sonar challenge problem

Sonar Challenge ProblemUpdated May 23, 2001

System Overview

• Use existing MFP demo design– Insert adaptive front end (MSEDR) between

incoming data stream and existing matched field design

• it will adapt the weights used in the existing k- beamformer

– Use the PE1 and PE2 subvoxel beamformers as-is as much as possible.

Challenge Overview

• Form covariance matrix at 3KHz rate– Requires 96GB/sec memory bandwidth !!!!!!

• Inversion of covariance matrix happens every 3-5 seconds– Don’t do inversion

• Determine K smallest eigenvalues – K << N, K ~= 10?

– Use Lanczos’ method to find eigenvalues• A subspace method which hopefully gives good

approximation to R

– Do this part on Pentium

Forming Covariance Matrix (R)

• Forming the covariance matrix (R)– 16MB memory locations updated @ 3KHz rate

• Requires 96GB/sec memory bandwidth• SV2 has 6GB/sec memory bandwidth

– Solutions• Downsample input data• Multiple boards

– 8 SV2 boards can form R10 PE’s per SV2read/MAC/write is kernel (2 cycles)pack 2 data elements per memory locationone SV2 can do 10 X 1 kernel per cycle = 1500M

kernels/secrequirement is 4M kernels @ 3KHz rate = 12000M

kernels/sec

Forming Covariance Matrix (continued)

• Bernecky has recently (since the Midway meeting) suggested a way of doing the R matrix computation in 50x50 chunks, and avoiding most off-chip memory accesses

• This is an interesting idea that may or may not work better than the previous slide’s approach. It will be investigated.

Rest of Algorithm

• Normalization may be required with this algorithm– to be determined by NUWC– prior work (2 years ago) showed normalization

doubled computation required, may be the case here

• Other things may be need to be included depending on results of NUWC Matlab experiments

Statement of 1-Year Goals

• Do a lab demonstration of AMFP system– Use recorded data at NUWC– In the absence of a full-up lab demo, have

enough of the pieces completed (sizing, design, simulation, or execution) to make a case for or against the approach.

Demonstration Architecture

• 5 SV2 boards– 4 for creating/updating R (PE0 replacement)– 1 for subvoxel interpolation (PE1 & PE2

replacement)

• External Pentium

• Requires ACS API + Myrinet

Tasks

• Determine whether subspace methods will work in this problem (NUWC)

• Determine time requirements for Lanczos method on a Pentium (Athanas)

• Determine effects of 2X downsampling on algorithm performance (NUWC)

• End-to-end MATLAB simulation (NUWC)

More Tasks

• MSEDR floating point requirements– Formats, rounding modes, etc. (VTech/BYU)

• Determine communications requirements/patterns (all)

• Create single board MSEDR FPGA design (BYU)• Expand to 4 boards (VTech)• Port 4K subvoxel beamformer design to SV2

(NUWC)

Yet More Tasks

• Integrate Lanzcos code (VTech)

• System integration/test (NUWC)

Schedule

• Determine whether subspace methods will work in this problem (NUWC) (2 months)

• Determine time requirements for Lanczos method on a Pentium (Athanas) (2 months)

• Determine effects of 2X downsampling on algorithm performance (NUWC) (3 months)

• End-to-end MATLAB simulation (NUWC) (3 months)

More Schedule

• MSEDR floating point requirements– Formats, rounding modes, etc. (VTech/BYU) (3 months)

• Determine communications requirements/patterns (all) (3 months)

• Virtex-II JHDL support (3 months)

• Create single board MSEDR FPGA design (BYU) (4 months)

• Expand to 4 boards (VTech) (10 months)

• Port 4K subvoxel beamformer design to SV2 (NUWC) (10 months)

Yet More Schedule

• Integrate Lanzcos code (VTech) (11 months)

• System integration/test (VTech/NUWC) (12 months)