SKA Signal Processing

1
SKA Signal ProcessingAlbert-Jan
BoonstraTask force members

• Albert-Jan Boonstra (chair from July06)),
  ASTRON, The Netherlands
• John Bunton, CSIRO/ATNF, Australia
• Roger Cappallo, Haystack Observatory, USA
• - Larry D'Addario (chair till July 06),
  JPL/Caltech, USA
• - Dick Ferris, CSIRO, Australia
• Andre Gunst, ASTRON, The Netherlands
• Mike Jones, University of Oxford, United Kingdom
  
• Alan Langman, SKA Project Office, South Africa
• Mohamed Missous, University of Manchester, UK
• Sergei Pogrebenko, JIVE, The Netherlands
• Jonathan Weintroub, Harvard-Smithsonian Center
  for Astrophysics, USA
• Dan Werthimer, Berkeley Space Sciences
  Laboratory, USA

2
Contents

• Analysis of current concepts from a DSP
• point of view key technologies
• Implications for other subsystems
• Some thoughts on costing
• Acqnowledgement
• SKA Signal Processing Task Force
• esp. John Bunton and Andre Gunst

3
Key technologies
Analysis of current concepts Digital design,
demonstrators for telescopes with Aeff 104 m2

• Dish based reflectors
• ATA (USA), (meer)KAT (SA) and ASKAP
  (Aus,Nl,Can,SA)
• Cylindrical reflectors
• SKAMP (Aus)
• Dense packed phased arrays
• SKADS-EMBRACE (EU)
• Sparse arrays
• LOFAR (Nl,EU), MWA (USA,Aus,India) and LWA (USA)
• And
• installation of EVLA (USA), e-Merlin (UK),
  CABB(Aus),
• and ALMA(Int.) correlators
• SP of FPA/PAFs not (much) considered

4
Key technologies

• Correlation
• Correlators mostly use FPGAs for signal
  processing, with the exception
• of LOFAR which use a high performance computer
  BlueGene.
• When hardware cost is small relative to the
  total cost of the system general
• purpose computers or clusters of computers are
  an ideal choice
• At the next level, pre-existing FPGA hardware
  can be used
• KAT is using (FPGA) hardware developed in
  Berkeley / CASPER
• Correlators such as ATA, ASKAP, SKAMP, MWA and
  CABB are
• building their own special purpose hardware
  based on FPGAs
• eVLA and ALMA combination of FPGAs and ASICs
• Older designs that are now being installed such
  as ALMA and eVLA use
• an XF approach with first stage filters, usually
  referred to as an FXF approach.
• Escoffier et al (2007) estimate the economic
  break even point from
• an FXF (XF) approach to a pure FX approach is
  about 20 antennas

5
Key technologies

• Beamforming
• The first stage of some designs is a beamformer
• digital beamforming ASKAP, LOFAR, and LWA
• analogue beamforming MWA
• both analogue and digital. EMBRACE
• For the digital beamforming two topologies
• ring structure for limited number of beams
  (LOFAR)
• cross connection approach for many beams (ASKAP)
• Ring structure approach (LOFAR)
• inside FPGAs new antenna inputs are weighted and
  added to beam data
• all FPGAs are connected in a ring to allow
  summation of all inputs
• this is a good/efficient approach when there are
  a limited number of beams
• but as the number of beams grows the amount of
  beam data that must
• be transported between FPGAs grows as well
• Cross connection approach (ASKAP)

6
Key technologies

• Filterbanks
• All correlators now incorporate a filter or
  filterbank operation
• before the cross multiply accumulate (CMAC)
• Even the ALMA correlator, originally a pure XF
  correlator, adopted the
• approach of feeding each of the 32 cross multiply
  units from a filter
• This increased the number of frequency channels
  measured
• for a full bandwidth observation by 32.
• As well as this there is greater flexibility as
  each of the cross multiply
• units can be operated at a different frequency
  and a different
• frequency resolution
• The eVLA correlator uses the same approach
• In both cases the output of the filters are
  inputs to an XF correlator unit

7
Key technologies

• Filterbanks
• Filterbanks
• in the LOFAR, SKAMP, and MWA correlator
  development this is taken to the extreme
• cross correlation compute load is independent of
  frequency resolution
• gt these correlators process the full bandwidth
  at a high frequency resolution.
• SKAMP
• initially in SKAMP a cascade of a 128 band and
  256 band filter bank is used
• to form 32,000 frequency channels
• higher resolution is limited by storage, but the
  basic hardware is capable of full-band
• operation at about half a million frequency
  channels (200 Hz resolution)
• however at this time there is no capability to
  process such a large data set.
• LOFAR
• in LOFAR a cascade of a 512 band and 256 band
  filter bank is used.
• a subset (maximal 216) of the 512 subbands are
  fed into the correlator
• maximal 55,296 frequency channels with a
  resolution of 610 Hz.
• ALMA
• half-million frequency channels is implemented
  in the ALMA compact array correlator

8
Key technologies

• Power consumption
• If half the eVLA compute load is in the cross
  multiplies then the
• power dissipation for the SKA correlator
  (ignoring any technology
• changes) would be 850 MW.
• For the SKAMP correlator the power dissipation
  is estimated
• at 5 kW (for 32 antennas) and it has half the
  cross multiply
• compute load of the eVLA.
• This suggests that a pure FX design would
  consume up to 50MW,
• using current technology.
• Paris SKA2006 presentation AJB/AG LOFAR-like
  approach
• (polyphase filterbank,subband beamforming,
  filtering, correlation)
• 100 MW(AA) 300 MW (PAF)
• Filterbank dissipation scales linearly with the
  number of antennas so scaling the eVLA numbers
  gives about 8MW.

9
Key technologies

• Packet switch interconnections
• signal re-ordering prior to correlation
• In An SKA Engineering Overview SKA signal
  processing chapter
• Interconnections lack of appropriate packet
  switches
• three companies that make 512 port IB switches
• (20 Gbit/sec per port, full duplex, 20
  Tbit/sec aggregate bandwidth).
• SUN Microsystems makes a 3420 port IB2 switch
• (140 Tbit/sec aggregate bandwidth)
• to build a 3420 station correlator, you'd need
  one of these
• SUN Microsystems switches for each 2 GHz of
  bandwidth.
• (assuming 4 bit real, 4 bit imag are transmitted
  from F to X)
• Source Dan Werthimer

10
Implications for other subsystems
Data transport number of beams Nb without station
beamforming
the variable ?a is approximately the same for all
concepts, ? for a given FoV and Aeff the total
number of signals to the correlator is
independent of the concept and the signal
processing
in case of station beamforming it is economic to
transport station beams if there are fewer beams
(2KM) than the number of beams from the antennas
in the stations (2KL). otherwise its more
efficient to transport the antenna beam(s) and
beamform at the correlator
11
Implications for other subsystems
Data transport the location of final
digitisation for the correlator data can also
affect the data rate. if all beamforming and
frequency channelisation is performed at the
antenna (station) then the data can be quantised
to its final resolution. this is 4 bits in
many recent correlators but 2 bit quantisation is
still common SKA monitoring campaign 3 to 7
effective bits are needed strong RFI station
nulling may be needed to reduce number of bits
(cf. LOFAR 4 bit mode) thus the resolution into
the correlator may be much less than that of the
analogue to digital converters, reducing the
data rate. if antenna data is transported to
the correlator, and beamforming is done there, it
may be necessary to transport the data at higher
resolution than that used at the input to the
correlator further simulation and analysis is
needed to determine if this effect will increase
the data rate significantly
12
Implications for other subsystems
SKA SSSM monitoring
Mode 1, 70 MHz 22 GHz

• Number of ADC bits
• - 3 to 7 effective bits
• depending on f, BW
• allow nonlinearities for short timescales?

13
Implications for other subsystems
Occupancy LOFAR HBA, 8-4-2007 two worst case
examples
loss _at_ 4 bits 30
loss _at_ 4 bits 60-80
?rfi ltlt ?sys 10 log10(Nch) dB ?rfi ltlt
?sys 24 dB for Nch 256, and assuming RFI is
located in one f-bin ony
14
Implications for other subsystems

• Computing
• FX correlator architecture easily generates a
  million frequency channels
• ? this does not change the cross multiply
  accumulate compute load, only storage.
• however this number of channels greatly exceeds
  the expected computing capacity of the SKA
• this effect is already being seen on correlators
  such as SKAMP and ALMA compact array
• it will be necessary to reduce the number of
  frequency channels by
• discarding channels
• reducing frequency resolution by averaging
  adjacent channels
• increasing integration time.
• for the last option the integration time might be
  variable with longer integrations for
• shorter baselines.

15
Costing models

• Operational aspects not yet included
• Especially power consumption
• 100 MW(AA) 300 MW (PAF) DSP 2006 (Paris SKA2006
  presentation), 8 Y Moores law 13-38 MW, using
  ASICS another factor 10 down?
• System design (memo 93 SKADS) _at_ station all
  signal paths transported to central (bunker)
  before digitization and before beamforning
  alternative strategy may reduce power consumption
• Ensure same approach for each of the concepts
  considered
• Best is to put it explicitly is a model
• Reliability
• E.g. all signal paths of 512 elements (9m2), all
  signal processing incl. ADC, on 1 PCB (8Tb/s).
  Very complex board
• Complexity may lead to a need of different designs

16
Costing models

• Extrapolation of cost
• (memo92 SKAcost, memo93 SKADS)
• Draft spec for the SKA Table 2 presents three
  possible combinations
• of technologies for Phase 2 implementation that
  are projected to fall
• below the cost target
• question what are the uncertainties, 20, 100
  1000?
• e.g. reduction of upper frequency 1.2/1.0 GHz to
  800 MHz for AA is unsufficiently justified - cost
  estimate of EMBRACE suggests AA fall below cost
  targets
• wonder whether more complicated relations are in
  the costing models
• e.g. lower signal processing efficieny of
  FPA/PAFs vs AA (in FPAs only part of the SP is
  used for lower frequencies)
• ? if frequency ranges for technologies are
  determined binary from (uncertain) cost
  estimates, then maybe introduce error bars on
  frequency ranges for technologies

17
Costing models

• Tracability, avoid black box
• Not just open source code, also
• clear, complets and self-contained
  descriptions
• of underlying formulas, assumptions etc.
• Connecting models vs integrating them
• Now several models connected together?
• Better one parametrized model in which
  different concepts are defined
• by different parameter values (as in SP white
  paper)
• - this also ensures a unique specification of
  components
• e.g. SKADS memo83 4xADC cost 10, 8Tb/s
  1 PCB 250 Euro,
• 150Euro 10TMAC/s (faster than Moores
  law) - different assumptions in SKAcost
• - System is then more suited to investigate
  different SP scenarios

C 2(L NL N1)I cd cb(I,K) 2K NL cb(L,M)
2 K M N (cD cFB(b) 2c?) 2 K
M ci(N,b) N 2cx cb(N,Q) cpd
18
more details in the TF white paper