VLSI Design

1

VLSI Designsmd154

Physical design (back-end)

2

Today’s Topics

• Introduction• Partitioning• Floorplanning• Placement• Routing

3

Intro. - Physical design

• Microelectronic system• ASICs• System partitioning• ASIC floorplanning• Placement• Routing

• A City• Buildings• City planner• Architect• Builder• Electrician

Physical design vs. Building a City

4

Intro.- History

In the beginning…

Front-end

Team

Back-end

Team

• Logic design (RTL)

• Synthesis

• Design for test

• Floorplanning

• Placement

• Routing

• Physical verification

• GDSII

5

Intro.- History

Front-end

TeamBack-end

Team

Smaller geometry + higher gate density results in:

•Front-end team and back-end team has to work closer together.

•Demand in new design methodologies.

•Demand for tools with smaller/faster data structures.

6

Partitioning

Goals• Divide circuit/system into smaller subcircuits/subsystems,

called blocks.– Speeds up design process– Can be designed independently– Original system functionality remains intact– Simplifies routing task– May degrade performance

Objectives• Minimize Interconnections Between Blocks (mincut problem)• Minimize delay due to partitioning.

Constraints• Area (not so important on chip level)• Number of terminals (depends on area) • Number of blocks (depends on capacity of placement

algorithm)

7

Partitioning

Partitioning Algorithms

Mincut problem is NP-complete!

Group Migration Algorithms• Start with some (usually random generated) partitions and

move components between blocks to improve results.• Quite efficient.

Simulated Annealing/Evolution Algorithms• Carry out partitioning by using a cost function.• Computationally intensive compared to other methods.

8

Partitioning

Initial Bisections

Cutsize = 9

Final Bisections

Cutsize = 1

1

2

3

4

5

6

7

8

1

5

8

4

2 6 3 7

9

Partitioning

After circuit partitioning phase

• Area occupied by each block can be estimated• Possible shapes of blocks can be ascertained• Number of terminals required by each block is

known• Netlist specifying connections between blocks is

available

Next step: Floorplanning

10

Floorplanning

Goals• Assign shape and location of blocks.• Decide location of I/O pads.• Decide location and number of power pads.• Decide type of power distribution.• Decide location and type of clock distribution.

Objectives• Keep highly connected blocks physically close to each other.• Minimize chip area.• Minimize delay.

11

Floorplanning

Input

Set of blocks.Area estimation.Possible block shapes.Number of terminals.Netlist.

Output

Shapes (Area & Aspect Ratio) and locations of blocks.

Soft Blocks• Flexible shape

• I/O positions not yet determined

Hard Blocks•Fixed shape

•Fixed I/O pin positions

12

Floorplanning

Aspect Ratio BoundsNo Bounds

With Boundslower bound ≤ height/width ≤ upper bound

Block 4

Block 3

Block 2

Block 1NOT GOOD!!

Soft Blocks

Hard Blocks

13

Floorplanning

Design Styles

Full Custom• Floorplanning is needed.

Standard Cell• Fixed cell dimensions. Floorplanning translates into a placement

problem.• Floorplanning may be required for large cells if they are partitioned

into several blocks.

Gate Array• Placement problem.

14

Floorplanning

Optimize Performance• Chip area.• Total wire length.• Critical path delay.• Routability.• Others, e.g. noise, heat dissipation.

Cost = αA + βL, A = total area, L = total wire length, α and β constants.

15

Floorplanning

Area

Deadspace

Minimizing area = Minimizing deadspace

Wire length estimation• Exact wire length not known until after

routing.

• Pin position not known.

• How to estimate?

• Center to center estimation.

16

Floorplanning

Floorplanning methods• Constraint Based.• Linear Programming.

– Mixed Integer Linear Programming (MILP) -> NP-complete.

• Rectangular Dualization Based.• Hierarchical Tree Based.• Simulated Annealing and Evolution Algorithms.• Timing Driven.

17

Floorplanning

Simulated Annealing Algorithm.• Represent floorplan by normalized polish expression.

V

V H

H H 3 4

H 7 2 5

1 6

7 5

6

1

4

23

E = 16H7H25HV34HV

18

Floorplanning

Try to optimize cost by moving blocks around.M1 : Swap two adjacent operands.M2 : Complement some chain.M3 : Swap adjacent two operand and operator.

cost = αA + βL

M1

=>

M3

=>

M2

=>

34V2H5V1H 32V4H5V1H 32V45HV1H 32V45VH1H

1 11

2

1

543 2 2

23

33

4

4

45

5 5

19

Placement

Goals• Arrange all logic cells within the flexible blocks.

Objectives (Ideally)• Guarantee the router can complete the routing step.• Minimize all critical net delays.• Make chip as dense as possible.• Minimize power dissipation.• Minimize cross talk between signals.

Ideally objectives are difficult to define for use in algorithms andeven harder to actually meet ...

20

Placement

Objectives (Compromised)• Minimize total estimated interconnect length.• Meet timing requirements for critical nets.• Minimize interconnect congestion.

Problems• No point in minimizing the interconnect length if we create a

placement that is too congested to route!• Trying to minimize both interconnect congestion and

interconnect length may result in long interconnection delays between some logic cells.

21

Placement

Placements Algorithms

• Constructive Placement– Uses a set of rules to create a constructed placement

• Interative Placement– Take an existing placement and improve it for each

iteration

Min-Cut Placement method1. Cut placement area into two pieces

2. Swap logic cells to minimize cut cost3. Repeat process from step 1, cutting smaller pieces until all logic cells are placed

22

Routing

Routing is• to connecting pins with wires• a NP-hard problem

Constraints• Minimize total wire length• Minimize knees in path• Meet timing budget

Steiner Tree • The optimal routing path is a

Steiner Tree • Finding a Steiner Tree is a NP-

complete problem

Types• Maze routing• Channel routing• Switchbox routing• River routing (Single layer)

Global Routing• Doesn't make any connections

only plans them.• Decides which channels is going to

be used for routing• Marks channels with wire

accommodate capacity

Detailed Routing• Connects pins and pads, creates

vias.

Special Routing• Clock routing

23

Routing

Obstructions • Areas that can only be crossed on a certain layer.

Obstacles • Areas that can’t be crossed on any layer, have to route around.

Channel • Rectangular area with connection points on two sides.

Switchbox• Area with connection points on four sides.

Channel

Switchbox

24

Routing

Lee-Moore (Maze routing) -Guaranties to find the shortest path between two nodes if it exist,

but it only holds for one net.-Demo http://foghorn.cadlab.lafayette.edu/cadapplets/MazeRouter.html

The Left Edge Algorithm (Channel routing)

-No guarantee to success.-Makes global considerations, often resulting in better routing then maze.-Demo http://foghorn.cadlab.lafayette.edu/cadapplets/ChannelRouter.html

25

Routing

Clock and Power routing– A part of floorplanning

CLOCK– Asynchronous/synchronous system– Clock skew and delay needs to be considered.

POWER– Low resistance in metal line need to be

considered.– Power and GND is routed in separate nets.

26

Questions?

?