Äîêóìåíò âçÿò èç êýøà ïîèñêîâîé ìàøèíû. Àäðåñ îðèãèíàëüíîãî äîêóìåíòà : http://www.atnf.csiro.au/vlbi/dokuwiki/lib/exe/fetch.php/difx/meetings/sydney2012/difx_askap.pdf Äàòà èçìåíåíèÿ: Tue Sep 25 06:35:20 2012 Äàòà èíäåêñèðîâàíèÿ: Tue Apr 12 12:41:53 2016 Êîäèðîâêà: Ïîèñêîâûå ñëîâà: molecules

ASKAP Signal Processing Overview
DIFX Users and Developers Meeting
Dr. Grant Hampson | ADE Team Leader 4 September 2012
CSIRO ASTRONOMY AND SPACE SCIENCE


ASKAP Overview
· 36x12-metre antennas
· 630 baselines

· 188-port PAF receiver
· " Tuneable" over 0.7-1.8GHz

· Digitisation
· Provides at least 300MHz processed BW

· 30deg2 FoV
· 36-dual pol. Beams (varies with frequency) · Carries through to correlator · Tied array outputs also

2 | DIFX Users and Developers Meeting | Grant Hampson | 25 Sep 12


ASKAP Differences
· BETA (Eng. Test Array)
· · · · Dual heterodyne receiver ADC at antenna Dragonfly-2 digitiser (IF sampling) Redback-2 DSP board inside ATCA chassis ­ Xilinx Virtex-6 FPGAs

· ADE (ASKAP Design Enhancements)
· · · · RFOF (RF over fibre) from PAF to central site All digital electronics in central site Dragonfly-3 digitiser (direct sampling) Redback-3 DSP board in 1U chassis ­ Xilinx Kintex-7 FPGAs

3 | DIFX Users and Developers Meeting | Grant Hampson | 25 Sep 12


Digital Back-end Blocks
PAF ADC Phase Switch

Digital Receiver

Coarse Filter Bank

Channel Select

PAF Cross Connect

192

36-antennas 192-ports 2x16-bits 300 MHz =66Tbps =6600 x 10G 36-antennas 2x36-beams 2x16-bits 300 MHz =25Tbps =2500 x 10G

Array Covariance Matrix Beamformer

Beamformer
Coarse Delay Fin1 e Filter Bank

Fringe

1

Antenna Cross Connect

Correlator
Correlator Tied Array Long Term Accumlator

36

4 | DIFX Users and Developers Meeting | Grant Hampson | 25 Sep 12


PAF Electronics ­ The Domino
· The "Domino" combines:
· Low noise amplifier · RF filtering ­ defines 3 bands · RF over SM fibre transmitter

· Each Domino contains electronics for 2 PAF ports:
· 1.8W power per port · 250grams per port

· Backplanes provide power, control and monitoring · Duplex LC/APC connections to MTP elite, connect RF to central site
5 | DIFX Users and Developers Meeting | Grant Hampson | 25 Sep 12


ADE Backend ­ Digital Receiver
· Upgrade to Xilinx Series-7 FPGAs
· Kintex-325 FPGAs (840-DSP/RAM) · Contains first stage coarse filter bank

· Direct sampling ADCs
· NatSemi 12-bit 1600MSPS ADCs

· Avago optical MiniPODs
· Multimode 120Gbps modules · FPGA connect directly at 10Gbps

· 1U chassis packaging
· No front or rear to boards · 16 ports per 1U · ~10W per PAF port
6 | DIFX Users and Developers Meeting | Grant Hampson | 25 Sep 12


ASKAP (BETA) Processed Spectrum
Filt erbank 0 -10 -20 -30 Magnitude (dB ) -40 -50 -60 -70 -80 -90 -100 -1 -0. 5 0 0. 5 Frequenc y (MHz) 1 1.5 2

Sampled Bandwidth = 384MHz Processed Bandwidth = 304MHz

Stage 1 filterbank
384MHz 768MHz

Beamformer

Second Nyquist zone First-stage filterbank output: 304 x 1MHz channels oversampled by 32/27

Stage 2 filterbank
54 x 18.52kHz fine channels (= 1MHz) sent to the correlator

Critically sampled filterbank

Second-stage filterbank output: 64 x 18.52kHz channels per 1MHz critically sampled

7 | DIFX Users and Developers Meeting | Grant Hampson | 25 Sep 12


Frequencies?
· Can't have a single 768MHz channel ­ need to channelise
· · · · Beamformer requires ~1MHz resolution to apply weights RFI ­ can remove frequencies that are troublesome ­ few MHz Allows looking at frequencies that are well separated Allows different simultaneous zooms on the same frequencies

· Zooms
· Fine delays/fringe stopping require ~100kHz to avoid smearing at band edges · To avoid aliasing near edges of 1MHz channels, and frequency rolloff ­ The 1MHz channels are oversampled by 32/27=18.5%

· Tied Array / VLBI
· Fine channels are stitched back together to form larger BWs (e.g., 16MHz) · To avoid aliasing and frequency rolloff, the fine channels are planned to also be oversampled · Don't require sample rates to be a power of 2
8 | DIFX Users and Developers Meeting | Grant Hampson | 25 Sep 12


ADE DSP: Beamformer and Correlator
· Common hardware platform · Xilinx Kintex-420 Series-7 FPGAs
· ~1680 DSP/RAM per FPGA · 2 DDR3 modules @ 1066MTS · 120Gbps optical transceiver

· Direct 10Gbps from FPGA to optics transceivers simplifies design · 1U chassis packaging
· No front or rear to boards · 6 FPGAs per Redback board

· Use of optical backplanes
· · · · Low capital cost ­ no power 1.5Tbps optical backplane Acts like "fixed" network switch Scalable and modular

9 | DIFX Users and Developers Meeting | Grant Hampson | 25 Sep 12


Why use FPGAs for Beamforming?
· Why use FPGAs?
· · · · · 10k multipliers per Redback-3 board (@300MHz = 3Tera ops/sec) We are using 36x6=216 Redback-3 boards in beamformer Beamforming, ACM and Fringe Large processed BW ­ 300MHz with 36 dual pol beams is significant Large bitwidths ­ linear up to beamformer output ­ preserving dynamic range · Lots of IO ­ guaranteed IO and lots of it · Power efficient

· Why not use a GPU?
· Network switch is a killer: 6600 input ports + 6600 output ports ­ size, price, power, cabling, ??? · Are GPU's IO constrained for such an application? · How many GPUs per PC?
10 | DIFX Users and Developers Meeting | Grant Hampson | 25 Sep 12


Why use FPGAs for Correlator?
· We are using 75 Redback-3 boards in correlator (225TOPS) · Why not use a GPU? · Network switch: 2500 input ports + 2500 output ports
­ size, price, power, cabling, ???

· Processing still very regular
· Correlation cells are not that different to beamformers (CMAC with DDR3)

· Have to also manage Tied Array outputs, inverse filterbank · The balance of on-chip resources: logic, RAM, DSP units along with the high IO bandwidth and relatively low power consumption still makes the FPGA a compelling choice in obtaining a cost effective solution for these types of processing applications.

11 | DIFX Users and Developers Meeting | Grant Hampson | 25 Sep 12


My two bits ...
· For single pixel feeds
· No beamformer · Correlator is so much smaller · Pretty simple ­ GPU?

· ASKAP ~1 Peta Op · Data rates going up · 10G to 25G · 10GbE pretty slow · 100GbE needs to become standard · Density going up · Better cooling

· However, single pixel feeds can be wider BW
· Maybe a factor of 10 larger? How many more dishes?

· Issues with GPUs???
· · · · · · Network switch ports and BW Watching bit widths carefully Conversion fixed-floating point Data rates into GPU cards? Density of processing? Software tools

· Issues with FPGAs???
· · · · · · Communications Programming Software tools Fixed interconnections? Compilation times Test benches

12 | DIFX Users and Developers Meeting | Grant Hampson | 25 Sep 12


We acknowledge the Wajarri Yamatji people as the traditional owners of the Observatory site.

Thank you
CSIRO Astronomy and Space Science Dr. Grant Hampson ADE Team Leader t + 61 2 9372 4647 e grant.hampson@csiro.au w www.csiro.au/cass

CSIRO ASTRONOMY AND SPACE SCIENCE