9 th International Front-End Electronics meeting Fermilab – Argonne 2014 A. Dragone*, P....
Transcript of 9 th International Front-End Electronics meeting Fermilab – Argonne 2014 A. Dragone*, P....
9th International Front-End Electronics meeting Fermilab – Argonne 2014
A. Dragone*, P. Caragiulo, B. Markovic, R. Herbst, K. Nishimura, B. Reese, S. Herrmann, P. Hart, G. Blaj, J. Segal, A. Tomada, J. Hasi, G. Carini, C. Kenney, G. Haller
ePix: a class of architectures for second generation LCLS cameras
RED Integrated Circuits Dept.
2
Outline
• Introduction- A few words on LCLS detector needs (more in Sven Herrmann’s talk)
• The ePix approach: platform based ASIC architectures- Architectural choices
- ePix platform
• ePix100 and ePix10k ASIC variants- Pixel designs
- Prototypes performance
• Summary and conclusions
FEE 2014, Fermilab-Argonne
LCLS: the first hard X-ray FEL source
LCLS - Undulators Hall
3FEE 2014, Fermilab-Argonne
• Operational since 2009 • Six instruments:
• AMO (500 – 2000eV)• CXI (5 – 10keV)• MEC (4 – 10keV)• SXR (250 – 2000eV)• XCS (5.5 – 10keV)• XPP (4 – 10keV)
LCLS – main characteristics
• Pulsed (up to120Hz)• Pulses are very short (~30 - 200fs) • Monochromatic • Extreme peak brightness • Spatially coherent
Other FELs
• FLASH - (since 2005, DESY, Germany)• SACLA XFEL - (since 2011, Japan) • PAL XFEL (South Korea)• SwissFEL (Swiss)• European XFEL (Germany)
Pulsed sources Integrating detectorsHigh repetition rate Fast readout
No matching commercial camera available:• Compromise• Dedicated detector development
4
Detector needs (more in S. Herrmann talk)
•High Dynamic Range Front End (1- >104 photons )- The extremely high brilliance of X-ray FEL sources
enables single shot experiments with sufficient resolution (ρ=1/√(Nph·DQE))
•Large Detector Area (common to several experiments)- Diffraction techniques with atomic resolution require
large angular coverage (~2rad)
•Small Pixel Size and Low Noise- Coherence property of the X-ray FEL sources
enables X-ray Photon Correlation Spectroscopy techniques. To resolve speckle patterns small pixel 50um or below and large area detector are often required.
FEE 2014, Fermilab-Argonne
S. Boutet et al., “High-Resolution Protein Structure Determination by Serial Femtosecond Crystallography”, Science.1217737 (2012).
5
Parallelism and time division multiplexing: some guidelines
- S
en
sor
arc
hite
ctu
re c
om
ple
xity
an
d d
eve
lop
me
nt
time
incr
ea
se
+ ASIC Development Time
Serial Readout
Column Parallel Readout
Hybrid Pixel
3D
Hybrid Pixel Hybrid Pixel
ADC
ADC
ADC
Column Parallel Readout
+ Frame Rate
Min Parallelism
Max Parallelism
External A to D conversion variants
- Input capacitance (noise)+ Power per channel+ ASIC front-end complexity- System complexity+ Radiation tolerance (technology)
+ Frame rate- Power distribution complexity+ ASIC back-end complexity+ System scalability
+ Analog pixel
+ No digital activity during integration
+ More space for filtering
+ Robust high res. low power ADC
+ Minimum clock rates
+ Frame rate not limited by the readout time- Long hold time in the analog memory- Larger balcony (gaps)FEE 2014, Fermilab-Argonne
6
Modularity and scalability: Platform based designsD
AQ
• Cameras require a top down design approach using common platforms.• The platform is defined by standardized interfaces and communication protocols between the
various components of the detector• A minimum set of key components has multiple variants (ASIC and assemblies)
• Several variants of ASIC/sensor combinations with a common communication BE up to the DAQ.
• Variants can be housed in several module assembly
• Facilitates integration of one or multiple detectors into a combined camera of various sizes and shapes (rectangular planes, curved arcs)
• Combinations can even include modules with different functionalities
Standard communication protocol between DAQ and BE electronics
Standard configuration and communication protocol with minimum I/Os between ASIC and BE electronics.
BE electronics
ASICs with a common BE and different functionalities designed at the pixel level
Sensors with different pixel size to match different pixel implementations
High Dynamic Range variant
Ultra low noise variant
Multiple compatible modules assembly
FEE 2014, Fermilab-Argonne
352
96
4
ΣΔ ΣΔ ΣΔ
DEC DEC DEC
ΣΔ
DEC
Temp
Row
sel
ecto
rSerializer
LVDS Output Buffers
CMD unit SACI
Bias
Global Registers
ePix architecture highlights
7
ePix100 ePix10kMode of Operation INT INT-with auto-rangingTechnology 0.25 µm 0.25 µmPixel size 50x50 µm2 100x100 µm2
Array 352x384 176x192Full Size Full Reticule Full ReticuleFrame rate 120Hz (> 1kHz) 120Hz (> 1kHz)Range 220ke- 22Me-
Effective ENC <100e- (50e-) <350e- (120e-)FE Gain 6.5µV/e- 6.5µV/e- ,64nV/e-
Polarity positive positiveFiltering LP + cds LP + cdsS/N 44 8.8DC Power cons. < 10µW/pix < 10µW/pix Power pulsing (PP) Yes YesPP Power cons. < 3µW/pix @ 120Hz < 3µW/pix @ 120Hz
Pixel Matrix: fully analog. Different designs for each detector variants• Pre-amplifies signals • Filters signals• Samples signals (with CDS)
Balcony (ePix platform): • Controls Configuration• Controls Acquisition and Readout• Performs Analog to Digital Conversion• Provides Calibration and Monitoring support
FEE 2014, Fermilab-Argonne
ePix(A) intermediate step – status of the project
352
96
4
Buf. Buf. Buf.
Temp
Row
sel
ecto
r
Diff. Analog Buffer
CMD unit SACI
Bias
Analog Monitor Global Registers
analog balcony
• Columns are read sequentially
• Limited read-out speed (120Hz - 8ms)
• 5MHz clock required• Back-end noise limited
(ePix100)
Column selector
ePix (P) - minimum area matrix prototypes with analog balcony
ePix (A) – final size with analog balcony
ePix (F) - full matrix with digital balcony
ADC
ADC
ADC
ePix100p designed, fabricated testedePix10kp designed, fabricated tested
To mitigate risks and to bridge the gap of the development time and intermediate version of ePix using external ADC
ePix design phases:
ePix100a designed and in fabricationePix10ka designed and ready for submission
Digital balcony in design
FEE 2014, Fermilab-Argonne 8
SACI
Acquisition
Idle phase: non vital resources kept in power saving mode
Wake up phase: all resources get ready to run
Acquisition phase: analog processing in pixels active, no digital activity (balcony in power saving mode)
Readout phase: matrix readout A/D conversion, data serialization and transmission (non vital pixel analog blocks in power saving mode)
Global Registers, In pix registers (gain, calib., analog monitor)
8ms
LCLS Pulses
8ms
… …ROclk
AQCUse only 3 signals
Res, AQC and ROclk
Use only 4 signalsand a control interface based on a
SLAC configuration protocol (standard platform)
Res
ePix phases of operation
9FEE 2014, Fermilab-Argonne
Pixel Matrix with Random Access
• In configuration phase a command unit pushes addresses into the counters providing access to the pixels. Single rows or columns can also be accessed in parallel as well as the full matrix
• In acquisition phase the row counter (both in the ver. A) scan the matrix in a windows defined by the addresses stored in the Rows and Columns start and stop registers
Pix Bank 1352x96
Pix Bank 2352x96
Pix Bank 3352x96
Pix Bank 4352x96
Row
de
cod
er
(Gra
y to
On
e H
ot)
352
96 96 96 96
Column Decoders (Gray to One Hot)
Bin
ary
Pro
gram
mab
le T
rue
Gra
y C
ount
er (
9 bi
ts)
9
Binary Programmable True Gray Counter (7 bits)
7
Row
Sta
rt R
eg.
Row
Sto
p R
eg.
9
9
Col Start Reg. Col Stop Reg.
7 7
ROclk
Pixel (0,0)
ePix pixel matrix access protocol
10FEE 2014, Fermilab-Argonne
SLAC ASIC Control Interface (SACI)
Motivation
• Need simple serial interface to ASICs for
configuring registers and sending
commands.
Standard Options Not a Great Fit
• SPI: No backpressure. No way for ASIC
to signal that it is done with a command
or ready for new data. Requires polling.
• I2C: Backpressure possible through clock
stretching, but complex protocol and
implementation.
Master/Slave Serial Interface
4 Signals
• 3 shared: saciClk, saciCmd, saciRsp
• 1 dedicated select line per slave: saciSelL
Allows multiple slaves on same SACI bus. (Similar to SPI.)
• 320 Standard Cells
• 98.52 µm x 540.20 µm
11FEE 2014, Fermilab-Argonne
Why sigma delta?
12
Noise Shaping
We are here
Ʃ-Δ
SAR
Pipeline
Flash
Ʃ-Δ
Sigma-delta present several advantages:• They are compatible with frame rates of 500-1kHz required by the application
• Allow trading resolution for clock speed (oversampling ratio). Different pixel variants may require different resolutions
• They provide noise shaping
• Have a pseudo-random nature which reduces cross-talk effects between channels
• Require low precision analog blocks and have simple implementation (small area)
• Have potential for upgrades in more advanced technologies (heterogeneous integration)
Optimal ADC architecture is a compromise among sampling rate, resolution, digital multiplexing, serialization rate and development time
FEE 2014, Fermilab-Argonne
• Gain calibration (per pixel)- charge can be injected in each single pixel or group of pixels. A 10-bit DAC can set the amount of
injected charge in constant or automatic scan mode.
• Frame common mode tracking (In frame)- at the end of each frame the baseline and the max charge values are digitized on each column.
This feature provides the possibility to track baseline shifts.
• Temperature tracking and compensation- a temperature variations compensation scheme is present.
• Analog and Digital monitors- critical internal nodes can be monitor during debugging and to check the correct functioning
during operation
ePix calibration, compensation and monitoring
13FEE 2014, Fermilab-Argonne
“Any detector is only as good as it’s calibration” S. Gruner
ePix100 pixel architecture
14FEE 2014, Fermilab-Argonne
ePix100 noise weighting function
15
Vout at 100ph @ 8keV Vout at 1ph @ 8keV
FEE 2014, Fermilab-Argonne
ePix100a pixel layout 50µm x 50µm
~ 28 million transistors
16
50 µm
50 µm
FEE 2014, Fermilab-Argonne
ePix100 prototype 96x96 pixels
ePix100p balcony close up ePix100p pixel close up
ePix100p pixel layout
ePix10k pixel architecture
17FEE 2014, Fermilab-Argonne
ePix10k pixel response
• High Gain = 42mV/fC or 6.5µV/e-
• Low Gain = 398mV/pC or 64nV/e-
• Auto-ranging• Forced Switched Low Gain
Non-Linearity < 0.05%
18
•1024 pixels•Auto-ranging (multiple gains), •Multiple sampling per train, •Power pulsing
D. Freytag et al., IEEE Nucl. Sci. Symp. Conf. Rec. (2008) 3447
KPiX SLAC; similar implementation successfully used in HEP:
4 Modes of operation:
FEE 2014, Fermilab-Argonne
Simulated Noise: 120e- r.m.s.
ePix10ka pixel layout 100µm x 100µm
~ 20 million transistors
19
100 µm
FEE 2014, Fermilab-Argonne
ePix10k prototype 48x48 pixels
ePix10kp balcony close up ePix10kp pixel close up
ePix10kp pixel layout
100 µm
20
ePix100a expected power consumption
DC mode Power pulsed mode
10 µW/pixel 3 µW/pixel
Eiger 8.8µW/pixel
Timepix3 14µW/pixel
Medipix3 9µW/pixel
FEI4 10µW/pixel
FEE 2014, Fermilab-Argonne
21
ePix100 pixel linearity and gain measurements
Gain slightly lower than nominal. Compatible with parasitic capacitances around the feedback capacitor:
Measured equivalent Cf = 27.5fF
FEE 2014, Fermilab-Argonne
22
Columns
Ro
ws
10 20 30 40 50 60 70 80 90
10
20
30
40
50
60
70
80
90370
375
380
385
390
395
ePix100 gain uniformity measurement
Good Uniformity of gain across the full matrix is achieved:
Flatness across rows < 1count/fC i.e. max 0.26%
Flatness across columns < 2count/fC i.e. max 0.52%
ePix100p gain map
FEE 2014, Fermilab-Argonne
365 370 375 380 385 390 395 4000
100
200
300
400
500
600
700
800
900
1000
Gain [ADC Counts per fC]
Nu
mb
er o
f P
ixel
s
y(x) = a exp( - ((x - x0) 2̂) / (2 .̂..
a = 983.72 = 1.1148x
0 = 382.36
R = 0.99754 (lin)
Overall gain distribution and dispersion:
Gain dispersion of 1.11 count/fC is achieved i.e. ~0.3%
The result is compatible with the typical dispersion of MIM capacitors and expected gain variations in the pixel analog chain
23
Dark frame uniformity: fixed patterns
Columns
Ro
ws
10 20 30 40 50 60 70 80 90
10
20
30
40
50
60
70
80
90 1300
1400
1500
1600
1700
1800
1900
2000
Fixed pattern noise is visible across columns as expected considering the layout arrangement of the pixels in the matrix (groups of 16 with mirror image every column)
Good flatness of the dark field is achieved:
Flatness across rows < 150 counts i.e. max 0.9%
Flatness across columns < 100 counts i.e. max 0.6%
ePix100p darkframe (1000 frames averaged)
FEE 2014, Fermilab-Argonne
Columns
Row
s
10 20 30 40 50 60 70 80 90
10
20
30
40
50
60
70
80
90
2
2.5
3
3.5
4
4.5
5
ePix100p noise (darkframe subtracted)
Good noise uniformity across the full matrix is achieved:
Flatness across rows < 0.2 count i.e. max 6%
Flatness across columns < 0.1 count i.e. max 3%
24
ePix100 noise performance
a sigma of 3.2 counts r.m.s. is equivalent to about 50e- r.m.s. i.e. a S/N ratio of 44 for
single photon at 8keV
-30 -20 -10 0 10 20 300
200
400
600
800
1000
1200
1400
ADC Counts
y(x) = a exp( - ((x - x0) 2̂) / (2 .̂..
a = 1301.4 = 3.0391x
0 = 0.01
R = 0.99807 (lin)
Single pixel noise distribution
-40 -30 -20 -10 0 10 20 30 400
1
2
3
4
5
6
7x 10
6
ADC Counts
y(x) = a exp( - ((x - x0)^2) / (2 ^...
a = 6.8752e+006 = 3.2997x
0 = 0.056757
R = 0.99997 (lin)
FEE 2014, Fermilab-Argonne
All pixels cumulative noise distribution
Distribution of the pixel noise (r.m.s.) across the matrix
Noise dispersion among pixels is on the order of 2%
3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.90
460
920
1400
ADC Counts
y(x) = a exp( - ((x - x0) 2̂) / (2 .̂..
a = 14.046 = 0.070021x
0 = 3.2543
R = 0.98227 (lin)
25
ePix100: first results with a bonded sensor
ePix sensor first test (left to right): P. Hart, K. Nishimura, S. Herrmann, G. Blaj, P. Caragiulo, B. Markovic
ePix100p sensor bump bonded on a ePix100p ASIC.
FEE 2014, Fermilab-Argonne
ASIC/Sensor Single pixel Noise performance under a x-ray tube with a Molybdenum target. (Integration time 10 times longer than planned for LCLS, no temperature control ~40 °C. Leakage fluctuations present).
Noise of 5.7 counts r.ms. ~ 90e- r.m.s.
(without optimization already compatible with the application requirement of 100e- r.m.s.) Charge sharing reconstructed - G. Blaj
26
ePix100: first results with a bonded sensor at LCLS
LCLS: 9.5keV beam hitting a Cu target with a Fe one on top
ePix100: set in x2 output gainC
ou
nts
ADU
Cu
Fe
G. Carini, S. Herrmann et al.
27
ePix10k pixel linearity and gain measurements
High Gain slightly lower than nominal. Compatible with parasitic capacitances around the feedback capacitor:
Measured equivalent Cf = 30fF
FEE 2014, Fermilab-Argonne
0 100 200 300 400 500 600 700 800 900 1000
4000
5000
6000
7000
8000
9000
10000
11000
12000
13000
14000
Pulser
AD
C C
ou
nts
0 100 200 300 400 500 600 700 800 900 1000
4000
6000
8000
10000
12000
14000
Pulser
AD
C C
ou
nts
0 100 200 300 400 500 600 700 800 900 1000
-8
-6
-4
-2
0
2
4
Pulser
INL
[A
DC
co
un
ts]
0 100 200 300 400 500 600 700 800 900 10000.4
0.6
0.8
1
1.2
1.4
1.6
x 104
Pulser
AD
C C
ou
nts
Pulser in HR
Pulser in LR
Pulser in HR
Fixed High Gain
Fixed Low Gain
Fixed Low Gain
Auto-range
28
ePix10k gain uniformity measurement
Good Uniformity of gain across the full matrix is achieved in both High Gain and Low Gain
ePix10kp gain maps:
FEE 2014, Fermilab-Argonne
Overall gain distribution and dispersion:
A gain dispersion of ~0.4% in High Gain and of ~ 0.3% in Low Gain have been achieved
137.5 138 138.5 139 139.5 140 140.5 141 141.5 142 142.50
20
40
60
80
100
120
140
Gain [ADU/Pulser DAC units]
Nu
mb
er
of
Pix
els
y(x) = a exp( - ((x - x0)^2) / (2 .̂..
a = 126.19 = 0.60442x
0 = 140.13
R = 0.9926 (lin)
High Gain Low Gain
1.685 1.69 1.695 1.7 1.705 1.71 1.715 1.72 1.725 1.730
20
40
60
80
100
120
140
160
180
Gain [ADU/Pulser DAC units]
Nu
mb
er
of
Pix
els
y(x) = a exp( - ((x - x0) 2̂) / (2 .̂..
a = 155.42 = 0.0044952x
0 = 1.7054
R = 0.99326 (lin)
Columns
Row
s
10 20 30 40
5
10
15
20
25
30
35
40
45135
140
145
Columns
Row
s
10 20 30 40
5
10
15
20
25
30
35
40
451.65
1.7
29
Dark frame uniformity: fixed patterns
Fixed pattern noise is visible across columns as expected considering the layout arrangement of the pixels in the matrix (groups of 16 with mirror image every column)
Good flatness of the dark field is achieved:
Flatness across rows < 80 counts i.e. max 0.4%
Flatness across columns < 120 counts i.e. max 0.6%
ePix10kp darkframe (1000 frames averaged)
FEE 2014, Fermilab-Argonne
ePix10kp noise (darkframe subtracted)
Good noise uniformity across the full matrix is achieved:
Flatness across rows < 0.4 count i.e. max 6%
Flatness across columns < 0.2 count i.e. max 3%
Columns
Ro
ws
5 10 15 20 25 30 35 40 45
5
10
15
20
25
30
35
40
45
1800
1900
2000
2100
2200
2300
2400
2500
2600
Columns
Ro
ws
5 10 15 20 25 30 35 40 45
5
10
15
20
25
30
35
40
45
4
5
6
7
8
9
10
6 6.2 6.4 6.6 6.8 7 7.2 7.4 7.6 7.8 80
50
100
150
200
250
300
350
400
ADC Counts
y(x) = a exp( - ((x - x0) 2̂) / (2 .̂..
a = 339.55 = 0.32752x
0 = 6.9506
R = 0.99451 (lin)
-40 -30 -20 -10 0 10 20 30 400
2
4
6
8
10
12
14
16x 10
4
ADC Counts
y(x) = a exp( - ((x - x0)^2) / (2 .̂..
a = 1.5725e+005 = 6.8069x
0 = -0.011614
R = 0.99996 (lin)
30
ePix10k noise performance
a sigma of 6.7 counts r.m.s. is equivalent to about 120e- r.m.s. i.e. a S/N ratio of 18 for
single photon at 8keV
Single pixel noise distribution
FEE 2014, Fermilab-Argonne
All pixels cumulative noise distribution
Distribution of the pixel noise (r.m.s.) across the matrix
Noise dispersion among pixels is on the order of 5%
-30 -20 -10 0 10 20 300
20
40
60
80
100
120
140
160
180
200
ADC Counts
y(x) = a exp( - ((x - x0)^2) / (2 .̂..
a = 180.35 = 6.7456x
0 = 0.16766
R = 0.99377 (lin)
31
ePix10k: first results with a bonded sensor
FEE 2014, Fermilab-Argonne
AD
U
Photons (8keV)
Linearity scan low gain mode at LCLSX-ray tube with a Molybdenum target (high gain)
G. Carini, S. Herrmann et al.
Co
un
ts
ADU
CumulativeCumulative with gain correction
Summary
32
• SLAC is working on a new generation of hybrid pixel detectors for photon
science based on a modular platform approach that:- facilitates integration, scalability, versatility
- mitigates risks
- reduces development time and costs
• ePix is a class of front-end ASICs for integrating hybrid photon detectors- Based on a column parallel readout architecture with sigma delta ADCs
- Fully analog pixel matrix with filtering optimized for different areas of application• ePix-100 is optimized for ultra low noise applications. It has pixels of 50µmx50µm
size arranged in a 352x384 matrix. A resolution of 50e- r.m.s. and a signal range of
35fC (100 photons at 8keV) has been achieved. In its final version it will be able to
sustain a frame rate of 1kHz.
• ePix-10k is optimized for high dynamic range applications. It has pixels of
100µmx100µm size arranged in a 176x192 matrix. A resolution of less than 120e-
r.m.s. and a signal range of 3.5pC (10k photons at 8keV) has been achieved. In its
final version it will be able to sustain a frame rate of 1kHz.
• Tests on the ePix100 and ePix10k prototypes have demonstrated the approach
is sound. - In particular the platform has been proven to behave as required, allowing to
proceed with the submission of the full size variants ePix100a and ePix10ka.
FEE 2014, Fermilab-Argonne
ePix100p with some pixel injected using the automatic scan mode based on the internal pulser. The pulser injects in every frame an increased amount of charge in the selected pixels up to the full scale (100Ph @8keV).The interesting aspect of this movie is that all functionalities of the ASIC are used.
33
ePix10k weighting functions
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-0.2
0
0.2
0.4
0.6
0.8
1
t/Tm0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
t/Tm
AD
U
33FEE 2014, Fermilab-Argonne
Weighting functions in high gain mode
For a small signal the filter has a long time constant and the function is quasi-trapezoidal
Passed the range switching the weighting function has a reduced amplitude. The effect of gating the comparator firing can be seen in the top right part.
Weighting functions in auto-range mode
CDS effects on auto-ranging – ePix approach
5 10 15 20 25 30 35 40 45 50 55
10
15
20
25
30
Signal / Noise vs R0 to Acq distance
Sig
nal /
Noi
se
R0 to Acq [us]
@ 1 8keV Ph
@ 1 8keV Ph fitted@ 100 8keV Ph
@ 100 8keV Ph fitted
34FEE 2014, Fermilab-Argonne
In auto-ranging if a signal triggers the comparator, the CDS will subtract a baseline weighted with a larger gain, resulting in an excess noise in the second range.
Solutions: - read the 2 samples separately and
subtract only if the signal is in the first range
- don’t use CDS at all- or the ePix approach
Because the intervals of time reserved to sample the baseline and the signal are continuously adjustable in ePix; the CDS timings can be unbalanced resulting in a filtered version of the baseline sample with an effective lower gain. This is equivalent of reducing the efficiency of the CDS and increasing the S/N ratio at the switching point.
Trade off between S/N ratio at the switching point and at the minimum signal in ePix10k auto-range mode
There is an optimal point at which the S/N is larger than 14 (150e- in the first range) in both ranges
35
20 30 40 50 60 70 80 90
1000
1500
2000
2500
3000
3500
4000
Carrier Temperature (°C)
Dum
my
Pix
el B
asel
ine
(AD
U)
y(x) = a x + ba = 20.826b = 1835R = 0.9941 (lin)
y(x) = a x + ba = -0.87013b = 1632.9R = 0.44481 (lin)
TCePix (uncompensated) = 20.8 ADU/°C
TCePix (compensated) = -0.9 ADU/°C
After proper calibration a TC between +- 0.2 ADU/°C is expected
FEE 2014, Fermilab-Argonne
ePix100 temperature compensation circuitry (preliminary)