Time to form one synthetic aperture sonar image

Forming one synthetic aperture sonar image of a 7,145 m² scene on one workstation takes 124 s at 20,000 scatterers per square meter with the Standard simulation fidelity, and 29.1 min with the Full reference, measured end to end. The sonar spent 71 s collecting the same data. At Standard fidelity the workstation forms the image in less time than that below about 3,617 scatterers per square meter. Each run propagates every ping to a seafloor of millions of individual scatterers and back, and then focuses the pings coherently into the image.

Last updated 26 October 2026 · version 8b9e256

Scene

Benchmark scene and trajectory

Every measurement on this page images the same patch of seafloor from the same trajectory. A HISAS-class sonar at 100 kHz passes once on a straight track along the near edge of the scene and looks across it, the standard strip-map geometry, so every object's shadow falls away from the track. The scene and the sonar are held fixed as a benchmark fixture, so later releases of the simulator are measured on exactly the same input and the results can be compared over time.

71 × 100 m

seafloor simulated, 7,145 m²

200

pings along a straight-line track

2.5 cm

image pixel, in both range and along track

71 s

the sonar's own collection time, at 2.83 pings per second

A simulated synthetic aperture sonar image of a seafloor scene holding boat hulls, a torpedo shape, containers, barrels, truss frames, a cable and a rock outcrop over rippled sand, with every shadow falling to the right, away from the track along the left edge — The scene at 80,000 scatterers per square meter, Standard fidelity. The sonar passes up the left edge and looks to the right; along track is up and range increases to the right.

Density

Image detail against scatterer density

Scatterer density sets how finely the seafloor is discretized. With too few scatterers per resolution cell the speckle is not fully developed; the image is grainy in a way real sonar imagery is not, and fine relief breaks up. Above some density the returns are statistically converged, and further scatterers add run time without changing the image.

The tiles show the same 20 m square of the scene, a rock outcrop with its acoustic shadow in a ripple field, at four densities, each formed at Standard fidelity. Each caption gives the density as scatterers per resolution cell, a cell being d / 2 = 2.0 cm along-track by 1.44 c / 2B = 3.6 cm of slant range (7.2 cm²), then the run's scatterer count and its end-to-end time.

A rock outcrop and its acoustic shadow on rippled sand, simulated at 200 scatterers per square meter — **200/m²** 0.14 per resolution cell 1 M scatterers · 56 s

A rock outcrop and its acoustic shadow on rippled sand, simulated at 2,000 scatterers per square meter — **2,000/m²** 1.4 per resolution cell 14 M scatterers · 66 s

A rock outcrop and its acoustic shadow on rippled sand, simulated at 20,000 scatterers per square meter — **20,000/m²** 14 per resolution cell 143 M scatterers · 124 s

A rock outcrop and its acoustic shadow on rippled sand, simulated at 40,000 scatterers per square meter — **40,000/m²** 29 per resolution cell 286 M scatterers · 4.2 min

Fidelity

Standard and Full simulation fidelity

The simulator offers two fidelity presets. Full is the reference: it keeps every scatterer the line of sight reaches, skips the transmit-beam prefilter and softens shadow edges with a Fresnel clearance test. Standard, the default, drops scatterers whose echo falls below 10⁻⁴ of the ping's peak or that lie outside the transmit beam's main lobe, and uses the hard line-of-sight test. Both presets image the same scatterers, so the two runs differ only in how much of the scene's echo they compute.

At 20,000 scatterers per square meter Standard forms the image 14.1 times faster than Full. Each time is the median of three runs per preset, interleaved so that any drift in the machine affects both presets alike.

Bar chart of end-to-end wall clock by preset at 20,000 scatterers per square meter: Standard 124 seconds and Full 1747 seconds, each split into simulation, beamforming, and scene build and file writing — End-to-end wall clock by preset, split into simulation, beamforming, and scene build and file writing.

The speedup is measured against the image Full forms. The difference map subtracts Full's magnitude from Standard's in decibels at every pixel, with time-varying gain off and no despeckling. Standard's image is on average 0.9 dB brighter than Full's; half of the pixels differ by less than 1.1 dB and 95 percent by less than 3.0 dB.

The magnitude of the focused image in decibels for Full and Standard side by side on one 60 dB scale — Image magnitude in decibels, Full and Standard on one scale, relative to Full's peak.

Map of the per-pixel decibel difference between Standard and Full over the image, on a scale from minus 10 to plus 10 dB, with red where Standard is brighter — Standard minus Full, in decibels, per pixel. Red marks pixels where Standard is brighter than Full.

Wall clock

Wall-clock time against scatterer density

The wall-clock time covers the whole path from the start of a run to the finished image: scene build, point-scatterer simulation, back-projection and writing the output image. Nothing is cached, and every density re-simulates from the scene file. All times were measured on an NVIDIA RTX 4500 Ada, 24 GB.

Standard runs from 56 s at 200 scatterers per square meter to 9.9 min at 80,000, which is 572 million scatterers. Full is measured from 200 to 20,000 per square meter.

Wall clock to form one image against scatterer density on log axes for Full and Standard, with a dashed line marking the sonar's 71 second collection time — The dashed line is the sonar's own collection time. Where a curve sits below it, the workstation formed the image faster than the platform gathered the data.

Real time

Wall-clock time against collection time

The sonar spent 71 s on this track collecting the data. A real-time factor above 1 means the workstation formed the finished image in less time than that.

At Standard fidelity the real-time factor falls to 1 near 3,617 scatterers per square meter. Below that density the workstation forms the image in less time than the sonar took to collect the data, and above it in more. Full keeps pace with the collection only at 200 per square meter, where it takes 69 s, and is slower at every higher density.

density	scatterers	Standard	real-time factor	Full	real-time factor
200	1 M	56 s	1.26x	69 s	1.02x
500	4 M	57 s	1.25x	94 s	0.75x
1,000	7 M	59 s	1.20x	137 s	0.52x
2,000	14 M	66 s	1.08x	3.7 min	0.32x
5,000	36 M	74 s	0.96x	8.0 min	0.15x
10,000	71 M	92 s	0.77x	15.0 min	0.08x
20,000	143 M	124 s	0.57x	29.1 min	0.04x
40,000	286 M	4.2 min	0.28x	not run	not run
80,000	572 M	9.9 min	0.12x	not run	not run

From 80,000 per square meter up, the scatterer cloud does not fit in the memory this workstation keeps free, so the run goes out of core: the cloud is held in disk-backed memory maps and streamed through the GPU in chunks, and the image is the same as an in-memory run would form. The 80,000 per square meter run held at most 6.2 GB of RAM. At 80,000 per square meter, 572 million scatterers, Standard forms the image in 9.9 min.

Method

Measurement method

Each run ran in its own process with its preset applied through the same settings the application uses, so the memory pressure of one run cannot distort the next.
Simulation was forced on every run, because a cache hit would measure the cache rather than the simulator.
The timer starts when a user presses Run and stops when the image is on disk, so scene build and file writing are included along with the GPU kernels.
At 20,000 scatterers per square meter each preset ran three times, interleaved, and the page reports the median; every other density ran once, in ascending order, so a failure at the top still leaves a usable curve.
The image comparison uses a second beamforming pass over each preset's own echoes with time-varying gain off, so the gain curve cannot hide or create a difference.
The scene is held fixed as a benchmark fixture, and every number on this page is read from the benchmark results rather than typed in.