Floorplanning in VLSI: Utilization, Aspect Ratio, Pins and Checks

Floorplanning in VLSI explained: die and core sizing, utilization and aspect ratio with worked examples, rows and sites, pin placement, blockages and checks.

Floorplanning in VLSI: Utilization, Aspect Ratio, Pins and Checks
On this page 16 sections
  1. What is floorplanning in VLSI?
  2. Inputs you need
  3. Die, core and the core-to-die margin
    1. Pad-limited vs core-limited chips
  4. Utilization: the most important number
    1. Worked example A: sizing a core (illustrative)
    2. Worked example B: the real utilization after blockages (illustrative)
  5. Aspect ratio and shape
    1. Worked example C: pins decide the shape (illustrative)
  6. Rows, sites and tracks
  7. I/O pin placement
  8. Macros, halos and channels
  9. Blockages, regions and power domains
  10. Physical-only cells and the power grid
  11. What to check before you hand off to placement
  12. Debug flowchart: when placement or routing fails because of the floorplan
  13. Interview questions and answers
  14. FAQ
  15. Summary
  16. References

Floorplanning is the first physical step of a place-and-route flow, and its mistakes are the most expensive to fix later. A congested channel, a pin on the wrong edge or a core that is too small can cost weeks in placement, CTS and routing. Interviewers ask about it because a good answer shows you understand how area, timing, congestion and power fit together.

This guide covers what a floorplan contains, how to size the core with utilization and aspect ratio, rows and sites, I/O pin placement, how macros, blockages and physical-only cells fit in, the checks to run before placement, a debug flowchart and interview Q&As. The worked examples are illustrative: the numbers show the method and are not from any foundry, library or project, except where an open-source library is named. Tool commands are simplified sketches; check your tool's documentation for exact options. Related reading on iVLSI: Physical Design Flow, Macro Placement in VLSI, Power Planning in VLSI and Routing Congestion in VLSI.

What is floorplanning in VLSI?#

Floorplanning turns a gate-level netlist into a physical starting point. Synthesis gives you cells and connections but no coordinates. The floorplan decides:

  • Die and core size and shape. The die is the full block (or chip) outline. The core is the inner area where standard cells and macros are placed.
  • Rows and sites. The placement grid that standard cells snap to.
  • Routing tracks. The preferred-direction grid for each metal layer.
  • I/O pins (for a block) or pads/bumps (for a full chip) and their layers.
  • Macro placement, with halos and channels between macros.
  • Placement and routing blockages, regions and fences for special logic.
  • Physical-only cells that must exist before placement, mainly endcaps and well taps.
  • Space for the power network, which is built right after.

In short, the floorplan fixes everything the placer is not allowed to move.

Figure 1. Anatomy of a block floorplan: die, core margin, rows, macros with halos, a placement blockage and I/O pins (not to scale, illustrative).

Inputs you need#

Floorplanning reads the usual PnR inputs:

  • Gate-level netlist from synthesis.
  • Technology LEF (layers, pitches, vias, sites) and cell/macro LEF (sizes, pins, blockages). See Physical Library (LEF).
  • Liberty libraries and SDC constraints, for early timing.
  • UPF power intent for multi-voltage designs; its power domains become physical areas.
  • Block outline and pin constraints from the parent floorplan in a hierarchical flow, often as a DEF.

Run the pre-placement sanity checks first; a floorplan built on a netlist with unresolved references is wasted work.

Die, core and the core-to-die margin#

The core holds the rows. The ring between the core edge and the die edge (the core margin or core-to-IO spacing) is used for the power ring, pin access, and in a full chip the I/O pad ring. Two ways to set the size are common, and the OpenROAD initialize_floorplan command supports both:

  1. Explicit coordinates. You give the die and core rectangles (or rectilinear polygons). This is normal for a block whose outline comes from the top-level floorplan.
  2. From utilization and aspect ratio. The tool computes the core size from the cell area, a target utilization and an aspect ratio, then adds the margin.

The OpenROAD documentation gives the second method as:

Formula
core_area = design_area / (utilization / 100) core_width = sqrt(core_area / aspect_ratio) core_height = core_width * aspect_ratio die = core + core_space on each side

Its documentation also warns that you should use one method or the other, not both, and the OpenROAD-flow-scripts floorplan script treats the methods (DEF, footprint, die/core area, or utilization) as mutually exclusive.

Pad-limited vs core-limited chips#

At chip level, a design is core-limited when the logic sets the die size, and pad-limited when the number of I/O pads around the perimeter forces a die larger than the logic needs. Flip-chip area bump arrays relax this limit because I/O is not restricted to the periphery.

Utilization: the most important number#

Core utilization is the fraction of placeable area that cells occupy. Definitions vary between tools and teams, so always say which one you mean:

  • Total utilization = (standard cell area + macro area) / core area.
  • Standard-cell utilization = standard cell area / (core area − macro area − halos − hard blockages).

The second one is what matters for placement and routability, because macros do not need routing space between their own cells. The old iVLSI Floorplan article shows both forms of the core-size equation.

There is no universal target; it depends on the library, routing layers, cell pin density and expected growth. Keep in mind:

  • The netlist grows. Optimization adds buffers, CTS adds clock buffers, hold fixing adds delay cells, and taps, endcaps and decaps consume sites. Start with headroom.
  • Utilization is not congestion. Two floorplans with the same utilization can route very differently.
  • Placement density is a separate knob. OpenROAD-flow-scripts, for example, has CORE_UTILIZATION for floorplanning and a separate PLACE_DENSITY for global placement.

As a concrete open-source data point, the OpenROAD-flow-scripts example configurations for the Nangate45 platform use CORE_UTILIZATION ?= 55 for the small gcd design and CORE_UTILIZATION = 65 with CORE_ASPECT_RATIO = 1 and CORE_MARGIN = 5 for swerv. These are settings for those flows, not general targets.

Worked example A: sizing a core (illustrative)#

A block has 120,000 µm² of standard-cell area and two hard macros of 400 × 300 µm. Each macro gets a 10 µm halo on every side, so each reserves 420 × 320 = 134,400 µm², or 268,800 µm² for both. The target standard-cell utilization is 65% and the aspect ratio is 1.0.

  • Area for standard cells = 120,000 / 0.65 ≈ 184,615 µm²
  • Core area = 184,615 + 268,800 ≈ 453,415 µm²
  • Core side = √453,415 ≈ 673.4 µm

Now snap to the placement grid. Using the Nangate45 site in OpenROAD-flow-scripts, whose LEF gives SIZE 0.19 BY 1.4, the height becomes 481 rows × 1.4 = 673.4 µm and the width 3,545 sites × 0.19 = 673.55 µm. Add a 10 µm core margin on every side and the die is about 693.55 × 693.4 µm.

Compare this with the total-utilization formula: (120,000 + 268,800) / 0.65 ≈ 598,154 µm². Including macro area in the denominator makes the core about 32% bigger than needed. That is why you should be clear about which utilization a script or a manager means.

Worked example B: the real utilization after blockages (illustrative)#

With the same core (about 453,569 µm²), suppose you add 10,000 µm² of hard placement blockage for a congested corner and an analog keep-out region. The area left for standard cells is 453,569 − 268,800 − 10,000 = 174,769 µm². The starting standard-cell utilization is now 120,000 / 174,769 ≈ 68.7%, not 65%. If optimization and CTS then add 10% more cell area (132,000 µm²), utilization reaches about 75.5%. Recalculate after every blockage or macro change.

Aspect ratio and shape#

Aspect ratio describes the core shape. In OpenROAD it is height / width, with 1.0 meaning a square; other tools may define it the other way, so check. A square minimizes perimeter for a given area, which tends to shorten wires. Choose another shape when the parent floorplan gives a fixed slot, when an edge needs more length for pins, or when macros fit a rectangle better. Rectilinear (L or T shaped) blocks are possible too (OpenROAD switches to rectilinear mode when the die is given as a polygon), but they make power, pins and congestion harder.

Figure 2. Same core area, different aspect ratios: a square has the shortest perimeter, wide or tall shapes trade it for longer edges for pins.

Worked example C: pins decide the shape (illustrative)#

The core of example A is 673 µm on each side. Its west edge must carry two 512-bit data buses plus 256 control and status signals, 1,280 pins in total. Assume a usable pin slot every 0.56 µm on one pin layer. The pins need 1,280 × 0.56 ≈ 717 µm, more than the 673 µm edge, before subtracting any corner keep-out. Options:

  • Use a second pin layer on that edge (check pin access and spacing rules).
  • Move some buses to the north or south edge, if the parent floorplan agrees.
  • Make the core taller. At height / width = 1.5 the same area gives about 550 × 825 µm, and the 825 µm west edge fits the 717 µm of pins with room for corner keep-outs.

Count pins per edge early; they often decide the shape more than area does.

Rows, sites and tracks#

A site is the smallest placement unit, usually one poly pitch wide and one standard-cell row high. A row is a line of sites. Standard cells have widths that are whole multiples of the site width and heights equal to the row height (or a multiple, for double-height cells).

Adjacent rows are usually flipped (orientation N, then FS, then N...) so that neighbouring rows share a VDD or VSS rail. Row parity can therefore matter; OpenROAD has -row_parity to force an odd or even row count, and -additional_sites to create rows for sites not yet used, such as double-height cells that flop clustering may add later.

Tracks are the routing grid for each layer. In OpenROAD, make_tracks builds them from the technology LEF pitch and offset. Pins and power stripes should line up with this grid; off-track pins cause pin-access problems later.

Figure 3. Rows, sites and shared power rails: alternate rows are flipped so neighbouring rows share a rail (illustrative).

I/O pin placement#

For a block, every top-level port becomes a pin shape on the boundary. Good pin placement:

  • Follows the parent floorplan: a pin should face the block it connects to, so the top-level route is short.
  • Uses allowed layers: usually a horizontal layer on the north and south edges and a vertical layer on east and west, matching preferred directions. In OpenROAD's place_pins you set these with -hor_layers and -ver_layers.
  • Keeps spacing and corner keep-outs: place_pins has -min_distance (by default two routing tracks between pins) and -corner_avoidance, plus -exclude for edge intervals where pins are not allowed.
  • Groups buses so related bits stay together (-group_pins). The OpenROAD pin placer puts pins on the track grid to minimize wirelength, treating unplaced cells as being at the die centre, so a common flow does a rough pin pass first and refines it after global placement. In a hierarchical flow the top level often fixes the pins. Boundary timing depends on the block SDC I/O delays; see Time Budgeting in VLSI.

Macros, halos and channels#

Memories and hard IP are placed during floorplanning because they are large and shape everything else: near the edges with pins facing the logic, grouped by data flow, with channels wide enough for the wires that must pass and halos around them. See Macro Placement in VLSI for worked channel sizing. Every macro move changes the available area, so re-check utilization after each change.

Blockages, regions and power domains#

  • Hard placement blockage: no standard cells at all. Used in narrow channels, near analog macros or in corners where placement would create unroutable spots.
  • Soft blockage: no cells during initial placement, but optimization buffers may be placed. Useful in macro channels that must still carry repeaters.
  • Partial blockage: limits density to a percentage in an area, to reduce local congestion.
  • Routing blockage: stops routing on chosen layers in an area.
  • Regions, fences and guides: keep a group of cells inside an area (fence: only those cells inside; region: those cells must be inside, others may enter; guide: soft attraction). Exact semantics and names differ between tools. See Blockages in VLSI and Module Constraints.
  • Power domains: each UPF domain becomes a physical area, often with a gap between domains (OpenROAD's -gap defaults to six times the minimum site height). See Power Gating in VLSI and Level Shifters in VLSI.

Physical-only cells and the power grid#

Once rows and macros are fixed, two steps complete the floorplan:

  1. Endcaps and well taps. Endcaps go at row ends and around macros; well taps are placed in a regular pattern within the library's tap distance. OpenROAD's tapcell does both (the Nangate45 platform script uses -distance 120). See Physical-Only Cells in VLSI and Tie Cells in VLSI.
  2. Power planning. Rings, stripes and rails are built on the finished outline. OpenROAD-flow-scripts runs these as floorplan sub-steps: 2_1_floorplan, 2_2_floorplan_macro, 2_3_floorplan_tapcell and 2_4_floorplan_pdn. See Power Planning in VLSI and IR Drop Analysis in VLSI.
Figure 4. Floorplanning steps from netlist to a placement-ready database.

What to check before you hand off to placement#

A floorplan is not finished when it "looks right". Check:

  • Geometry: die and core on grid, rows inside the core, no overlapping macros, macros in legal orientations.
  • Utilization: recalculated after macros, halos and blockages, with headroom for growth.
  • Pins: all ports placed, on allowed layers and tracks, no overlaps, spacing and corner rules met, matching the parent floorplan.
  • Physical-only cells: endcaps on every row end, well-tap distance met everywhere, including next to macros.
  • Power: rails, macro pins and switch cells on the right nets; early static IR drop acceptable.
  • Routability: a trial placement plus early global route shows no large overflow hotspots, especially in macro channels and near pin-dense edges.
  • Timing: trial placement shows no impossible paths, such as macro outputs crossing the whole block in one cycle.
  • Power intent: every domain has its area and always-on routing is possible.

Save the floorplan as a DEF or database snapshot; you will iterate it.

Debug flowchart: when placement or routing fails because of the floorplan#

Work through the usual causes in order:

  1. Legalization fails: recalculate standard-cell utilization with real macro, halo and blockage areas; grow the core or drop needless hard blockages.
  2. Congestion in channels or corners: widen channels, add soft or partial blockages, turn macro pins toward the logic.
  3. Long macro or I/O paths: move pins or macros to follow the data flow.
  4. Boundary pin DRCs: check pin layers, tracks, spacing and corner keep-outs.
  5. IR or EM failures: revisit the ring, stripes and macro power hookups.
  6. Tap or endcap DRCs: re-run insertion after any row, macro or halo change.

After each fix, rerun trial placement and early global route.

Figure 5. Debugging floorplan-related placement, congestion, timing, pin, IR and tap problems.

Interview questions and answers#

Q1. What is floorplanning and why is it important? The first physical step: sizing the die and core, creating rows and tracks, placing pins and macros, adding blockages and physical-only cells and preparing the power grid. Placement, CTS and routing must live with it, so a poor floorplan causes congestion, timing and IR problems later.

Q2. What is core utilization and what value would you use? Cell area divided by available area; state whether it is total or standard-cell only. Justify the value with trial placement and congestion results, not a fixed number.

Q3. How do you calculate the core size? Standard cell area divided by target utilization, plus macro area including halos and any hard blockages, then pick the aspect ratio, snap to row height and site width and add the core margin.

Q4. What is the aspect ratio and how does it affect the design? The core height-to-width ratio (check the tool's convention). It changes perimeter, pin edge length, wire lengths and how macros fit.

Q5. Why are alternate rows flipped? So adjacent rows share a VDD or VSS rail, which saves area and makes the rail structure regular.

Q6. What is the difference between a pad-limited and a core-limited design? In a pad-limited design, the number of I/O pads around the edge sets the die size; in a core-limited design, the core logic sets it.

Q7. What is the difference between hard, soft and partial placement blockages? Hard blocks all standard cells, soft blocks initial placement but allows optimization buffers, and partial limits local density to a percentage.

Q8. After placement you see heavy congestion between two memories. What would you change in the floorplan? Check channel width against the nets that must pass, the macros' pin orientation, and whether the channel needs a soft or partial blockage; then widen the channel or move a macro and rerun trial placement and global route.

More questions: Floorplan Interview Questions and Answers.

FAQ#

Is floorplanning done before or after synthesis? After, because it needs the gate-level netlist, though teams often do an early floorplan from estimates.

Is power planning part of floorplanning? Many flows treat it as the last floorplan sub-step, because the power grid depends on the outline, macros and rows, and placement depends on the grid.

Can I change the floorplan after placement? Yes, but every change to the outline, macros or rows invalidates placement, taps, endcaps and the power grid.

What is the difference between utilization and placement density? Utilization is a global area ratio used to size the core. Placement density is the target local packing the placer aims for.

Summary#

Size the core from standard-cell area and a utilization you can defend, include macro halos and blockages, and snap to rows and sites. Choose the shape from pins, macros and the parent floorplan, not only from area. Place pins on allowed layers and tracks, add endcaps and well taps, build the power grid, and check geometry, utilization, routability, timing and power before placement.

References#

  • OpenROAD documentation, Initialize Floorplan (ifp) README: initialize_floorplan (utilization, aspect ratio, core space, die/core area, rows, -row_parity, -flip_sites, -additional_sites, -gap, rectilinear floorplans) and make_tracks: https://github.com/The-OpenROAD-Project/OpenROAD (openroad/ifp/README.md)
  • OpenROAD documentation, Pin Placer (ppl) README: place_pins options (-hor_layers, -ver_layers, -min_distance, -corner_avoidance, -exclude, -group_pins): https://github.com/The-OpenROAD-Project/OpenROAD (openroad/ppl/README.md)
  • OpenROAD documentation, Tapcell (tap) README: tapcell endcap and tapcell insertion: https://github.com/The-OpenROAD-Project/OpenROAD (openroad/tap/README.md)
  • OpenROAD-flow-scripts, flow/scripts/floorplan.tcl, flow/scripts/variables.yaml, flow/Makefile and the Nangate45 platform files (config.mk, tapcell.tcl, NangateOpenCellLibrary.tech.lef), plus the gcd and swerv design configurations: https://github.com/The-OpenROAD-Project/OpenROAD-flow-scripts

Written by

Swasti Pujari

Associate Director, Backend Engineering at CoreIC Technologies, with 15+ years in VLSI Physical Design and backend engineering. Previously Practice Head at UST Global, leading Physical Design, STA, EMIR, DFT, and Physical Verification. Delivered 3nm, 6nm and 7nm projects for AMD, Intel, and Microchip, and published IEEE research on hybrid AI-EDA floorplanning.