How your Verilog becomes a manufacturable chip: synthesis, floorplan, placement, clock tree, routing, signoff and the GDS file sent to the foundry.
The big picture
In the last chapter you described circuits in Verilog. A chip factory can't read Verilog. It needs a layout: the exact shape of every transistor and wire, layer by layer, stored in a GDSII file. The chain of tools in between is the RTL to GDS flow, and the people who run its back half are physical design engineers.
Step
What it decides
Result
Synthesis
Which gates
Netlist of standard cells
Floorplan
How big, where the pins and power go
Die, core, rows, power grid
Placement
Where each gate sits
Every cell at an (x, y)
Clock tree synthesis
How the clock reaches every flip-flop
Balanced buffer tree
Routing
Which wires, on which metal layers
Connected layout
Signoff
Is it fast enough, manufacturable, and correct?
Clean STA, DRC and LVS
GDS export
Nothing new: write it out
GDSII file → tape-out
> Loading the animation…
Synthesis
A foundry offers a standard cell library: a few hundred pre-drawn gates (NAND2, NOR3, flip-flops, buffers) in several drive strengths, all the same height so they line up in rows. Each cell has a liberty file entry with its area, its delay for each input slope and output load, and its power.
Synthesis turns your RTL into a netlist that uses only those cells. It first builds generic logic, optimises it (the Boolean algebra and K-map ideas from earlier chapters, done by algorithms), then maps it onto real cells, picking bigger, faster cells for critical paths and small ones elsewhere.
// RTL
assign y = ~(a & b);
// after synthesis (Sky130)
sky130_fd_sc_hd__nand2_1 U1 (.A(a), .B(b), .Y(y));
One line of RTL, one standard cell. A real design maps to thousands.
Floorplanning and power
Before placing anything, the floorplan fixes the die (the whole chip), the core (the area where cells go), the standard-cell rows, the I/O pin positions, and any large blocks such as memories.
It also builds the power distribution network: a grid of wide VDD and VSS straps on the top metal layers, connected down to rails along every row. If the grid is too thin, the voltage sags under load (IR drop) and cells slow down.
utilization = total cell area ÷ core area
Typically 40–70%. The rest is room for the router and for buffers added later.
Placement
Global placement treats the netlist like a system of springs: connected cells pull together, crowded regions push apart, and the solver finds positions that keep wires short. Legalisation and detailed placement then snap every cell into a row slot with no overlaps.
Placement is timing-driven: the tool knows which paths are critical (from static timing analysis) and pulls their cells closer, because wire length is delay.
Clock tree synthesis
Until now the clock was ideal: it reached every flip-flop at the same instant. Real clock nets drive thousands of flip-flops, far too much load for one wire. CTS builds a tree of buffers so that every flip-flop gets a sharp clock edge at nearly the same time.
The leftover difference in arrival times is skew. You met its effect in the timing chapter: skew helps setup on some paths and hurts hold on others. After CTS, tools fix hold violations by inserting delay buffers on paths that are now too fast.
> Loading the animation…
Routing
A chip has a stack of metal layers separated by insulator and joined by vias. Lower layers are thin and used for short local connections; upper layers are thick, for long wires, clocks and power. Each layer prefers one direction (horizontal or vertical), alternating up the stack.
Global routing plans which region and layer each net uses and spots congestion. Detailed routing then draws exact tracks and vias that obey every design rule: minimum width, minimum spacing, via overlap. After routing, the real resistance and capacitance of every wire are extracted and timing is recomputed.
NoteOn advanced nodes wire delay can exceed gate delay. That's why placement, CTS and routing all care about timing, not just connectivity.
Signoff and tape-out
STA: every path meets setup and hold at every process, voltage and temperature corner, using the extracted wire delays.
DRC: the layout obeys all of the foundry's geometric rules.
LVS: the transistors and connections in the layout match the netlist exactly.
Antenna, density, ERC: manufacturing hazards are fixed.
When everything is clean, the layout is written as GDSII and sent to the foundry: tape-out. The name dates from when the file was literally written to magnetic tape.
Doing it yourself
All of this used to need million-dollar tools. Today an open-source flow covers every step on the SkyWater 130 nm process: Yosys for synthesis, OpenROAD for floorplan, placement, CTS and routing, OpenSTA for timing, Magic and KLayout for DRC and GDS, Netgen for LVS. OpenLane wraps them into one script, and TinyTapeout puts small community designs on a real shuttle.