Transcript ppt

Modeling and Design for Beyond-the-Die
Power Integrity
Yiyu Shi, ECE Dept., Missouri Univ. of Science and Technology
(formerly University of Missouri-Rolla)
Lei He, EE Dept., Univ. of California, Los Angeles
Importance of Power Integrity
●
Power supply noise is a major threat for circuit reliability
in 45nm and beyond
–
reduces noise margin of digital circuits
–
shifts the operating point of analog circuits
–
decreases the effective driving strength of the gates
–
causes output signal distortion (e.g. jitters) impairing signal
integrity
ICCAD’10
2
Simultaneous Switching Noise (SSN)
●
●
●
●
a major threat to the power integrity
occurs due to a very large amount of instantaneous P/G
current from simultaneously switching gates
mainly inductive
most significantly observed around the output pads of
the chip
–
–
–
ICCAD’10
large I/O buffers
clock synchronized I/O
Large inductance in package
3
Power Delivery System
●
three distinct peaks
–
–
–
●
●
~kHz (power
regulator/board)
~MHz (package/board)
~100MHz
(chip/package)
significant noise near
the largest peak
need accurate models
to capture it
ICCAD’10
4
Shi et al, “stochastic current prediction enabled frequency actuator for runtime resonance noise reduction”, ASPDAC’10
Outline


Modeling

Chip Models

Package and Board Models
Design

I/O planning and placement

Decap Allocation

Layer Stacking and P/G Plane Stapling
ICCAD’10
5
Models for Chip/Package/Board

impossible to put detailed models of chip, package and board
together for the simulation due to the high complexity

need some simplified models that preserve only necessary
information for the simulation

but how?
ICCAD’10
6
Transistor Models
●
most accurate
●
require detailed info about the circuit and process
parameters, which vendors are reluctant to provide
●
not all simulators are fully compatible
●
slow simulation speed
●
no convergence guarantee
ICCAD’10
7
Current Source Model
●
model the chip I/O as a time variant/invariant current
source with parasitic R and C
voltage drop ↓
↑
switching current ↓
●
the non-linearity of the I/O buffer is ignored => negative
feedback effect is ignored
ICCAD’10
8
IBIS Models

I/O Buffer Information Specification

a universal standard for describing the buffers
using data in ASCII text format
– Not really models
– just behavioral data to be used by
simulators
started in the early 90s to promote toolindependent I/O models for system-level signal
integrity work
IBIS 3.2 is standardized: ANSI/EIA-656-A and
IEC 62014-1
IBIS 4.1 incorporates links to VHDL-AMS and
Verilog-AMS



ICCAD’10
Wiki: IBIS is a group of longlegged wading birds in the
family Threskiornithidae
9
Elements of an IBIS Model
ICCAD’10
10
Pros and Cons of IBIS Models

Pros
–
–
–

Cons
–
–
–
●
simulate much faster than SPICE model
protect circuit and process intellectual properties
easy portability and guaranteed convergence
extrapolation required when load is out of the range (inaccurate)
model regeneration required when the package parasitics change
cannot capture the dynamic characteristics as the data relies
primarily on static characteristics
Only good when the I/O speed is not high!
ICCAD’10
11
Other Models for Chip I/O…

use radial basis function (RBF) to represent the I/O
dynamic behavior
–
–

accurate
intractable for complex driver circuits with multiple ports
use spline functions with a finite time difference
approximation
–
–
ICCAD’10
include the previous time instances of the buffer output
voltage/current
cannot be extended to highly nonlinear buffers
12
Lumped/Distributed Models for Package/Board
●
Lumped models
–
–
–
●
use simple geometry with a few RLC
elements (e.g. π equivalent circuit)
efficient but lack accuracy
should only be used for low
performance/speed design
Distributed models
–
–
–
–
ICCAD’10
run parasitic extraction
huge number of RLC elements
model reduction or other simplification
techniques are needed to reduce
complexity
High computational cost
13
S-Parameters 101




measured by sending a single frequency signal into the network
and detecting the exit waveform at each port
frequency dependent, load dependent
can be obtained using a 3D full-wave EM simulator such as
HFSS or using vector network analyzer (VNA)
By sweeping over a wide frequency range, they can reveal
frequency-dependent characteristics (e.g. skin effect and
dielectric conductance effect)
 S11
S  
 S 21
ICCAD’10
S12 

S 22 
14
Simulation with S Parameters
●
●
simulated directly using convolution-based methods in
frequency domain
or synthesize an RLC circuit from S-parameters in time
domain
–
–
–
●
create a circuit template with a certain topology
convert the measured S parameters to Y or Z parameters
matching the Y/Z parameters of the template and the
measured Y/Z parameters to determine the element values in
the template
put some stringent requirements on S-parameters
–
–
ICCAD’10
passivity (and thus stability and causality)
but hard to satisfy while maintaining accuracy
15
Importance of Co-Simulation
●
A differential pair from chip to package to board
Comparison of the S11 parameter and the power supply voltage from chip,
package and board co-simulation and these from separate simulation.
ICCAD’10
16
Possible Co-Simulation Flows
Frequency domain
Frequency model
for circuit I/O
Time domain
IBIS model
for circuit I/O
S parameters
for package/board
Ckt realization of
S parameters
for package/board
Inverse Transform
ICCAD’10
17
Outline


Modeling

Chip Models

Package and Board Models
Design

I/O planning and placement

Decap Allocation

Layer Stacking and P/G Plane Stapling
ICCAD’10
18
I/O Planning and Placement

Flip-chip design
 Assign
pins and pads to signals and power/ground supply
Xiong et al, ““constraint driven I/O planning and placement for chip-package co-design”, ASPDAC’06
ICCAD’10
19
Rule #1
●
●
separate the P/G pins and pads for analog and digital signals
whenever possible
minimize the digital noise coupled to the analog portion
ICCAD’10
20
Rule #2
●
●
●
●
SSN is negatively correlated to the ratio of # of P/G
pads/pins to # of signal pads/pins
insert as many P/G pads and pins as possible
total inductance ↓ (parallel connection)
the slew of the SSN v.s. # of switching I/O buffers curve ↓
obtained from Q3D extraction
ICCAD’10
21
Decoupling Capacitor Allocation
●
●
●
short power and ground planes at high frequencies to control voltage
fluctuations
discrete passive components with a given capacitance with parasitic
resistance and inductance
Determine the optimal decap allocation strategy
ICCAD’10
22
Decap Allocation
●
●
considering the congestion from signal and power routing,
decaps can be inserted only at selected slots
usually minimize the total decap cost subject to power
integrity and congestion constraints
cost(D )
minD
i
i
G0 x(D,t)  C0
dx(D,t)
dx(D,t)
  (Di M i x(D,t)  Di N i
)  Bu(t)
dt
dt
i
y(D,t)  LT x(D,t)
  y (D,t)dt  U
i
i

Before decap allocation
ICCAD’10
After decap allocation
Hao et al, “Off-chip decoupling capacitor allocation for chip package co-design,” DAC’07
Chen et al, “Noise-driven in-package decoupling capacitance insertion,” ISPD’06
23
Layer Stacking and P/G Plane Stapling
●
●
in high performance flip-chip package, multiple layers are typically
used for P/G planes and signal routing
Determine the number of layers and the locations of the vias to staple
them
ICCAD’10
24
Determine the Number of Layers
●
The # of layers depends on
–
–
–
–
cost
the # of the signals to be routed
the cross-talk constraints of these signals
the #of voltage domains, which constraints


●
the # of power plane layers
how a layer should be partitioned and shared by multiple voltage domains
usually multiple P/G planes are used to keep the power
supply noise low and to shield the signal routing layer
–
ICCAD’10
If affordable, shield every routing layer by alternated
power/ground planes in between
25
Stapling Rules
●
●
●
the resonance frequency ↑ as the number of vias ↑
the locations of the vias do not have a significant impact on
the resonance frequency. Instead, they change the
inductance of the package.
a centered via distribution always has a lower inductance
than a uniform via distribution  Always cluster P/G vias
for each power domain!
ICCAD’10
centered
uniform
Zhao et al, “Effects of power/ground via distribution on the power/ground performance of C4/BGA packages,” epep’98.
26
Conclusions
●
●
●
Power integrity has become an increasingly important
design consideration for circuit designs in 45nm technology
and beyond
We have provided an overview of power-integrity driven
modeling and design issues for beyond the die
We have discussed
–
–
–
ICCAD’10
background of simultaneous switching noise (SSN) and its
significance to the circuit designers
various models of different accuracy and complexity for the
board, package and chip
different design techniques to suppress SSN
27