single station correlator(uni)
DESCRIPTION
Single Station Correlator(UNI). A Widefield Imager. Science: Large field, short time resolution imaging for transient events. Daily/Weekly sky survey Engineering: Scalable FPGA based correlator design GPU image processing backend Prototype for large-N EoR arrays and SKA-Low Names: - PowerPoint PPT PresentationTRANSCRIPT
Griffin Foster [email protected]
A Widefield ImagerScience: Large field, short time resolution imaging for transient events. Daily/Weekly sky survey
Engineering: Scalable FPGA based correlator design GPU image processing backend Prototype for large-N EoR arrays and SKA-Low
Names: SEPCAM (SEPnet CAMera) LASI (LOFAR All Sky Imager) UNI (UnNamed Imager) …
Griffin Foster [email protected]
Instrument Overview
RSP
RSP
x24
GPU Imaging Node
GPU Imaging Node
FX Correlator
LOFAR HBA
Real Time Imager
Post-Processing/
Storage Server
x192
FFT
FFT
ANT 0
ANT 192
Cornerturn(192 Antennas x 1024 Channels)
CH 0
CH 1023
CMAC
CMAC
Griffin Foster [email protected]
Correlator Specifications
FPGA Based (UniBoard) 96 dual-pol HBA antennas → 18624 baselines 4 bit correlation 15-30 MHz Bandwidth 1024 FFT Channels Integration times ~10 ms Minimal modification to LOFAR station
Griffin Foster [email protected]
UniBoard Developed at ASTRON 8 × Altera Stratix IV 40nm FPGA, type
EP4SGX230KF40C2, 1288 multipliers, 1517 pins
One front node → all back nodes mesh 14 layers 8 × 2 DDR3 modules 4 × 4 10GbE links in 4 × 4 8-bit LVDS out Estimated maximum power consumption
280 W
Griffin Foster [email protected]
Correlator Design
Back Node FPGA48 antpols/BN4 BNs
XAUI TX
CLK @ 4 x Data CLK
Low Band FFT
EQ/4b Quant
Cornerturn
48 antpol X Engine
Xeng Out
LVDS Mesh
Front Node FPGA256 Channels/FN4 FNs
CLK @ 4 x Data CLK
LVDS Mesh
FN1
FN2
FN3
BN1
BN2
BN3
Mux
192 antpol X Engines
Vacc
Vacc
Xeng Out
Daterates:Input @ 25 MHz BW: 76.8 GbpsOutput: 141 MB per integration
Processing:6984 MACs @ 200 MHz: 1.4 TMACsAt 25 MHz BW complete correlation possible with 1 UniBoard
UniBoard
Griffin Foster [email protected]
RSP Interface
AP0
AP1
AP2
AP3
BP
RCU dataRCU ctrl
LCU
Inter board interface (IBI)
Inter board interface (IBI)
ring
ring
CEP
serdes
serdes
RCU dataRCU ctrl
RCU dataRCU ctrl
RCU dataRCU ctrl
Current RSP Design: Connected to adjacent RSP boardsvia a 4 lane(2.5 Gbps per lane) XAUI interface
Utilizing:Beamlets: 195312.5 Hz × 248 × 4 × 24b = 4.65 GbpsCrosslets: 195312.5 Hz × 96 × 4 × 18b = 1.35 GbpsOverhead: 0.025 GbpsTotal: 6.025 Gbps
Modified RSP Firmware Design: Signal Channel correlator dropped, utilize 2 lanes forthe beam ring and 2 for UNI. Requires a passive connectorto split/combine XAUI CX-4 interface
Utilizing:Beamlets: 195312.5 Hz × 248 × 4 × 24b = 4.65 GbpsOverhead: 0.025 GbpsTotal: 4.675 Gbps
RSP Board
CX-4 XAUI Connector
Griffin Foster [email protected]
Firmware Design Options12b Voltage samples1
12.5 MHz BW/ XAUI Lane Simplest firmware modification
8b Voltage samples (quant) 1
18.75 MHz BW/ XAUI Lane Maximize BW/ XAUI lane, require EQ control interface1Limited band selection
18b Subband samples2
8.5 MHz BW/ XAUI Lane Simpler firmware modification
8b Subband samples (quant) 2
18.75 MHz BW/ XAUI Lane Maximize BW/ XAUI lane, require EQ control interface2Selectable band, quantization required
Griffin Foster [email protected]
Fast Transient Imaging Channelized data is independent and is easily parallelized, GPUs
offer a good solution Short timescale images can be crude, thus traditional imaging
steps can be simplified/ignored But short timescale images must be formed faster then the
integration time, the wide field of view also introduces a number of challenges.
RFI calib
Array Model
Sky Model
Antenna Model
phase φ0 grid/w-proj fft
facet
clean
phase φ1
phase φn
Compare Threshold
Prev. Image
1 s Image
10 s image
Storage
GPU
FX
Detections
grid/w-proj fft clean
grid/w-proj fft clean
Griffin Foster [email protected]
Widefield Sky Survey Longer timescale transient events New viewable sky survey every few days Take advantage of the fast transient pipeline by using
the post RFI, calibrated data Perform secondary integration/data compression Follow a more traditional imaging pipeline Analysis will focus on rare, bright, transient events
which occur on timescales of minutes to days
Griffin Foster [email protected]
Where we are atFirmware: Main modules (FIR,FFT,X Engine) currently being written LOFAR RSP modifications and passive XAUI splitter done by
ASTRON
UniBoard: First round of boards(7) have been produced and shipped
to institutes
Imaging: Collaborators working on GPU imaging pipeline Widefield imaging techniques in development on
prototype array in Medicina
Related: AARTFAAC (Super Terp Correlator, ASTRON/U Amsterdam)