Kristian Zarb Adami

Pathfinders for the SKA:Nlog(N) vs. N2 Imaging

Instruments:

N log N Astronomy

In fact... Japan designed the SKA in 1994

8x8 Images in 1994 with Waseda telescope

Extrapolating with Moore’s Law (doubling every 18 months)

2016 is 1x106 antennas

Which is equivalent to SKA-phase-1

Remit of the talk Science Justification for SKA1-Low

Science and Technical simulations towards implementation of the SKA

Physical Implementation on Medicina as a flexible DSP test-bed and a comparison between spatial-FFT and N2 imaging

Industrial Engagement

SKA Phase-1 Specifications

Memo 125

Sensitive (-ity) Issues..

[SKA Memo 100]

Roadmap to the SKA-loN

BW

(16,32)

(16,60)

(1,768)

(25,192)

(400,50x2xNbeams)Super Terp

LOFAR UK

GMRT

Medicina

SKA-1

(32, 64)

MWA-32

LOFAR

(8,50x2xNbeams)

LWA

(78,100)

(32, 1024)

MWA-512

PAPER

(100,128)

MITEOR

(25,16)

H1-Power Spectrum (z≈8)

Theoretical 21-cm Power Spectrum @ 150 MHz

Power Spectrum from a (100,256) instrument

Foregrounds suppressed by frequency/angledifferencing

NlogN vs. N2

LOFAR2010

Super-Terp2011

SKA-Phase 1SKA-Phase 2

HI Power Spectra (SKA-Phase-II)

Blue: HI > 108

Green: HI > 20’

Linear Bias = 1.0Linear bias = 0.8

Co-moving Volume = (500MPc/h)3

SKA1 Low Layout

100km

200m

Bandwidth 70 – 450 MHz (Instantaneous B/W 380 MHz)

ADC Sampling at 1 GSa/s @ 8-bit

Antenna Spacing ~ 2.6m

Array Configuration:

50 stations

11,200 antennas per station (~10,000)

Output beams of 2-bit real; 2-bit imag

The numbers game (SKA1-low)

Numbers cont... SKA-1 ~ 50 stations of 10,000 antennas each Station diameter ≈ 200m Station beam @ 70 MHz ≈ 1○, @ 450 MHz ≈ 0.2○

Nbaselines = 5,000 (50^2/2 *4)

Input data rate to station 160 Tb/s (total data rate 8 Pb/s for the SKA-1 lo) Output rate?

Assume 10 Tb/s off station = 100 x 100Gb/s fibres

Output beams 2+2 bits, ~100kHz channels (1.6Mbps per beam-channel) 6.25 million beam-channels – by DFT need 0.1 Pop/s (6250 beams @ 1000 channels)

Equalise sky coverage so N(f) ~f2 – 100 beams in lowest (70 – 70.1 MHz) channel 100 sq deg instantaneous coverage.

Correlator has to do 1,000 baselines for each 1 kHz beam-channel (for a total ~ 10 Pop/s)

Station Architecture

Station Layout

Richard Armstrong – richard.armstrong@astro.ox.ac.uk

TileProcessorTile

Processor

TileProcessor Tile

Processor

StationProcessor

Optical Fibre

Optical Fibre

Copper

Hierarchical Architecture

Antennas

Multiply and add by weights

Multiply and add by weights

Cross correlation of sub-arrays (for station calibration and ionospheric calibration)

Hierarchical Beam Forming (tiles then station)

Tile Level Weights

Station Level Weights

Direct Station Beam Forming

Station Weights

Sub-Station Cross-correlation (calibration)?

Sub-Station Weights

Tile level

Electronic Calibration

Field or Strong Source Calibration

~CAS-A

Source & Polarisation Calibration

Polarisation Calibration

Tile processor box

RF in (coax) 16 x dual pol

Multi-chip module

Fibre:Data outClock and control in

RegDC in

Tile Processor

ADCADCADCADC

Coarse freqsplitting

1st LevelBeamforming

RFIMitigation

&4-bit

Quantisation

Tile Processor

Inputs: 16 dual-pol antennasADC @ 1GSA/s @ 8-bit

Coarse frequency splittingInto 4 channels

Outputs: dual-pol beams@ 1GSA/s @ 4-bit re/4-bit imag

Output is optical

Control and Calibration Interface

Space-Frequency Beamforming

Time-delay beamforming is now an option…

Dense mid-freq array: Antenna sep ~ 20cmTime step ~ 1ns ~ 30 cmAngle step > 45 deg

Sparse low-freq array: Antenna sep ~2 mTime step ~ 1ns ~ 30 cmAngle step ~10 deg – less if interpolate

Front end unit can combine space-freq beamforming in a single FIR-like structure

Golden Rule: throw away redundant data before spending energy processing/transporting it

Station processor

Optical-electro

Heirarchicalprocessor

Electro-optical

Multi-chip module

M&C

Optical-electro

Heirarchicalprocessor

Electro-optical

Multi-chip module

M&C

Optical-electro

Heirarchicalprocessor

Electro-optical

Multi-chip module

M&C

Optical-electro

Heirarchicalprocessor

Electro-optical

Multi-chip module

M&C

Optical-electro

Heirarchicalprocessor

Electro-optical

Multi-chip module

M&C

Clock & control

Station Processor

2nd LevelBeamforming

2nd LevelChannelisation

CornerTurner

Station Calibration and Correlator

Inputs: 64-dual pol 1st stage beams

Outputs: selectable dual-pol beams@ 1GSA/s @ 2-bit re/2-bit imag

Channelisation to 4096 channelsWith a 1024 channeliser

Station Calibration and station correlator

Output is optical and correlator ready

Simulations

Multi-Level Beamforming

Split the problem to be hierarchical and parallel.

Station divided into tiles (can be logical).

Dump as much unwanted data as we can early on.

Tile beam

Station beams

Simple Beam Patterns

80 x 80 degrees:

Station beam at (45, 87) degrees.Tile beam at zenith.

Visualisation of beams

Elevation 85 - 90 degrees

1000 MHz65536 antennas, 256 tiles

Station beams 0.05 degrees apartTile beams 2 degrees apart27 tile beams, 31707 station beams

Run time: 5.67 seconds

Station beams 0.20 degrees apartTile beams 2 degrees apart27 tile beams, 8005 station beams Run time: 2.18 seconds

Dynamic Range SimulationCourtesy: S. Schediwy & Danny Price

This is the reason a correlator is required for a beamformer

Auto-power beam Peak power 0 dBArray station sparsed x3

Cross-power beam 3deg rotationPeak power -20dB

Cross-power beam 30 deg rotation Peak power -50dB

Examples of Implementation

Introduction564m24 segments

640m64 cylinders

32m dish

Medicina Radio Telescopes

BEST-2

BEST-3Lo

BEST-2 specs

N cylinders 8N receivers 32Total collecting area

1357.98 m2

Total effective area 964.17 m2

Central Freq. 408 MHzFrequency BW 16 MHzIF 30MHzLongest baseline N/S E/W

70m 17.04m

Primary FOV 37.65 deg2

Sensitivity / Antenna Gain 0.363 K/Jy

Aeff / Tsys 11.651 m2/K

Transit time at delta = 45 deg

2353.3 sec.Marco Bartolini, IRA - INAF

64- C

hann

el A

DC

F - R

OAC

H

X -

ROAC

HS

- RO

ACH

B -

ROAC

H

HOST - PC GPU Imaging & Calibration

GPU Transient

1Gb-E

10 Gb-e

Richard

Griffin

Jack

PCI-XJack

Alessio

Dickie

OeRC

Medicina Radio Telescopes

Medicina Backend: Spatial FFT

Danny Price – Jack Hickish

Medicina Fringes…

Medicina Fringes (Cas. A.)

Cas. A. Image

Industrial Engagement It is NOT the intention of the SKA community to deliver 'finished' chip

designs yet. Aiming for detailed device specifications ready to start prototype

manufacture when NRE money available There are basic engineering processes that have to be done to enable

meaningful sizing, cost & power estimation IP identification and development – potential industrial involvement Development of strategic technology partnerships

ADC design IP macros for eg FFT, switch fabric Embedded controllers Non-packaged device mounting

Identification of key architectural features Identify appropriate optimisation opportunities and trade-offs. Development of accurate models for cost and power analysis at the

wider system level. Identify key interface 'Hot Spots' and apply effort accordingly

Industrial Engagement Multi-Chip Module (One Chip to Rule them all!)

4 x 4 antenna array (currently) – easily extended to 8x8

Can also be used for Phased Array feeds for dishes

Current Chip RFI protection shows -57dB/m (in air)

ADC FIR-FFT

ProcessorBeam Combiner

&Calibrator

Optical I/ORF IN

OpticalOUT

16-8 bit1GS/s

1024 channelsplitter

16 elementBeam combiner

OpticalChip

UWBRX

10mW/FFT10mW/channel 4mW/Beam ??

Requirement Specifications