“performance model for inter-chip busses”1 performance model for inter-chip busses considering...

“Performance Model for Inter-chip Busses” 1

Performance Model for Inter-chip Busses Considering Bandwidth and Cost

ISCAS 2005

Authors: Brock J. LaMeres University of Colorado, Boulder, CO

Sunil P. KhatriTexas A&M University, College Station, TX.


Problem : Packaging Limits Performance

• Transistor Technology is Faster than Package Technology

IC

“Moore’s Law” - # of transistors will double every 18 months

Package

“Rent’s Rule” - # of I/O will double in next 10 years


• Inductive noise in packaging• Package performance model

– Relationship with inductive parasitics

– “Bandwidth per unit cost” metric

• Experimental results – Model versus SPICE

• Conclusions

Outline


1) Supply Bounce

• Switching current through inductive packaging induces voltage:

• Multiple Signals Switching Increase the Problem:

11bnc

diV L

dt

111

nk

bnck

diV L

dt

Inductive Noise in Packaging

n = # of drivers sharing the power/gnd pin (L11).

L11 = Inductance of pwr/gnd pin that current is being switched through.


2) Pin-to-Pin Coupling

• Switching Signals Couple Voltage onto Neighbors:

• Multiple Signals Switching Increase the Problem:

Inductive Noise in Packaging

1k

couple k

diV M

dt

11

nk

couple kk

diV M

dt

M1k = Mutual Inductance between pwr/gnd pin and kth signal pin.


Analytical Model

• Bus Parameters

WBUS : # of signals in the bus

G S PS S S G S S P S S G S S PS S

WBUS

NG : # of Grounds in the bus

G S PS S S G S S P S S G S S PS S


Analytical Model

• Bus Parameters

SPG : (# of Signals) : (# of PWR’s) : (# of GND’s)

= 4:1:1 in above example

G S PS S S G S S P S S G S SP S S

Repetitive Pattern of Signal, Power, and Ground Pins


Analytical Model

• Bus Performance Description

dt

dv v(t)

t

load

dv dislewrate Z

dt dt

Slewrate


Analytical Model


v(t)

t

(0.8) DDrise

Vt

slewrate

Risetime

(0.8)VDD

90%

10%

VDD


DATA

Analytical Model


DATA DATA

Minimum Unit Interval

UI

min minmax

1(1.5) riseUI t

DR


Analytical Model


DATADATA DATA

Bus Throughput

max maxBUSTP W DR

DATADATA DATA

DATADATA DATA

Tx WBUS Rx


NOISE

Analytical Model

• Bus Performance Limits

v(t)

t

Maximum Acceptable Ground Bounce

pVDD

VDD

bnc MAX DDV p V (ptypical = 5%)


Analytical Model

• Model Development

Maximum Ground Bounce

111

2

busWbus

gnd bnc DD kkg

W L di diV p V M

N dt dt

Self Contribution

Coupling Contribution


Analytical Model


Maximum Slewrate

- pull out (di/dt)

- convert to (dv/dt)

max 111

2

bus

DD load

Wbus

kkg

p V Zdv

dt W LM

N


Analytical Model


Minimum Risetime : But since

- convert slewrate to risetime

111

2

min

0.8busW

busk

kg

riseload

W LM

Nt

p Z

min

max

(0.8) DDrise

Vt

slewrate


Analytical Model


Maximum Datarate : Also, since

- convert Risetime to Datarate

max

111

2

1.5 0.8bus

load

Wbus

kkg

p ZDR

W LM

N

Maximum Throughput : Finally, we have

max maxBUSTP W DR

minmax

1(1.5) riset

DR


Experimental Results

• Per-pin and Bus throughput values computed for 3 packages

– Compared our model with SPICE simulations

QFP – Wire Bond BGA – Wire Bond BGA – Flip-Chip



• QFP Wire-Bond Package Simulations

Per-Pin Data-Rate Bus Throughput

- Throughput reaches an asymptotic limit as channels are added

ModelSimulation



• BGA Wire-Bond Package Simulations


- Level 1 : BGA Increases Performance Over QFP



• BGA Flip-Chip Package Simulations


- Level 2: Flip-Chip Increases Performance Over Wire-Bond



• Cost Must Also Be Considered in Analysis

Bandwidth Per Cost

• This Metric Represents “Cost Effectiveness of the Bus”

bus

TPBPC

Cost

Units = (Mb/$)



• Bandwidth Per Cost Results

Faster Narrower Busses = More Cost Effective


• Presented a model for bus performance– Accounts for inductive parasitics in packaging

– Developed a “bandwidth-per-unit-cost” metric to evaluate packages

• Experiments indicate – Strong agreement with SPICE simulation

– Throughput for a bus reaches asymptotic limit as channels added• Increase in channels compensated by decrease in per-pin throughput

• Optimal number of channels is small (4-8)

• “Bandwidth-per-unit-cost” experiments indicate– Flip-chip packages are actually most cost-effective

– Narrower buses have better bandwidth-per-unit-cost

Conclusions


Thank you!


Backup Slides


Example

PACKAGE - Rent’s Rule

IC Core - Moore’s Law

On-Chip - 8 bit Data Bus - 300 Mb/s

Package - Need

(8)(300M) = 2400 Mb/s


2400 Mb/s

Example

Need:

QFP – Wire Bond BGA – Wire Bond BGA – Flip-Chip- 4 bits wide, SPG=2:1:1 - 1 bit wide, SPG=2:1:1 - 1 bit wide, SPG=2:1:1 - 16 bits wide, SPG=4:1:1 - 1 bit wide, SPG=4:1:1 - 1 bit wide, SPG=8:1:1

X

X X

X


Example

• Cost of Each Bus Configuration

Most Cost Effective:

- BGA-WB- Wbus = 1- SPG = 2:1:1



• Cost per Bus Configuration

• Performance Increases with Cost (Package, SPG)

$

“performance model for inter-chip busses”1 performance model for inter-chip busses considering...

Documents