Correlation Receivers for Detecting
Cosmological Signals
by
Girish B. S.
Srivani K. S., Saurabh Singh, Mayuri S.
Udaya Shankar N., Ravi Subrahmanyan
Raman Research Institute
Bengaluru
A National Workshop
Cosmology With the 21-cm HI Line
June 23 – 26, 2015
Introduction
 RRI is working towards the goal of building precision receivers for
detecting signals from the Epoch of Recombination & the Epoch of
Reionization
 Array of Precision Spectrometers for the Epoch of Recombination APSERa - a project to detect recombination lines from the Epoch of
Cosmological Recombination.
 Detecting Epoch of Recombination signatures in the CMB spectrum is
extremely challenging, since the estimated magnitude of fluctuations is 8
to 9 orders of magnitude weaker than the CMB radiation temperature
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
About APSERa
 Feasibility study has shown that optimal frequency range for operating
APSERa to be around 2 - 6 GHz
 APSERa is likely to consist of an array of 128+ purpose-built small
telescopes, operating in the above frequency range
 Rapid advances in technology allows development of broadband
precision receivers to attempt such challenging tasks
 Currently, a prototype system consisting of antennas, precision analog
front-end & digital receivers, is being designed & developed at RRI.
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
In this talk….
This talk highlights the design challenges in the
development of a broadband correlation spectrometer
based on pSPEC (RRI’s high speed signal processing
platform) and preliminary results obtained from
pSPEC prototype
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
Correlation Spectrometer - Requirements
 Capability to digitize and process broadband signals having a
bandwidth of about 2 to 3 GHz
 Channelize the sampled signal into narrow sub-bands (8192) by
implementing efficient polyphase filter banks which provides a
handle on shape of pass band and adjacent channel rejection 
A high dynamic-range, RFI-tolerant receiver
 Channel-wise cross-correlation to obtain cross-power spectra of the
sampled signal
A common DSP platform that can cater to requirements of both EoR
and Recombination experiments is being developed
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
Time-interleaved ADCs to enhance sampling rate
 High bit precision, 5 to 6 GSps commercial ADCs not easily available.
Large BW requirements are met by time interleaving multiple, highspeed ADC modules
 Sampling rate increased by a factor of M, by
time interleaving M ADCs with sampling CLKs
phase offset by
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
Challenges in time interleaving ADCs
Offset mismatch
Gain mismatch
Phase mismatch
 Sampling Fin= 1 GHz (1000 ps), a 10 ps (3.6°) skew
in sampling clock between TI-ADCs degrades
SNR to ~25 dB
 Offset mismatch is independent of Fin
and its amplitude
 Gain mismatch: Largest error in channel
outputs occurs at peak amplitude
 Phase mismatch: Largest error in
channel outputs occur during largest
slew rate
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
Gain and Phase mismatch errors
 If two 10-bit ADCs are time-interleaved, to obtain a system
performance of ˜62 dBc, gain matching and phase matching (clock
skew) between time-interleaved ADCs should be better than 0.1%
and 0.04 degrees (~120 femtoseconds at 1000 MHz)
 Offset, Gain and Phase matching between channels/cores of
modern ADCs can be realized by:
• Using ADC’s on-chip compensation features
• Matching flight-times of signals routed between ADCs and
FPGA to within timing-budget
• Finer path delay adjustment features inside the FPGA
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
Sampling Clock Jitter
Flatness of pass band
 Each sample of autocorrelation function
is perturbed using a normally distributed
jitter
 Nulls in the autocorrelation function do not average to zero due to
asymmetric nature of the autocorrelation function around each null
 Resulting bias due to averaging, alternates between positive and
negative values for odd and even nulls
 Sampling clock jitter (rms) of about 100 fs
 Results in a peak-peak pass band ripple of about 0.0004 dB
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
Overview of RRI’s high-speed DSP platform
Digital Signal Processing platform (pSPEC) built around multi-GSps
ADCs and Virtex 6 FPGA
 Processing of 4 analog signals at Fs = 2 GSps (2 ADCs in time- interleaved mode)
MWA Technical Meeting, The University of Melbourne, June 9 – 10, 2013
Choice of ADC for pSPEC
 Quad, 10 –bit ADC
Block Diagram of EV10AQ190CTPY
 Analog BW = 3 GHz
 Fclk (max) = 1.25 GHz
 LVDS output format
A
B
C
D
 DDR output protocol
 500 mV p-p analog input (-2 dBm)
 +3.3 V, +1.8 V supplies
 1.4 W per channel
 Modes of operation:
a. 4 independent channel mode (BW = 625 MHz, each)
b. 2, 2-channel time-interleaved mode (BW = 1.25 GHz)
c. 1-channel time-interleaved mode (BW = 2.5 GHz)
 In-built analog cross-point switch (multiplexed)
 SPI interface for Gain matching (steps of 0.02%, ±10 %), Offset matching (steps
of 40 µV, ±40 LSBs), Phase matching (steps of ~30 femtoseconds, ±15 ps)
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
pSPEC: Evaluation of ADC module
Effective Number Of Bits of 7 bits and Signal-to-Noise Ratio (SNR) of 44dB for an
input tone at 1GHz (-3dBm), compares favourably with datasheet specification.
Advantageous to use Quadrature Sampling
Instead of time interleaving multiple ADC cores to enhance sampling
rate, Quadrature Sampling is used
 In Quadrature Sampling, Fs at
each ADC needs to be ≥ signal
bandwidth, as opposed to twice
the bandwidth in real sampling
 When we need to digitize wide bandwidths, Quadrature Sampling is
very attractive, especially, when the aim is to used FFT-like processing
for spectral analysis.
 FFT engine can be used efficiently by feeding complex numbers as input,
rather than real numbers.
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
Channelization – Divide & Conquer approach
 De-multiplexed ADC data:1 GSps/4
@ 250 MSps feed 8 smaller pipelined FFTs
 Modern,
high-performance
FPGAs
allow parallelization of FFT algorithm
(M=8, N=1024)
 A combination of optimized, pipelined
FFT core and custom-designed parallel
FFT.
 Implementation of large BW, high-
resolution spectrometer possible.
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
Preliminary results from pSPEC
Tone i/p (Fin)
Fs/2
Fs/2 - Fin





Input to pSPEC: Tone at 78.125 MHz, -22 dBm
Two ADCs each in TI mode-of-operation -I, Q(set to 0)
Effective Sampling rate = 1 GSps x2 = 2GSps
FFT length = 8192
On-chip averaging for 16.77ms
For a tone i/p at 78.125
MHz (-22 dBm), achieved
SNR value matches with
the
ADC
datasheet
specifications.
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
Preliminary results from pSPEC
 Broadband noise input
(LPF, with fc = 450 MHz)
 Two ADCs each in TI modeof-operation -I, Q(set to 0)
 Eff. Sampling rate = 2 GSps
 FFT length = 8192
 On-chip averaging for
16.77ms
 Self-power spectra and
cross-power spectrum
For broadband noise i/p, we get
the expected band shape and
SNR
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
SARAS (for EoR experiment)
Shaped Antenna measurement of the
background RAdio Spectrum (40 – 200 MHz)
FREQUENCY -Hz
FREQUENCY -Hz
 Input: Tone at 78.125 MHz, -26 dBm
 Fs = 500 MSps, FFT length = 8192
(weighted by Blackman-Nuttal Window)
 On-chip averaging for ~67ms
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
SARAS
Picture of integrated EoR spectrometer
A National Workshop – Cosmology With The 21-cm H1 Line, June 23 – 26, 2015
S
Thank you