, 2026, pp. 1273–1279.DOI
- [10]L. Bamberger, A. García-Ortiz, L. Zhu, S. Pentapati, D. E. Shim, and S. K. Lim, “Macro-3D: A physical design methodology for face-to-face-stacked heterogeneous 3D ICs,” in Proceedings of DATE. Piscataway, NJ, USA: IEEE, 2020, pp. 37–42.
- [11]L. Zhu, J. Hu, G. Murali, and S. K. Lim, “Hetero-3D: Maximizing performance and power delivery benefits of heterogeneous 3D ICs,” in Proceedings of ISLPED. New York, NY, USA: Association for Computing Machinery, 2024, pp. 1–6.DOI
- [12]Y. Shi, C. Gao, W. Ren, P. Xie, S. Xu, K. Xu, M. Yuan, C. Qian, and Z.-H. Zhou, “Open3DBench: Open-source benchmark for 3D-IC backend implementation and PPA evaluation,” 2026, version 2, April 5, 2026. [Online]. Available: https://arxiv.org/abs/2503.12946v2
- [13]K. Chang, S. Sinha, B. Cline, R. Southerland, M. Doherty, G. Yeric, and S. K. Lim, “Cascade2D: A design-aware partitioning approach to monolithic 3D IC with 2D commercial tools,” in Proceedings of ICCAD. New York, NY, USA: Association for Computing Machinery, 2016, pp. 1–8.DOI
- [14]S. S. K. Pentapati, K. Chang, V. Gerousis, R. Sengupta, and S. K. Lim, “Pin-3D: A physical synthesis and post-layout optimization flow for heterogeneous monolithic 3D ICs,” in Proceedings of ICCAD. New York, NY, USA: Association for Computing Machinery, 2020.
- [15]G. Kim and H. Park, “A unified design flow for homogeneous and heterogeneous 3D integration with fine-pitch hybrid bonding,” in Proceedings of ICCAD. Piscataway, NJ, USA: IEEE, 2025, pp. 1–9.
- [16]L. Jiang, X. Shi, R. Xiao, C. Yan, R. Jiang, Z. Wang, and K. Zhu, “Differentiable partitioning with dual-max 3D wirelength model and hybrid bonding terminal awareness for optimized 3D placement,” arXiv:2609.06408v2, 2026. [Online]. Available: https://arxiv.org/abs/2609.06408v2
- [17]L. Zhu, C. Jo, and S. K. Lim, “Power delivery solutions and PPA impacts in micro-bump and hybrid-bonding 3D ICs,” IEEE Transactions on Components, Packaging and Manufacturing Technology, vol. 12, no. 12, pp. 1969–1982, 2022.DOI
- [18]M. Wang, L. Zou, G. Li, F. Gu, Y. Lu, T. Yang, B. Han, J. Mu, H. Li, S. K. Lim, and B. Yu, “Distill3D: Cross-stage knowledge distillation for early-stage 3D IC timing prediction,” in Proceedings of ICCAD, 2026, to appear. [Online]. Available: https://lanchengzhou.github.io/
- [19]Y. Ji, Y. Shi, T. Hou, Y. Zhao, C. Zhao, P. Liao, Z. Guo, C. Qian, Y. Lin, and B. Yu, “Integrated timing-driven placement for hybrid-bonding-based face-to-face 3D ICs,” in Proceedings of DAC. New York, NY, USA: Association for Computing Machinery, 2026.
- [20]Y. Ji, Y. Lu, Z. Zhang, Z. Guo, Y. Lin, and B. Yu, “DiffCCD: Differentiable concurrent clock and data optimization,” in Proceedings of ICCAD. Piscataway, NJ, USA: IEEE, 2025, pp. 1–9.
- [21]Y. Zhu, D. Feng, Z. Luan, L. Ren, W. Chen, and Z. Tan, “Open3DFlow: An open-source EDA platform for 3D chip design with AI enhancement,” in Proceedings of ICCAD. Piscataway, NJ, USA: IEEE, 2025, pp. 1–9.DOI
- [22]Open3DFlow Authors, “Open3DFlow: Open-source project and demonstrations,” n.d., project artifact; accessed October 28, 2026. [Online]. Available: https://github.com/bzhang24/Open3DFlow
- [23]L. Jiang, A. B. Kahng, Z. Wang, and Z. Zheng, “Invited: Toward sustainable and transparent benchmarking for academic physical design research,” 2026. [Online]. Available: https://arxiv.org/abs/2601.17520
- [24]S. Patnaik, M. Ashraf, O. Sinanoglu, and J. Knechtel, “Best of both worlds: Integration of split manufacturing and camouflaging into a security-driven CAD flow for 3D ICs,” in Proceedings of ICCAD. New York, NY, USA: Association for Computing Machinery, 2018, pp. 1–8.DOI
- [25]L. T. Clark, V. Vashishta, L. Shifren, A. Gujja, S. Sinha, B. Cline, C. Ramamurthy, and G. Yeric, “ASAP: A 7-nm FinFET predictive process design kit,” Microelectronics Journal, vol. 53, pp. 105–115, 2016.DOI
- [26]V. Vashishta, M. Vangala, and L. T. Clark, “ASAP7 predictive design kit development and cell design technology co-optimization: Invited paper,” in Proceedings of ICCAD. Piscataway, NJ, USA: IEEE, 2017, pp. 992–998.





