Speed on real recordings

From the benchmark of release 0.0.4 (2026-10-29): read and process, nothing written, in multiples of real time, on an RTX 2070 Max-Q and an i7-9750H.

Recording	Chain	CUDA	wgpu	CPU runtime
Neuropixels, 385 channels	zero-phase + CMR	12.5×	10.1×	0.4×
Neuropixels, 385 channels	causal + CMR	10.5×	8.5×	0.7×
HD-EMG, 32 channels, 300 s	zero-phase + CMR	56×	45×	13×
HD-EMG, 32 channels, 300 s	causal + CMR	76×	59×	38×

With a Zarr file written (dspk run), Neuropixels dropped to 2–4×.

These numbers are from before two improvements

Since then, vector kernels handle odd row lengths, which is what held the CPU runtime back on 385 channels, and the writer runs on its own thread and compresses in parallel. The benchmark hasn’t been re-run yet, so expect these numbers to be pessimistic.

Latency

Streaming on CUDA, 1 ms of Neuropixels took 0.11 ms, and 10 ms took 0.36 ms. wgpu has a floor of about 0.4 ms per chunk, from submission overhead. CPU p99 jitter was up to about 10 times the median.

Device time alone

Data already on the device, autotuned, one block of 384 channels × 60,000 samples (2 s at 30 kHz):

Operation	CUDA	wgpu (Vulkan)	CPU runtime
IIR band-pass, 5th order, forward	5.3 ms	12.3 ms	39 ms
FIR, 63 taps, causal	3.1 ms	6.4 ms	33 ms
Running median, width 9	0.59 ms	0.69 ms	20 ms
Common median reference	10.3 ms	16.5 ms	239 ms
Common average reference	1.0 ms	1.2 ms	11 ms
resample_poly 1 / 3	1.4 ms	1.7 ms	16 ms

Reproduce with cargo test -p dspk-benchmarks --profile validate --test bench_kernels -- --ignored --nocapture.