Experimental GPU Acceleration Benchmark

This report is an experimental GPU evidence lane. It is intentionally not a production speed claim: the Metal path brute-forces packed Hamming k=1 distances and is compared against DotMatch’s existing CPU indexed Hamming assignment with identical output checks.

For the shipping dotmatch crispr-count CLI on MAGeCK/Yusa and Sanson/Brunello, see Production CRISPR CPU vs Metal. That report records when Metal is eligible, when --metal-validate passes, and when CPU remains faster or authoritative.

The decision rule is simple: GPU rows must have zero mismatches before any speed result is considered, and CPU-indexed throughput remains the production baseline unless the GPU path is faster end-to-end on real workloads.

Synthetic Figure

GPU speedup

Public CRISPR Figure

Public CRISPR GPU speedup

Synthetic CPU vs Metal Rows

n_reads	n_targets	len	k	gpu_reads_per_sec	cpu_reads_per_sec	gpu_vs_cpu_kernel	gpu_vs_cpu_total	mismatches
20000	737	20	1	4146940.0	297594.0	13.93x	13.50x	0
50000	4096	20	1	3649260.0	366657.0	9.95x	10.05x	0

Public CRISPR CPU vs Metal Rows

total_reads	packable_reads	n_targets	target_start	target_length	gpu_reads_per_sec	cpu_reads_per_sec	gpu_vs_cpu_kernel	gpu_vs_cpu_total	mismatches	count_delta
20000	19954	87437	23	19	433598.0	191601.0	2.26x	1.92x	0	0

Synthetic Raw Rows

tool	backend	status	workload	n_reads	n_targets	len	k	prep_seconds	seconds	total_seconds	reads_per_sec	total_reads_per_sec	pairs_per_sec	checksum	device	notes
dotmatch_cpu_index	cpu	ok	synthetic_hamming	20000	737	20	1	0.000604375	0.0672056	0.06781	297594	294942	2.19327e+08	124483641	Apple M5	CPU indexed Hamming k=1 baseline; prep is index build
dotmatch_gpu_metal	metal	ok	synthetic_hamming	20000	737	20	1	0.0001995	0.00482283	0.00502233	4.14694e+06	3.98221e+06	3.05629e+09	124483641	Apple M5	Metal brute-force packed Hamming k=1; prep is shared-buffer allocation and copy
dotmatch_cpu_index	cpu	ok	synthetic_hamming	50000	4096	20	1	0.00291654	0.136367	0.139284	366657	358979	1.50183e+09	1714574030	Apple M5	CPU indexed Hamming k=1 baseline; prep is index build
dotmatch_gpu_metal	metal	ok	synthetic_hamming	50000	4096	20	1	0.000155292	0.0137014	0.0138567	3.64926e+06	3.60836e+06	1.49474e+10	1714574030	Apple M5	Metal brute-force packed Hamming k=1; prep is shared-buffer allocation and copy

Public CRISPR Raw Rows

tool	backend	status	workload	total_reads	packable_reads	n_targets	target_start	target_length	k	input_seconds	prep_seconds	seconds	total_seconds	reads_per_sec	total_reads_per_sec	assigned_unique	assigned_exact	assigned_corrected	ambiguous	unmatched	invalid_windows	non_acgt_windows	checksum	mismatches	count_delta	device	notes
dotmatch_cpu_index	cpu	ok	public_crispr_yusa_hamming	20000	19954	87437	23	19	1	0.0747252	0.0536802	0.104384	0.232789	191601	85914.6	18376	17894	482	0	1578	0	46	15101300162	0	0	Apple M5	CPU indexed Hamming k=1 public CRISPR baseline
dotmatch_gpu_metal	metal	ok	public_crispr_yusa_hamming	20000	19954	87437	23	19	1	0.0747252	0.000191333	0.0461257	0.121042	433598	165232	18376	17894	482	0	1578	0	46	15101300162	0	0	Apple M5	Metal public CRISPR FASTQ extract-pack-dispatch-readback-count lane

Scope

This lane tests whether GPU compute is worth productizing. It currently covers fixed-length packed A/C/G/T Hamming k=1 assignment only. The public CRISPR row includes FASTQ parsing, guide-window extraction, packing, Metal dispatch, readback, and count/QC aggregation. It does not cover Levenshtein indels, BCL conversion, N/IUPAC GPU fallback, CUDA deployment, or production scheduling.

A future production GPU path needs additional real-workload gates for feature-barcode and BCL lanes, plus CPU fallback for non-A/C/G/T windows before promotion out of experimental status.