–
FFTs per second
–
MS/s transformed
–
MS/s from the source
–
blocks dropped (total)
–
accelerator busy
–
CPU used (4 cores)
Signal (PL test generator)
Stream
What runs where
The pattern generator in the PL writes 64-bit beats of int16 I/Q into a DDR ring; the accelerator transforms each block of 16 × 4096 samples as it lands. The A53 only hands out ring slots (a small C loop) and turns ~60 blocks a second into the rows you see; everything else is the accelerator.
Image
2-D spectrum (log magnitude, centred) — paint here
Inverse FFT of the painted spectrum
removedboosted ×4symmetric strokes keep the image real
–
filtered images / s
–
round trip, ms
–
inverse FFT call, ms
–
accelerator, ms
–
CPU used (4 cores)
Select Image
Brush
Filters
–
volumes / s
–
accelerator per 3-D FFT, ms
–
3-D FFT call incl. copies, ms
–
CPU spectrum + reduce, ms
–
CPU used (4 cores)
Field (on the board)
Rendering (in your browser)
–
224-point FFTs / s
–
accelerator per FFT, ms
–
FFT call incl. copies, ms
–
span shown
–
bins per pixel
–
CPU used (4 cores)
Fly to
fs = 224 Hz, so each of the 16,777,216 bins is exactly 1 Hz wide. Moving: a tone drifting across the band, one hopping between 6.4 and 8.2 MHz every frame, one swinging around 3.5 MHz.
Signal (made on the board)
Aperture (512²) — draw on it
Far-field intensity, |2-D FFT|², log scale
–
patterns / s
–
round trip, ms
–
FFT call incl. copies, ms
–
accelerator, ms
–
CPU used (4 cores)
Aperture
Optics
Azimuth slice (x–depth)
Elevation slice (y–depth)
C-scan (x–y at one depth)
–
volumes / s
–
accelerator per volume, ms
–
CPU Stolt remap, ms
–
time per volume, ms
–
CPU used (4 cores)
Probe (simulated on the board)
3 MHz, half-wavelength pitch, 256 samples per element (16 mm). Phantom: speckle, an anechoic cyst, four point targets and a bright sphere orbiting through the volume.
View (in your browser)
Range–Doppler map, all channels integrated · circles: CFAR detections
Bird’s-eye view · outlines: the true objects
–
frames / s
–
points per frame
–
CFAR detections
–
accelerator per frame, ms
–
radar pipeline per frame, ms
–
front-end simulation, ms
–
CPU used (4 cores)