IAE Academy

FOR CORE AND IT JOBS · CHAPTER 7 OF 7

The chip design flow, from RTL to silicon

This chapter follows a chip from a written specification to the layout sent for manufacture. It shows the timing checks that decide when a design is done, and when an FPGA costs less than an ASIC. Front-end and back-end chip roles each work on part of this flow.

Updated 30 Oct 2026

What you will be able to do

You will list the steps from RTL to silicon in order and say what each takes in and gives out. You will compute setup and hold slack and decide whether a path meets timing. You will say which steps are front-end and which are back-end work. You will compare an FPGA with an ASIC and find the volume above which the ASIC costs less per chip.

Before you start

  1. Chapter 6, Describing hardware with Verilog

  2. Chapter 5, Fabrication and layout

  3. A flip-flop, which stores its input at a clock edge (Chapter 6)

  4. Clock period

  5. Adding and subtracting times in ps and ns (1 ns = 1000 ps)

  6. Dividing one cost by a number of units

  7. Solving a simple equation for one unknown

From specification to a gate-level netlist

A chip is built by turning one description into a more detailed one until it is a set of shapes for the masks. Steps 1 to 5 live in the logic world: nothing has a position on silicon yet.

Functional verification runs the RTL in a simulator against a testbench: Verilog code, like Chapter 6's, that applies inputs and checks the outputs. Coverage measures how much of the design the tests touched. Logic synthesis maps the RTL onto a standard-cell library: pre-drawn gates and flip-flops of one fixed height, so cells sit side by side in rows.

Step

Goes in

Comes out

1. Specification

Product need: function, speed, power, area

Written spec of behaviour and limits

2. RTL design

Spec

Verilog modules

3. Functional verification

RTL and a testbench

Pass or fail, plus a coverage report

4. Logic synthesis

RTL, standard-cell library, timing targets

Gate-level netlist of library cells

5. Design for test

Netlist

Netlist with test structures, plus test patterns

Design for test makes each manufactured chip testable. In full scan, each flip-flop becomes a scan flip-flop: a multiplexer at its input picks either normal data or the previous flip-flop's output. Linked this way, the flip-flops form a shift register called a scan chain. In test you shift a pattern in, clock once to capture the logic's response, and shift it out. Automatic test pattern generation (ATPG) computes patterns that expose modelled manufacturing faults.

FAULT COVERAGE

Holds for a stated fault model, such as each node stuck at 0 or at 1. is the number of modelled faults the patterns expose; is the number of modelled faults in the netlist.

SCAN TEST LENGTH

Holds for one scan chain of flip-flops and patterns, each new pattern shifting in while the previous response shifts out. Each pattern takes shift cycles and 1 capture cycle; the last response needs more to shift out. counts test clock cycles.

Physical design and timing closure

From step 6 on, the netlist gets a place on silicon. Placement, the clock tree and routing add real wires, so delays change and timing is checked again. Fixing and rechecking until every path passes is called timing closure.

Power planning lays a grid of wide metal straps that brings and ground to every row with little voltage drop. Clock tree synthesis builds a network of buffers so the clock edge reaches all flip-flops at nearly the same time; the small difference left is the skew. Physical verification makes two checks. The design rule check (DRC) compares every width and spacing with the rules of Chapter 5. Layout versus schematic (LVS) rebuilds the circuit from the layout and compares it with the netlist.

Step

Goes in

Comes out

6. Floorplanning and power planning

Netlist, die size, pin positions

Block outlines, cell rows, power grid

7. Placement

Floorplan, netlist

Every cell at a position in a row

8. Clock tree synthesis

Placed design

Buffered clock network

9. Routing

Placed design with clock tree

Metal wires for every connection in the netlist

10. Static timing analysis

Routed design, wire parasitics, cell delays from the library

Timing margin (slack) of every path

11. Physical verification

Layout, netlist, design rules

DRC and LVS reports

12. Tape-out

Layout that passes DRC, LVS and timing

Final layout file sent for mask making

Static timing analysis adds up delays along each path between flip-flops without simulating any input. A launching flip-flop sends data after its clock edge (its clock-to-output delay, one value here); the logic takes between and . To store data reliably, the capturing flip-flop needs it steady for the setup time before its clock edge and the hold time after it. The skew is the capture clock's arrival time minus the launch clock's.

SETUP SLACK

Holds for two flip-flops on one clock, capture one period after launch. Data arrives at and is required by ; slack is required minus arrival. All terms in ps or ns. The path meets setup when slack is zero or more.

HOLD SLACK

Holds for the same pair, checked on the same clock edge. The earliest new data arrives at and must not come before ; slack is the difference. Negative slack means new data corrupts the value being captured.

WORKED EXAMPLE

A path has T = 1 ns (1 GHz), t_cq = 60 ps, t_max = 700 ps, t_su = 40 ps and s = 0. After routing, the wires add 180 ps to t_max and skew becomes s = -30 ps. Find the setup slack before and after routing.

Solution
  1. Work in ps so every term has one unit: T = 1000 ps. Before routing, slack = 1000 + 0 - 40 - 60 - 700 = 200 ps, so the path passes with room.

  2. Routing adds the delay of the real wires, which the earlier estimate did not include: t_max = 700 + 180 = 880 ps.

  3. Negative skew means the capture clock arrives 30 ps before the launch clock, which takes time away from the data: s = -30 ps.

  4. After routing: slack = 1000 - 30 - 40 - 60 - 880 = -10 ps.

  5. Negative slack is a setup violation. The path must get faster, such as with larger cells, or the period must rise to at least 30 + 40 + 60 + 880 = 1010 ps. Then the analysis runs again.

Answer: Before routing 200 ps; after routing -10 ps, a setup violation.

WATCH OUT FOR

Negative setup and hold slack need different cures. Test: change . If the slack moves, it is setup: a slower clock or faster logic fixes it. If not, it is hold: the fastest path needs added delay.

Who does what, and FPGA against ASIC

Front-end roles cover specification, RTL design, verification and often synthesis (steps 1 to 4): they work on what the chip does. Back-end roles cover steps 6 to 12: where everything goes, whether it meets timing, and whether it can be manufactured. Design for test, step 5, sits between them.

A field-programmable gate array (FPGA) is a chip of configurable logic blocks and wires. Your Verilog is synthesised, placed and routed onto those blocks and loaded as a configuration. An application-specific integrated circuit (ASIC) goes through all twelve steps. You prototype on an FPGA to test the design on real hardware, then decide whether the volume justifies an ASIC.

Property

FPGA

ASIC

One-time cost

Low: no masks

High: masks, tools, verification

Cost per unit

High

Low

Speed and power, for the same logic

Lower speed, higher power

Higher speed, lower power

Change after delivery

Load a new configuration

Needs a new mask set

Time to first hardware

Hours to days

Months through the full flow

COST PER CHIP

Holds when the one-time cost (non-recurring engineering) is spread evenly over chips. is the cost to make one chip; all costs in one currency.

BREAK-EVEN VOLUME

Holds when and ; the terms are unit costs. Setting the two equal gives ; solve for . Above the ASIC is cheaper per chip.

WORKED EXAMPLE

An ASIC has a one-time cost of USD 900,000 and costs USD 10 per chip. An FPGA solution has no one-time cost and costs USD 55 per chip. At what volume does the ASIC become cheaper per chip?

Solution
  1. Write each cost per chip in USD: the ASIC's is 10 + 900000/N and the FPGA's is 55 + 0/N = 55. They cross where they are equal: 10 + 900000/N = 55.

  2. Subtract 10 from both sides to isolate the one-time share: 900000/N = 45.

  3. Multiply by N and divide by 45 to solve for N: N = 900000/45 = 20000.

  4. Check: 900000/20000 = 45, so 10 + 45 = 55, matching the FPGA.

Answer: N* = 20,000 chips. Above that volume the ASIC costs less per chip.

Check yourself

CHECK YOURSELF 1

Two flip-flops share one clock with period T = 800 ps. The launch flip-flop has t_cq = 70 ps. The logic has t_max = 600 ps. The capture flip-flop has t_su = 50 ps. The skew is s = +20 ps. Find the setup slack in ps.

Answer and solution

Answer: (accepted range )

  1. Use slack = T + s - t_su - t_cq - t_max, with every term in ps.

  2. slack = 800 + 20 - 50 - 70 - 600.

  3. 800 + 20 = 820; 820 - 50 = 770; 770 - 70 = 700; 700 - 600 = 100.

  4. The slack is 100 ps, so the path meets setup.

CHECK YOURSELF 2

After routing, a path shows negative hold slack. Changing the clock period T has no effect on it. Which action fixes it?

  1. (A)

    Raise the clock period T

  2. (B)

    Add delay to the fastest path through the logic

  3. (C)

    Replace the logic with faster cells on the slowest path

  4. (D)

    Increase the setup time of the capturing flip-flop

Answer and solution

Answer: (B)

  1. The hold slack is t_cq + t_min - t_h - s. T does not appear, so a slower clock (A) cannot change it.

  2. Hold slack rises when t_min rises, so adding delay to the fastest path fixes it (B).

  3. Faster cells (C) lower t_max, which helps setup. Faster cells can only lower delays, so they cannot raise t_min and cannot fix hold.

  4. A larger setup time (D) is a setup quantity and does not enter the hold slack.

  5. Answer: B.

CHECK YOURSELF 3One or more options are correct

In the split this chapter uses, the twelve-step flow divides into front-end and back-end work, with design for test between them. Which of these steps are front-end work?

  1. (A)

    Logic synthesis

  2. (B)

    Routing

  3. (C)

    Specification

  4. (D)

    Physical verification

Answer and solution

Answer: (A), (C)

  1. Front-end roles cover steps 1 to 4: specification, RTL design, functional verification and, often, logic synthesis.

  2. Logic synthesis is step 4, so A is front-end work in this split.

  3. Specification is step 1, so C is front-end.

  4. Routing is step 9 and physical verification is step 11. Back-end work covers steps 6 to 12, so B and D are back-end.

  5. Answer: A and C.

CHECK YOURSELF 4

A design uses one scan chain of L = 100 flip-flops and P = 50 test patterns. Each new pattern shifts in while the previous response shifts out, with 1 capture cycle per pattern, and the last response needs L more shift cycles. Find the number of test clock cycles.

Answer and solution

Answer: (accepted range )

  1. Use N = P(L + 1) + L.

  2. Each pattern costs L + 1 = 101 cycles (100 shifts and 1 capture).

  3. P(L + 1) = 50 x 101 = 5050.

  4. Add the final shift-out of L = 100 cycles: 5050 + 100 = 5150.