SDRmm

Develop

GPU measurements

Apple M4 Max wgpu compute and WebGL2 waterfall benchmarks showing where GPU offload wins.

Apple M4 Max (40 GPU cores, Metal, wgpu 30.0.1, release; 8 warmups + 101 iterations including upload and readback). Raw CSV: gpu-m4-max.csv. Passive radar numbers: Hardware tests.

Workload CPU median GPU median Default
Spectrum, 64 × 4,096 0.48 ms 0.33 ms CPU
Spectrum, 64 × 32,768 4.89 ms 1.23 ms GPU
Spectrum, 64 × 65,536 11.3 ms 2.29 ms GPU
Spectrum, 64 × 262,144 56.0 ms 7.89 ms GPU
Filter bank, 13 bands × 8,192 79.5 µs 177.3 µs CPU
  • gpu-fft: portable WGSL (f32); software adapters rejected; CI gpu verifies Vulkan parity on Lavapipe.
  • Policy: spectra < 32,768 pts stay on CPU; larger offload only if startup probe beats CPU by ≥ 20% (capped at 25% GPU time). Radar caps at 512 MiB and falls back to CPU on any error.
  • WebGL2 waterfall (Chrome 153 Metal vs SwiftShader): 640×240 @ 2× = 0.076 ms vs 2.236 ms; 1280×720 @ 2× = 0.246 ms vs 12.242 ms. Repaints on change only (zero idle draws). Raw JSONL: before, after, WebKit.
Reproduce benchmarks
cargo test -p sdrmm-engine --no-default-features --features gpu-fft --release --lib -- --ignored --nocapture --test-threads=1 gpu::benchmarks
cargo test -p sdrmm-engine --no-default-features --features gpu-fft --release --lib -- --include-ignored --nocapture --test-threads=1 spectrum::tests gpu::caf::tests array::radar::gpu
node web/scripts/benchmark-gpu.mjs

Edit this page on GitHub

Esc