Hardware comparison

Compare hardware and output modes

8,192 correlations × 4,096 points: 16 data spectra and 512 templates. Warm public run() calls, one CPU thread, setup and upload excluded; GPU synchronization included. Hierarchical rows use the captured PyCBC reference profile and are recorded at several SNR thresholds. Recorded October 3, 2026.

Enable JavaScript to change the comparison.

Full output → peaks avoids writing every lag; hierarchy also skips most full transforms. Automatic coarse choices vary by device; requested dismissal budgets are not measured here. Whiskers show timing-block 10th–90th percentiles, not confidence intervals.

FFT-only references: FFTW on CPU, cuFFT on NVIDIA, rocFFT on Radeon 8060S, MLX on M2. Missing references and failed checks are shown explicitly; software GPUs are excluded. The Xeon result is from a virtual machine. All timing samples and configuration details are in the recorded JSON (teaser-fleet-20261003.json and gpu-fleet-512x512-20261003.json).