Telescope Concepts for Very Large Arrays. John D Bunton Workshop on Novel Telescopes for 21cm Cosmology June 2011

Transcription

1 Telescope Concepts for Very Large Arrays John D Bunton Workshop on Novel Telescopes for 21cm Cosmology June 2011

2 Early Radio Astronomy L Parkes 64m 1961 Power meter Sea Cliff Interferometer 1948 Power meter Mills Cross 1954 Phased array Single cross correlator Molonglo Radio Observatory Cylindrical reflector 11 beams and correlations

3 Transition to parabolic dish antenna Sea Cliff 4 Yagis, 6 sources Mills Cross, 1000 dipoles, sources Molonglo 408 MHz, 7000 dipoles, sources 186 pulsars But 1980 to now is the era of the dish with single feeds Parkes and Jodrell ~1960, Westerbork 1973, VLA 1983 Why? - Cryogenically cooled LNA Not possible on phased arrays or cylinders Eg Parkes new receivers yearly, continual decrease in Tsys, increase in BW, also increasing frequency coverage While Molonglo (similar cost and time) managed a single upgrade. The more complex multi-feed telescope could no keep up (VLA initially cryo cooled and such a large collecting area that it prospered without significant upgrades till now)

4 The new millennium Existing parabolic dish antenna are starting to run out of steam Tsys no longer dominated by LNA and bandwidths approaching octave. Era of improving performance by installing a new dual polarisation feed is rapidly fading. Astronomy community see this coming One response is SKA spend money to get collecting area (super VLA at high frequencies). Second response increase field of view

5 Increasing Field of View First success - Parkes Multibeam 13 beams increased speed or equivalently for surveys increased sensitivity by 3.6 Breaths new life into an old telescope keeping very active in terms of publication/science Resurgence of low frequency astronomy Aperture Arrays (children of the Mills Cross and similar early phased array instrument) LOFAR, MWA, PAPER Cylindrical reflectors BAO, CHIME proposed Decreasing cost of electronics is increasing maximum frequency Mills Cross 85MHz to LOFAR 250MHz, MWA 300MHz Fully electronic now - no manual tuning

6 Aperture Array Field of view LOFAR 2.5 o at 150MHz single high band station beam 32 MHz Tradeoff between beams and bandwidth 16 2MHz gives 10 o MWA 25 o at 150MHz bandwidth 32 MHz PAPER full sky SKA phase 1 AA low proposals approximately LOFAR but 100s of beams Field of view ~25 o across For all designs full sky with ~100 pointings or less (cf 10,000 to 100,000 for dish at 1.4GHz)

7 Field of View comes to the SKA (high) Cylindrical Reflectors proposed for SKA Field of view of elements ~50 sq deg at 1.4GHz How can this be utilised BW beam tradeoff proposed at 2003 SKA meeting Dish BW at max frequency ~10GHz Cylinder at 1.4 GHz can have ~20 beams at 500MHz for the same correlator Order of magnitude more field of view. Problem Paralactic Angle changes with HA + and -2 Hours FOV= 6-10 times similar width dish

8 Cylinder to PAF Cylindrical Reflector did not get traction in SKA community. But wide field of view did Phased array feeds on ~10m dishes proposed Concept was needed for the Large Adaptive Reflector Phased array feeds allow up to 50 times the survey speed (centre feed blockage limit) compared to a single feed on a dish. In practice PAF and beamformer should not cost more than the dish (optimum survey speed at this point) What is possible is limited by technology. Current practical designs (ASKAP and EMBRACE) survey speed times more than single dish.

9 Receptor Electronics tradeoff Number of signal paths (N) per m 2 varies with RF beamforming on Aperture Array or Cylindrical Reflector Width of Cylindrical Reflector Ratio of PAF area to dish area. Assume field of view proportional to N/m 2 Total cost A with C spent on electronics Total Area = (A-C), Field of View =C/(A-C) Survey speed = (A-C) 2 * C/(A-C) = (A-C)C Maximum with C=A/2 Normally C<A/2

10 Cost of a signal path including beamforming ASKAP ~$3000 per signal path from LNA to Beamforming (2008 technology) Including, RF on copper to pedestal, dual heterodyne receiver, ADC, filterbank, beamformer In 2012 technology RF over fibre (eliminate coaxial cable Direct conversion (eliminate hetrodyne RX) Two generations of FPGA improvement (1/4 cost) Low cost optical SNAP12, Avargo MiniPOD Estimate ~$1000 per signal path including Beamformer (but excluding correlator) Cost not very sensitive to BW. ~500MHz

11 Cost of metal Aim for SKA is ~$1700/m 2 for a dish Cost of steerable Cylindrical reflector estimated from Molonglo cost ~$130/m 2 for 12m wide reflector from SKA Cylindrical Reflector Proposal No Slewing ring Elevation from two widely space supports Adjacent bays support each other Surface curved in one dimension easy to make Add a significant margin for error ~$400/m 2 Cylindrical reflector less than 1/4 cost of dish but mesh surface Fmax 2-3GHz Fixed Cylinder much less again

12 Comparison dish/paf and Cylinder Assume electronics=metal Dish $2000 per dual pol chain/($1700 per m 2 ) ~1.4m 2 /signal chain 100 dual pol element PAF 140m 2 ~13m reflector Parabolic Reflector $2000/($400 per m 2 ) ~5m 2 /signal chain 100 dual pol feed = 500m GHz one dual pol feed every 0.12m 5m 2 /0.12m = 40m wide reflector

13 Comparison continued Field of view Dish ~20 square degrees Area ~140m 2 Survey speed 6x10 5 /T sys 2 m 4 deg 2 K -2 Cylindrical Reflector (signal chain per feed element) Field of view 120 deg x 0.25 deg 2 = 30 sq deg Area 20mx12m =500m 2 Survey speed 7.5x10 6 /T sys 2 m 4 deg 2 K -2 Higher survey speed with Cylindrical Reflector But only as transit instrument for full field of view Also ignore loss of sensitivity at 60 deg off centre.

14 Cylindrical Reflector with RF beamforming If line feed RF beamformed so the beam is less 20 times that of the optical beam (~ λ/d radians) (use to -1dB) Then can observer many hours. RF beam is ~1/N radians if beamformed over a distance of Nλ/2 m 1/N <20λ/D N > D/20λ = Let D=8m N>2 Length of line feed for an area of 5m 2 is 0.625m Beamform over 5 elements (number of elements OK) For 100 dual pol element area is 8x62.5m = 500m 2 Field of view is 6 O x 1.5 O = 9 deg 2 Survey speed 2.2x10 6 /T sys 2 m 4 deg 2 K -2 Lost survey speed as RF beam used to -1dB only Can decrease width and increase no elements beamformed

15 Effect of Cost Changes If cost of area halved Area per signal chain doubled But Field of View per signal chain halved Survey Speed Doubles per unit cost If cost of signal chain halves Two signal chains for same cost Area per chain is halved (optimum Survey Speed) For same cost same area Field of View per chain is doubled Survey Speed Doubles per unit cost

16 Aperture Arrays. Comment on Cylinder regarding Signal Chains per square metre and Survey Speed to arrays of dipoles. Dipoles are cheap but need many to maximise survey speed. Or very cheap signal processing chains. PAPER has single processing chain per dipole optimised for science not survey speed. MWA RF beamforming over 16 dipoles Adds cost of RF beamformer Better survey speed optimisation

17 Signal processing challenge For small system can bring all data into single FPGA/CPU/GPU If more processing needed distribute from here to multiple nodes But as systems grow data distribution becomes the major problem ASKAP has 6912 signal processing chains MWA 1024 VLA has 128 each 2GHz wide All require data from every input to arrive at the same correlator module (FPGA)

18 Forming the image Consider 20 cylinders 5mx100m total area 100mx100m (single pol for simplicity) RF beamforming over 1m = 100 signals per cylinder, 2000 total signal form telescope Full correlator = 2M correlation Ouch - but Standard packages available to reduce data Standard techniques for to determine antenna based errors. High quality imaging possible High quality imaging need for future astronomy

19 Fourier Telescope Optical telescopes have superb imaging Image is Fourier transform of aperture Very large number of aperture elements ~(λ/d) 2 or ~10 8 in a 10cm reflector but each with large errors ~20 o Form all beams using Fourier Transform across corresponding time samples (Fourier Telescope) For our example ~20,000 complex multiplication -good First used at Molonglo using a Butler Matrix which is the analogue equivalent of a Fourier Transform. Good imaging for the time with ~100 elements But now well behind what is achieved by VLA, Westerbork, AT Not enough elements - much larger number needed, unless there is a method to calibrate elements better. Full correlator installed at Molonglo, immediately identified, mistracking, phase jumps etc Fourier telescope equivalent to full beamforming within RF beam. Only beams within -1dB contour of RF beam usable (~1/4 RF field of view)

20 Omiscope If technique is based on correlations F Full field of view becomes available Ominscope arranges antenna elements so that there is a large degree of redundancy in uv plane and calculates only the summed redundant values. 201x41 = 8000 multiplies (single pol) to form correlations Also need Fourier Transform of cylinder voltages So for our example similar computations to Fourier Telescope Each output correlation is formed from subsets of elements Should be easier to untangle errors but techniques need to be developed. Fourier telescope directly gives an image, Omniscope needs processing of correlations. But this extra processing may allow calibration of elements

21 Fast Synthesis Array Omniscope is like an autocorrelation. All data Fourier Transformed and multiplied by itself. Break the array up into identical OmniScope subarrays Can now multiply Fourier data from one array with Fourier data from another Effectively a cross correlation version of Omniscope Fast Synthesis Array IEEE Antennas and Propagation 2011 Antennas arrays need not be uniformly spaced allowing better uv coverage Increase in number of correlation 201x200 = 40k correlations (single pol) for example telescope Five times Omniscope

22 Calibration Fourier Telescope calibration of elements using output data may not be possible With Omniscope 4N values for N signal chain. Calibration possible but not a lot of degrees of freedom Correlation error is sum of the product of complex error pairs Information tangled together but may be separable Fast Synthesis Array, ~20N values Also Correlations of opposite elements are ordinary correlations. These should be solvable with ordinary selfcal Others are tangled to some degree Suggest Full or Partial correlation along Cylinder For full correlation 20 x100 2 /2=100k correlations - similar to Fast Synthesis Array correlation (decreases with number of cylinders) Now can easily calibrate each cylinder Single complex gain error per cylinders Use opposite end elements or develop techniques using tangled correlations.

23 HARDWARE

24 Example of current generation system ASKAP beamformer MHz bandwidth Array covariance matrix (sub sample time and frequency) 36 dual polarisation beams frequency channels CROSS CONNECT

25 DragonFly system Digitiser and filterbank system installed in Pedestal DragonFly-2 2 Dual channel 8-bit ADC Four coaxial inputs + 768MHz clock Virtex 6 LX130T FPGA for coarse filterbanks Four SPF+ optical links each transporting 76MHz bandwidth for all four ports Command/Control and Power system moved to separate boards Full system 48 Dragonfly-2 4 Control 4 Power In 4 3U racks 18Tops/antenna

26 Redback Beamforming Base on industry standard AdvancedTCA shelf with fully cross connected back plane 12 10G inputs to RTM To Crosspoint Switch to backplane Data from single Dragonfly2 distributed to 16 Redback2 19MHz each 4 LX240T processing FPGAs per board. 8 x10g and dual 1GE ports for output 27 Tops/antenna Same hardware for correlator RTM AdvancedTCA shelf with full cross connect backplane Redback-2

27 ASKAP Beamformer hardware (1 of 36) 48 DragonFly-2 16 Redback processing FPGAs

28 The digital challenge Tasks Decide topology Get sufficient compute power (Flops, MIPS) Get the data where it is needed Can we afford operational cost If correlator based choose FX. For large numbers of signal chains Filterbank cost is small and FX minimises computation Cross connect so correlator module process part of bandwidth For Fourier Telescope 2D FFT process ~ N signal chains in a module From 10,000s to 100s. Time domain processing may be possible. Still need cross connect between row and column computation First task can we get the data where we need

29 AdvancedTCA cross connection AdvancedTCA Telecommunication card cage and backplane Backplane near limit of circuit board fabrication ~$10k for backplane and cardcage 14 and 16 slot Four high speed connection from each slot to all others on backplane 16 slots to 15 other slots x 4 connections = 960 Standard 2.5GB/s per connection = 2.4Tb/s $4 per Gb/s Standard inputs are 10Gb/s (SFP+, Infiniband) One card cage can handle 240, possibly 500 inputs Not enough for our 20 cylinder telescope

30 Pure optical cross connect 12 Fibre optical links now available 10Gb/s per fibre Next Gen FPGAs can handle ~4 of these Two for input and output Two for cross connections between FPAGs Can connect to 24 other FPGAs 24 FPGAs x 24 Fibres = 576 fibres = 5.76Tb/s Pure optical cross connect exceeds electrical backplane connections

31 Hybrid Cross Connect Optical between boards Electrical on boards Two hops to get any where 1/3 per FPGA connections for I/O 1/3 per FPGA connections for optical cross connect 1/3 per FPGA connections for electrical Say 48x10Gb/s per FPGA, N FPGAs Input to a board ~Nx16 fibres Connections to ~Nx16 boards ~256N 2 fibres into system 8 FPGAs per board 16k Fibre inputs, 160Tb/s

32 Data per Fibre Assume 8 bit data 600MHz bandwidth on each 10Gb/s fibre Put data from multiple ADCs on one fibre Bandwidth less than 600MHz multiple ADC fibres MWA 32 ADCs per fibre For wide bandwidth use filterbank and put part of bandwidth on a fibre ASKAP puts ¼ of BW per fibre, 4 ADC per fibre If we are using 12 Fibre ribbons Data for 12 or more ADC per fibre Cross connect between fibres mechanically Systems with more 100,000 ADCs possible with current technology

33 Next Gen FPGAs 2011/12 generation of FPGAs Stratix V, Virtex multipliers 2TMACs 4T arithmetic ops per second Gb/s SERDES 1Tb/s per FPGA But high end parts are very expensive Choose midrange devices Cost per FPGA ~ $1000 ~2000 Multipliers ~30 SERDES (300Gb/s) Why FPGAs Lower power At $1-4/w/year with CPU/GPU you will spend more on the power than the capital cost

34 FPGA Technology Roadmap Moores law starting to slow down 2 years moving to 3 years per technology step 5 years for next two steps 2016/17 Multipliers x4 SERDES speed 25Gb/s per link More links per FPGA ~ Estimated midsized FPGA 8000 Multipliers 8Top/s 100 x 25Gb/s 2.5Tb/s Power ~50W Cost still ~$1000

35 Optical connections (short haul) 12 fibre multimode short haul Expect devices to match FPGA SERDES 10Gb/s to 25Gb/s - 300Gb/s per ribbon cable Cost now less than electrical backplane technology Electrical backplanes replaced by optical Large increase in production volume Decrease cost to ~$200 for TX and RX Less than $1 per Gb/s (cf $4 per Gb/s in ATCA) Optical backplanes more flexible than electrical

36 Optical connections (Long Haul) Sweet spot Gb/s - SFP+ Reach ~$200 each Sweet spot Gb/s Sweet spot Gb/s Four wavelengths at 25Gb/s per wavelength Cost per unit constant with improved performance

37 Optical Connection (Analogue) Until last year putting RF signals onto optical links required external modulators to get required performance Must use single mode fibre Too many modes in multimode fibre Mode conversion gain and phase variations Directly Modulated DFB laser with suitable linearity available this year ~$100 per DFB laser Work on VCSEL laser much cheaper again 2GHz+ of bandwidth at better than 8bit converter dynamic range Equivalent data rate over

38 ADC Up to 5 GS/s (eg E2V) Power and cost proportional to clock rate ~1W per GS/s (8bit converter) Clock rate reduced if operated in second Nyquist zone 1.6GS/s and cheap antialiasing filter GHz bandwidth (less than an octave) Second band 1GS/s GHz 3.2GS/s for in one hit. ADC improvement slow Maybe factor of 2 in 2016!/2W per GS/s 4 channels at 2.5 GS/s Hope for better than this 100Gb Ethernet need very high speed ADC Novel design Time Domain ADC (Haslett et al, U Calgary)

39 Using the ADC Direct sampled Clock rate > 2Fs Everything in one go Direct sampled 2 nd and 3 rd Nyquist zone ½ to 1/3 clock rate Octave bandwidth Large saving on ADC but multiple filters IQ demod ADC clock rates match BW not Fmax IQ balance? Missing band at demod frequency

40 ASKAP beamformer 2016/17 RF over Fibre replaces coax (10km reach) ~$100 per link 4 channels high performance ADC 40 inputs per ADC module 5 modules (down from 48) 12 x 8000 multiplier FPGA beamformer- 2 boards 200 x 500MHz processing system $500 per signal path

41 Opportunities SKA 1 will fill a many of niches AA-low will be hard to compete with below 450MHz Above 450MHz dishes have a fill factor of at most ~0.1 Only 1/10 or less of the area enclosed by an array of dishes is collecting photons. This is a problems for surveys using tied array beams such as pulsar surveys and transient detection Pulsar survey want to form a beam using all antennas then analyse the resulting signal Poor fill factor = small beam area and slow survey. Let us think of a filled aperture telescope. Let us say 250,000m 2, 450-1GHz

42 The Receptor At low frequencies the dipole is king LOFAR, MWA, LWA, PAPER Number needed for squ metre proportional to frequency 2 Rapdily numbers become unmageable Dish constant area per receptor Expensive, Surface supported from one central point Mount is two axis one above the other Roughly 1/3 surface, 1/3 mount 1/3 pedestal/ foundation

43 Cylindrical reflector Use reflector to concentrate in one direction Use Array of dipoles in other Hybrid of dish and single dipoles Number of dipoles proportional to frequency (not f 2 ) Proportionality can be changed by changing width of surface. If steerable surface support at two points, adjacent bays share support Mount is simple Plummer blocks no azimuth bearaing Pedestal takes torque in one axis and can be trangulated Cost of equivalent mount and Pedestal low Surface cheaper as it is curved in one direction If non-steerable cost even lower

44 The Transient telescope 250,000m 2 = 500m x500m Cheap collecting area Fixed cylinders on ground say ($100/m 2????) $25M for reflector Fixed means transit instrument NS Can t be too wide otherwise observation time too small.- say 10m (1.5 deg@1ghz 6+min dwell time) 25km of line feed $25m on electronics = 50,000 signal paths One per metre dual polarisation Can see 8 degree swathe in DEC 1db ~$50M capital cost

45 Beams, beams, beams ~250 beams in DEC for each cylinder These are fixed 500 dual pol inputs 250 dual pol beams out FFT to form these beams Gain variation between beams less than 1dB For each corresponding beam Beamform across cylinders 50 beams tracking in RA, requires delay tracking Continuous RA coverage Total of 12,500 beams Pulsar search a say 5 DEC beams at once 250 search engines, 48 deg in DEC per year Transients limited by dedispersion resources HI in the early universe possibly confusion limit

46 Is it Easy?

47 ASKAP John Bunton Project engineer Phone: John.bunton@csiro.au Web: Thank you Contact Us Phone: or enquiries@csiro.au Web: