1. Pete Zampardi, Dave Nelson, Steve Rohlfing
Application of a Design of Experiment Approach to Handset Power Amplifier Design
Pete Zampardi, Dave Nelson, Steve Rohlfing

2. DOE Gets a Bad Rap Why Does DOE Get Such a Bad Reputation?
Bad implementation (we don’t advocate it where it doesn’t make sense)
People using them as an excuse not to think (best DOEs are by people that understand the problem!)
People try to use it for crank finding problems rather than crank turning problems (best for optimization, not innovation)
Need a New Level Beyond Black-Belt. We propose “Ninja” that comes in and kicks the black-belts butt!

3. Used Correctly, It Makes Your
Why Use DOE?
MINIMIZES the number of variants necessary to evaluate how factors and their INTERACTIONS affect measured outcomes
Make what you are already doing easier!
Learn more in the process
Provides systematic method for modeling these effects
Example: PAE, GAIN, ACP vs. factors
Reduce design time by providing designers with critical information for new projects (how much do you need to change something to meet new spec?)
Do this on a REAL design. Have something improved when you are done!
Used Correctly, It Makes Your
Your Life Easier!

4. Make-up of Cellular Handset PA
9/12/2018
Make-up of Cellular Handset PA
DOE Factors
Input stage, Output stage,
Bias circuits, Matching
Cint
Rb,out
Ain
Aout
Re,in
For a PA, the circuit is not very complicated
Need to understand how different technology parameters affect the bias circuits, the power transistors, and the passives (matching)
Must look at what each one requires separately then trade-off for best total design
Tseng, UCLA
Can be replaced
With Tuner
Can be replaced
With Tuner
Most of the RF Chain is a Prime Candidate for
Using a Design of Experiments Approach!

5. What Factors Did We Pick?
Minimize Number of On-wafer Variants (record is something like 78!)
On-Chip Factors (27 Layouts)
Input Stage Area (Ain) (+/- 25%)
Output Stage Area (Aout) (+/-7%)
Input Stage Ballast (emitter) (Re,in) (+/-50%)
Output Stage Ballast (Rb,out) (+/- 33%)
Inter-stage Cap Value (Cint) (+/- 20%)
Off-Chip Factors (Not varied here)
Vref Resistor (Icq’s)
Input Match (Source pull)
Output Match (Load pull)
IMPORTANT CAVEAT NUMBER 1:
Realistically, this was too many factors.
because of uncontrolled
variation (laminates, SMT, assembly),
quite a few parts of each variant needed to be
tested making this experiment
more cumbersome than intended.
to avoid this, use SAME panel for variants and
SMTs from same batches
KISS=Keep It Simple Stupid!
Important Safety Tip - Egon
Picked On-Chip Factors Designers
Vary Most (or Asked About Most)

6. Test Results 5 Parts from each variant were tested (maybe not enough).
We tested one subset of the DOE first (just as a check), then completed
the full set
Conditions
Variants (same wafer, etc)
IS-95 Modulation
Temperature=25 C
Vref=4.75 Volts
Vbias=3.2 Volts
Freq=836.5 MHz
Other data (over temperature, modulation, freq, also available)
IMPORTANT CAVEAT NUMBER 2:
Analysis of DOE data requires single point
values, but PA data is continuous.
comparing at a single power, you may
get screwed up someplace else.
To deal with this, we defined new parameters
that describe the curves for the data we
were trying to fit.

7. Parameter Definitions
Gain Slope, Gain Peak, PAE_14, PAE_peak, ACPR_belly,
ACPR_MIN, ACP1_Slope, ACP2_Slope, ACP2_29

8. Summary of Behavior (for preceding data only)
Gain
PAE
ACPR1
ACPR2
Cint   -
1st Area  
2nd Area 
None of the variables affected PAE much
Larger Cint and 1st stage increase gain
Larger Cint and 2nd stage improve ACPR1 and ACPR2
Without changing anything else, larger Cint improves
Gain, PAE, and linearity for this circuit
Score!

9. Data for All Variants IMPORTANT CAVEAT NUMBER 3&4:
(3) Just because there is *Correlation* doesn’t
mean it’s important. Need to check the range
to see if you really care or can use it!
(4) Make sure you vary stuff A LOT – Aout was space
limited in this experiment

10. Position of belly affected
Higher
Lower
Position of belly affected
by A1
Depth of belly
depends
on Cint & A2

11. Cint and A2 improve ACPR1 (at 29 dBm)
ACPR1 (29 dBm)
Higher
Lower
Cint and A2 improve ACPR1 (at 29 dBm)
bigger A1, R1, R2 degrade

12. ACPR2 (29 dBm) ACPR2 improved by Cint, A1, A2, R1 degraded by R2
Higher
Lower
ACPR2 improved
by Cint, A1, A2, R1
degraded by R2

13. Summary of Effects (For This Design)
Peak Gain
Gain Slope
PAE
ACP
Belly
P_ACBelly
ACP1
ACP1Slope
ACP2
ACP2 Slope
A1 ()  
A2 () -
R1 ()
R2 ()
Cint ()
Improves
 Degrades

14. Design Figure: R1, R2, Cint Fixed
Cint=Nominal
If ACPR (belly) < ACPR_29 Part Fails
Design contours can be generated from the equations/software

15. Summary
A DOE design study was performed on a 29 dBm power amplifier. There were a LOT of interactions.
Dependencies of PA performance parameters on RF chain parameters have been identified and should allow better optimization of these circuits (for performance margin, yield)
This approach has since been applied to optimizing ballasting, stage sizes, and interstage matching. Also identified areas outside of die for improvement (laminate, SMT).
Key Lessons Learned:
Keep DOEs simple! Still effective since you can see interactions.
Vary factors enough to impact stuff.
Don’t be surprised other stuff you didn’t vary impacts results.
Need some creativity in parameter definitions for analysis

16. THANK YOU!! Final Thoughts on DOE! Random Acts of Varying Stuff is
NOT a DOE
Best DOE’s Require Careful Thought
“Designing” an Experiment Doesn’t Make It a DOE!
DOE Follows a Certain Methodology
THANK YOU!!