Performance and Speedups

Zero-code-change acceleration with cuml.accel delivers an approximately 5.8× median speedup across training benchmark cases completed on both CPU and GPU, measured on NVIDIA RTX Pro 6000. Training performance varies with the estimator, dataset size and shape, and hyperparameters, but most measured workloads are faster on GPU, with the largest gains appearing on wider or larger datasets.

Speedup by operation and workload

Rows are ranked by the median exact speedup. Green indicates a measured gain, gray is centered at 1×, and warm colors indicate a slowdown. Hatching marks CPU-timeout lower bounds and unavailable results.

A timeout indicates that the complete isolated CPU benchmark case—not a single estimator call—exceeded its wall-clock limit. When the GPU run completed, the displayed value is a conservative lower bound (≥), not an exact speedup.

Training and combined operations

PCA performance depends strongly on the requested component rank. The table below holds the input at 61,035 rows by 4,096 features and varies only n_components. The timings show an approximate crossover from transfer-dominated low-rank work to positive speedup at higher ranks.

See the detailed PCA results for the remaining workloads.

Medium-wide PCA fit-transform by component rank

Components	PCA CPU / GPU / result
128	1.42 s / 2.97 s / 0.48×
256	2.15 s / 2.91 s / 0.74×
512	3.34 s / 2.98 s / 1.12×
1,024	8.14 s / 3.1 s / 2.62×

Inference and transforms

The heatmap shows the 10 operations with the highest median exact speedup. Detailed estimator tables retain every inference and transform result, including short operations where dispatch and data-conversion overhead dominate.

Detailed benchmark results

Use the workload guide to interpret the heatmaps, then open an estimator for exact shapes, throughput or wall times, slowdowns, and timeouts. Detailed results express inference and transform operations as throughput in samples per second, abbreviated as M/s, k/s, or /s, with one input row counted as one sample. The wall time used to calculate throughput is shown below it. Training and combined-operation rows show wall time only.

Five relative workload categories, from transfer-bound to compute-heavy

Every operation uses the same five relative workload categories, ordered from smaller or more transfer-sensitive cases to larger or more compute-intensive cases. Concrete shapes follow operation-family policies, with workload sizes chosen to complete within reasonable runtime and resource limits. These sizes are illustrative examples across several workload regimes, not scaling limits or maximum supported sizes. The same category can therefore represent different rows, features, and input sizes across operations; exact values appear in the estimator tables.

Workload dimensions and decimal float32 X size

Label	Rows	Features	Input
small.balanced	2,000–781,250	32–1,024	0.256 MB–205 MB
medium.thin	8,000–15,625,000	8–128	0.256 MB–500 MB
medium.balanced	8,000–3,906,250	32–1,024	1.02 MB–2 GB
medium.wide	8,000–976,563	128–4,096	8.19 MB–2 GB
large	32,000–15,625,000	32–1,024	4.1 MB–10 GB

Results by estimator

Open an estimator for every workload, throughput or wall time, slowdown, timeout, and actual shape.

Expand all Collapse all

Linear models

ElasticNet

ElasticNet results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit	small.balanced	50,000	128	25.6 MB	24.4 ms	16.2 ms	1.51×
fit	medium.thin	1,562,500	16	100 MB	216.7 ms	47.7 ms	4.55×
fit	medium.balanced	500,000	128	256 MB	388.4 ms	314.8 ms	1.23×
fit	medium.wide	244,140	1,024	1,000 MB	2.66 s	2.47 s	1.08×
fit	large	5,000,000	128	2.56 GB	3.67 s	3.2 s	1.15×
predict	small.balanced	195,313	128	100 MB	32.1M/s 6.1 ms	23.4M/s 8.3 ms	0.73×
predict	medium.thin	7,812,500	16	500 MB	251M/s 31.2 ms	176M/s 44.3 ms	0.70×
predict	medium.balanced	976,563	128	500 MB	32.6M/s 29.9 ms	23.8M/s 41.1 ms	0.73×
predict	medium.wide	488,281	1,024	2 GB	4.74M/s 103.0 ms	3.05M/s 160.0 ms	0.64×
predict	large	5,000,000	128	2.56 GB	41.1M/s 121.6 ms	24.1M/s 207.2 ms	0.59×

Lasso

Lasso results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit	small.balanced	50,000	128	25.6 MB	22.6 ms	16.7 ms	1.35×
fit	medium.thin	1,562,500	16	100 MB	130.2 ms	46.7 ms	2.79×
fit	medium.balanced	500,000	128	256 MB	398.2 ms	322.9 ms	1.23×
fit	medium.wide	244,140	1,024	1,000 MB	2.62 s	2.45 s	1.07×
fit	large	5,000,000	128	2.56 GB	3.75 s	3.19 s	1.17×
predict	small.balanced	195,313	128	100 MB	35M/s 5.6 ms	23.5M/s 8.3 ms	0.67×
predict	medium.thin	7,812,500	16	500 MB	247M/s 31.6 ms	176M/s 44.3 ms	0.71×
predict	medium.balanced	976,563	128	500 MB	34.4M/s 28.4 ms	23.8M/s 41.1 ms	0.69×
predict	medium.wide	488,281	1,024	2 GB	4.9M/s 99.6 ms	3.04M/s 160.5 ms	0.62×
predict	large	5,000,000	128	2.56 GB	41.1M/s 121.6 ms	24M/s 208.0 ms	0.58×

LinearRegression

LinearRegression results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit	small.balanced	50,000	128	25.6 MB	54.9 ms	7.3 ms	7.50×
fit	medium.thin	1,562,500	16	100 MB	223.0 ms	17.0 ms	13.1×
fit	medium.balanced	500,000	128	256 MB	805.3 ms	46.3 ms	17.4×
fit	medium.wide	244,140	1,024	1,000 MB	4.99 s	113.0 ms	44.2×
fit	large	5,000,000	128	2.56 GB	10.6 s	1.04 s	10.2×
predict	small.balanced	195,313	128	100 MB	22.7M/s 8.6 ms	23M/s 8.5 ms	1.01×
predict	medium.thin	7,812,500	16	500 MB	248M/s 31.5 ms	178M/s 43.9 ms	0.72×
predict	medium.balanced	976,563	128	500 MB	32.6M/s 30.0 ms	23.8M/s 41.0 ms	0.73×
predict	medium.wide	488,281	1,024	2 GB	4.87M/s 100.2 ms	3.05M/s 160.3 ms	0.63×
predict	large	5,000,000	128	2.56 GB	41M/s 121.9 ms	24.1M/s 207.2 ms	0.59×

LogisticRegression

LogisticRegression results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit	small.balanced	50,000	128	25.6 MB	29.4 ms	6.3 ms	4.64×
fit	medium.thin	1,562,500	16	100 MB	683.6 ms	18.0 ms	38.0×
fit	medium.balanced	500,000	128	256 MB	369.5 ms	29.6 ms	12.5×
fit	medium.wide	244,140	1,024	1,000 MB	533.0 ms	96.6 ms	5.52×
fit	large	5,000,000	128	2.56 GB	3.32 s	266.2 ms	12.5×
predict	small.balanced	195,313	128	100 MB	39.2M/s 5.0 ms	23.2M/s 8.4 ms	0.59×
predict	medium.thin	7,812,500	16	500 MB	188M/s 41.6 ms	180M/s 43.4 ms	0.96×
predict	medium.balanced	976,563	128	500 MB	31.7M/s 30.8 ms	23.8M/s 41.0 ms	0.75×
predict	medium.wide	488,281	1,024	2 GB	3.82M/s 127.9 ms	3.04M/s 160.7 ms	0.80×
predict	large	5,000,000	128	2.56 GB	38.6M/s 129.5 ms	24.1M/s 207.4 ms	0.62×

Ridge

Ridge results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit	small.balanced	50,000	128	25.6 MB	27.6 ms	7.1 ms	3.86×
fit	medium.thin	1,562,500	16	100 MB	106.0 ms	18.4 ms	5.75×
fit	medium.balanced	500,000	128	256 MB	221.4 ms	49.8 ms	4.44×
fit	medium.wide	244,140	1,024	1,000 MB	537.9 ms	129.7 ms	4.15×
fit	large	5,000,000	128	2.56 GB	1.54 s	490.0 ms	3.15×
predict	small.balanced	195,313	128	100 MB	35.1M/s 5.6 ms	23.6M/s 8.3 ms	0.67×
predict	medium.thin	7,812,500	16	500 MB	253M/s 30.9 ms	178M/s 43.8 ms	0.71×
predict	medium.balanced	976,563	128	500 MB	34.1M/s 28.7 ms	23.8M/s 41.1 ms	0.70×
predict	medium.wide	488,281	1,024	2 GB	4.93M/s 99.0 ms	3.04M/s 160.5 ms	0.62×
predict	large	5,000,000	128	2.56 GB	40M/s 125.1 ms	24.2M/s 206.7 ms	0.61×

Clustering and manifold learning

DBSCAN

DBSCAN results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit_predict	small.balanced	2,000	32	0.256 MB	5.7 ms	1.1 ms	5.00×
fit_predict	medium.thin	8,000	8	0.256 MB	70.5 ms	2.0 ms	35.1×
fit_predict	medium.balanced	8,000	32	1.02 MB	19.5 ms	1.9 ms	10.3×
fit_predict	medium.wide	8,000	256	8.19 MB	62.4 ms	3.1 ms	19.9×
fit_predict	large	32,000	32	4.1 MB	168.5 ms	13.9 ms	12.2×

HDBSCAN

HDBSCAN results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit_predict	small.balanced	2,000	32	0.256 MB	40.3 ms	7.4 ms	5.46×
fit_predict	medium.thin	8,000	8	0.256 MB	105.0 ms	10.2 ms	10.3×
fit_predict	medium.balanced	8,000	32	1.02 MB	211.2 ms	10.6 ms	19.9×
fit_predict	medium.wide	8,000	256	8.19 MB	4.76 s	14.7 ms	325×
fit_predict	large	32,000	32	4.1 MB	1.09 s	34.3 ms	31.7×

KMeans

KMeans results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit_predict	small.balanced	50,000	128	25.6 MB	66.5 ms	24.1 ms	2.77×
fit_predict	medium.thin	1,562,500	16	100 MB	492.9 ms	81.8 ms	6.02×
fit_predict	medium.balanced	500,000	128	256 MB	526.2 ms	143.4 ms	3.67×
fit_predict	medium.wide	244,140	1,024	1,000 MB	6.8 s	477.5 ms	14.2×
fit_predict	large	5,000,000	128	2.56 GB	13.5 s	1.3 s	10.4×
predict	small.balanced	195,313	128	100 MB	36.1M/s 5.4 ms	21.4M/s 9.1 ms	0.59×
predict	medium.thin	7,812,500	16	500 MB	107M/s 72.7 ms	150M/s 51.9 ms	1.40×
predict	medium.balanced	976,563	128	500 MB	32.9M/s 29.7 ms	22.9M/s 42.6 ms	0.70×
predict	medium.wide	488,281	1,024	2 GB	5.76M/s 84.8 ms	3M/s 162.5 ms	0.52×
predict	large	5,000,000	128	2.56 GB	37M/s 135.3 ms	23.4M/s 213.8 ms	0.63×

t-SNE

t-SNE results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit_transform	small.balanced	2,000	64	0.512 MB	1.62 s	397.6 ms	4.07×
fit_transform	medium.thin	10,000	8	0.32 MB	8.88 s	468.1 ms	19.0×
fit_transform	medium.balanced	10,000	64	2.56 MB	8.71 s	581.8 ms	15.0×
fit_transform	medium.wide	10,000	512	20.5 MB	9.26 s	614.1 ms	15.1×
fit_transform	large	50,000	64	12.8 MB	44.2 s	904.7 ms	48.9×

UMAP

UMAP results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit_transform	small.balanced	2,000	64	0.512 MB	2.43 s	112.2 ms	21.7×
fit_transform	medium.thin	10,000	8	0.32 MB	5.27 s	108.8 ms	48.4×
fit_transform	medium.balanced	10,000	64	2.56 MB	5.88 s	264.8 ms	22.2×
fit_transform	medium.wide	10,000	512	20.5 MB	6.29 s	325.4 ms	19.3×
fit_transform	large	50,000	64	12.8 MB	13.4 s	252.0 ms	53.0×
transform	small.balanced	390,625	64	100 MB	Timeout at 70 s	2.97M/s 131.3 ms	≥529×
transform	medium.thin	15,625,000	8	500 MB	Timeout at 100 s	8.76M/s 1.78 s	≥57.0×
transform	medium.balanced	1,953,125	64	500 MB	Timeout at 100 s	2.64M/s 739.3 ms	≥138×
transform	medium.wide	244,141	512	500 MB	Timeout at 100 s	369k/s 661.1 ms	≥154×
transform	large	7,812,500	64	2 GB	Timeout at 14 min	1.47M/s 5.32 s	≥160×

Neighbors and density estimation

KNeighborsClassifier

KNeighborsClassifier results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
predict	small.balanced	2,000	128	1.02 MB	146k/s 13.7 ms	2.09M/s 0.96 ms	14.3×
predict	medium.thin	8,000	16	0.512 MB	310k/s 25.8 ms	8.35M/s 0.96 ms	26.9×
predict	medium.balanced	8,000	128	4.1 MB	177k/s 45.1 ms	3.64M/s 2.2 ms	20.5×
predict	medium.wide	8,000	1,024	32.8 MB	32.7k/s 244.5 ms	707k/s 11.3 ms	21.6×
predict	large	32,000	128	16.4 MB	82.2k/s 389.5 ms	1.8M/s 17.8 ms	21.9×

KNeighborsRegressor

KNeighborsRegressor results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
predict	small.balanced	2,000	128	1.02 MB	149k/s 13.4 ms	2.74M/s 0.73 ms	18.4×
predict	medium.thin	8,000	16	0.512 MB	226k/s 35.4 ms	12.2M/s 0.66 ms	54.0×
predict	medium.balanced	8,000	128	4.1 MB	187k/s 42.8 ms	4.13M/s 1.9 ms	22.1×
predict	medium.wide	8,000	1,024	32.8 MB	34k/s 235.6 ms	730k/s 11.0 ms	21.5×
predict	large	32,000	128	16.4 MB	89.6k/s 357.2 ms	1.88M/s 17.0 ms	21.0×

KernelDensity

KernelDensity results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
score_samples	small.balanced	2,000	128	1.02 MB	3.29k/s 608.7 ms	1.97M/s 1.0 ms	599×
score_samples	medium.thin	8,000	16	0.512 MB	3.1k/s 2.58 s	13.5M/s 0.59 ms	4372×
score_samples	medium.balanced	8,000	128	4.1 MB	831/s 9.62 s	4.11M/s 1.9 ms	4942×
score_samples	medium.wide	8,000	1,024	32.8 MB	Timeout at 100 s	741k/s 10.8 ms	≥9426×
score_samples	large	32,000	128	16.4 MB	168/s 190 s	2.18M/s 14.7 ms	12969×

NearestNeighbors

NearestNeighbors results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
kneighbors	small.balanced	2,000	128	1.02 MB	135k/s 14.9 ms	2.98M/s 0.67 ms	22.2×
kneighbors	medium.thin	8,000	16	0.512 MB	282k/s 28.4 ms	12M/s 0.67 ms	42.4×
kneighbors	medium.balanced	8,000	128	4.1 MB	207k/s 38.6 ms	3.72M/s 2.2 ms	17.9×
kneighbors	medium.wide	8,000	1,024	32.8 MB	33.2k/s 241.2 ms	721k/s 11.1 ms	21.7×
kneighbors	large	32,000	128	16.4 MB	88k/s 363.7 ms	1.85M/s 17.3 ms	21.0×

Decomposition

PCA

PCA results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit_transform	small.balanced · 256 components	10,000	1,024	41 MB	347.9 ms	114.6 ms	3.04×
fit_transform	medium.thin · 32 components	195,313	128	100 MB	115.6 ms	256.0 ms	0.45×
fit_transform	medium.balanced · 256 components	100,000	1,024	410 MB	2.19 s	1.67 s	1.31×
fit_transform	medium.wide · 128 components	61,035	4,096	1,000 MB	1.42 s	2.97 s	0.48×
fit_transform	medium.wide · 256 components	61,035	4,096	1,000 MB	2.15 s	2.91 s	0.74×
fit_transform	medium.wide · 512 components	61,035	4,096	1,000 MB	3.34 s	2.98 s	1.12×
fit_transform	medium.wide · 1,024 components	61,035	4,096	1,000 MB	8.14 s	3.1 s	2.62×
fit_transform	large · 256 components	1,000,000	1,024	4.1 GB	16.7 s	20 s	0.83×
transform	small.balanced · 256 components	50,000	1,024	205 MB	1.71M/s 29.3 ms	173k/s 289.5 ms	0.10×
transform	medium.thin · 32 components	976,563	128	500 MB	14M/s 69.6 ms	1.39M/s 700.9 ms	0.10×
transform	medium.balanced · 256 components	488,281	1,024	2 GB	2.05M/s 237.9 ms	89.7k/s 5.44 s	0.04×
transform	medium.wide · 1,024 components	122,070	4,096	2 GB	280k/s 436.5 ms	23.2k/s 5.27 s	0.08×
transform	large · 256 components	2,441,406	1,024	10 GB	2.49M/s 982.0 ms	91k/s 26.8 s	0.04×

PCA performance depends strongly on both input feature width and the number of retained components; results can vary substantially across these dimensions.

Ensembles

RandomForestClassifier

RandomForestClassifier results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit	small.balanced	50,000	128	25.6 MB	439.4 ms	140.3 ms	3.13×
fit	medium.thin	1,562,500	16	100 MB	7.81 s	778.6 ms	10.0×
fit	medium.balanced	250,000	128	128 MB	2.25 s	284.9 ms	7.90×
fit	medium.wide	244,140	1,024	1,000 MB	7.95 s	791.2 ms	10.0×
fit	large	1,953,125	128	1 GB	28.6 s	1.12 s	25.6×
predict	small.balanced	195,313	128	100 MB	6.67M/s 29.3 ms	15.8M/s 12.4 ms	2.37×
predict	medium.thin	7,812,500	16	500 MB	8.79M/s 888.6 ms	106M/s 74.0 ms	12.0×
predict	medium.balanced	976,563	128	500 MB	6.91M/s 141.3 ms	16.5M/s 59.3 ms	2.38×
predict	medium.wide	488,281	1,024	2 GB	2.36M/s 206.8 ms	2.13M/s 229.1 ms	0.90×
predict	large	5,000,000	128	2.56 GB	6.21M/s 804.8 ms	16.8M/s 297.9 ms	2.70×

RandomForestRegressor

RandomForestRegressor results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit	small.balanced	50,000	128	25.6 MB	2.95 s	716.6 ms	4.12×
fit	medium.thin	1,562,500	16	100 MB	19.8 s	1.63 s	12.2×
fit	medium.balanced	250,000	128	128 MB	17.7 s	1.8 s	9.80×
fit	medium.wide	244,140	1,024	1,000 MB	Timeout at 100 s	10.6 s	≥9.59×
fit	large	1,953,125	128	1 GB	208 s	7.6 s	27.4×
predict	small.balanced	195,313	128	100 MB	5.07M/s 38.5 ms	16.4M/s 11.9 ms	3.23×
predict	medium.thin	7,812,500	16	500 MB	7.15M/s 1.09 s	111M/s 70.3 ms	15.5×
predict	medium.balanced	976,563	128	500 MB	5.04M/s 193.9 ms	16.6M/s 58.7 ms	3.30×
predict	medium.wide	488,281	1,024	2 GB	2.17M/s 224.7 ms	2.13M/s 229.2 ms	0.98×
predict	large	5,000,000	128	2.56 GB	3.91M/s 1.28 s	17M/s 293.5 ms	4.36×

Preprocessing

StandardScaler

StandardScaler results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit_transform	small.balanced	50,000	128	25.6 MB	21.7 ms	11.6 ms	1.87×
fit_transform	medium.thin	1,562,500	16	100 MB	142.0 ms	76.8 ms	1.85×
fit_transform	medium.balanced	500,000	128	256 MB	218.2 ms	101.0 ms	2.16×
fit_transform	medium.wide	244,140	1,024	1,000 MB	788.0 ms	232.5 ms	3.39×
fit_transform	large	5,000,000	128	2.56 GB	2.1 s	958.2 ms	2.20×
transform	small.balanced	195,313	128	100 MB	8.73M/s 22.4 ms	11M/s 17.8 ms	1.25×
transform	medium.thin	7,812,500	16	500 MB	54.6M/s 143.1 ms	92.8M/s 84.2 ms	1.70×
transform	medium.balanced	976,563	128	500 MB	8.67M/s 112.6 ms	11.6M/s 83.9 ms	1.34×
transform	medium.wide	488,281	1,024	2 GB	1.14M/s 428.8 ms	1.47M/s 331.6 ms	1.29×
transform	large	5,000,000	128	2.56 GB	8.83M/s 566.3 ms	11.8M/s 423.1 ms	1.34×

TargetEncoder

TargetEncoder results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit_transform	small.balanced	195,313	32	25 MB	541.1 ms	5.35 s	0.10×
fit_transform	medium.thin	3,125,000	8	100 MB	1.76 s	1.49 s	1.18×
fit_transform	medium.balanced	781,250	32	100 MB	2.53 s	5.81 s	0.43×
fit_transform	medium.wide	250,000	128	128 MB	3.27 s	24 s	0.14×
fit_transform	large	7,812,500	32	1 GB	25.6 s	8.14 s	3.15×
transform	small.balanced	781,250	32	100 MB	679k/s 1.15 s	1.15M/s 679.1 ms	1.70×
transform	medium.thin	15,625,000	8	500 MB	4.29M/s 3.64 s	27.5M/s 568.2 ms	6.41×
transform	medium.balanced	3,906,250	32	500 MB	673k/s 5.81 s	2.23M/s 1.75 s	3.31×
transform	medium.wide	976,563	128	500 MB	158k/s 6.2 s	130k/s 7.52 s	0.82×
transform	large	15,625,000	32	2 GB	662k/s 23.6 s	2.69M/s 5.82 s	4.06×

Kernel methods

SVC

SVC results for all measured operations and workloads

Operation	Workload	Rows	Features	Input	CPU	GPU	Result
fit	small.balanced	2,000	64	0.512 MB	46.4 ms	19.6 ms	2.37×
fit	medium.thin	8,000	8	0.256 MB	145.4 ms	21.6 ms	6.75×
fit	medium.balanced	8,000	64	2.05 MB	502.6 ms	32.0 ms	15.7×
fit	medium.wide	8,000	512	16.4 MB	3.42 s	60.5 ms	56.6×
fit	large	32,000	64	8.19 MB	8.59 s	76.8 ms	112×
predict	small.balanced	2,000	64	0.512 MB	26.8k/s 74.6 ms	2.33M/s 0.86 ms	87.0×
predict	medium.thin	8,000	8	0.256 MB	32.6k/s 245.2 ms	6.12M/s 1.3 ms	188×
predict	medium.balanced	8,000	64	2.05 MB	10.7k/s 750.6 ms	2.86M/s 2.8 ms	268×
predict	medium.wide	8,000	512	16.4 MB	1.26k/s 6.36 s	425k/s 18.8 ms	338×
predict	large	32,000	64	8.19 MB	3.97k/s 8.06 s	1.59M/s 20.2 ms	399×

What determines GPU speedup

Work per call

cuml.accel benefits when each estimator call has enough parallel work to outweigh dispatch and compatibility overhead. Training and compute-intensive operations usually accelerate more than short prediction or transformation calls in practice.

Shape and parameters

Performance depends on feature width, operation, and estimator parameters, not just total workload size. Similar-sized datasets can produce different speedups when shapes or configurations differ. We selected workloads to illustrate these differences.

Data movement

cuml.accel typically ingests and returns NumPy arrays, converting data for GPU execution and scikit-learn compatibility. Native cuML can reduce this overhead with GPU-resident arrays, especially when pipelines perform repeated estimator operations.

Methodology and reproducibility

These benchmarks compare the corresponding CPU implementations with cuml.accel on NVIDIA RTX Pro 6000 across five relative workload categories. CPU measurements use the corresponding scikit-learn estimators, except for UMAP, which uses umap.UMAP from umap-learn, and HDBSCAN, which uses hdbscan.HDBSCAN from hdbscan. Each isolated case used one warmup, the median of three measured repetitions, operation-appropriate correctness validation, and a complete-case timeout.

Test system, validation, and timing policy

System. NVIDIA RTX PRO 6000 Blackwell Workstation Edition (102.0 GB), AMD Ryzen Threadripper PRO 7975WX 32-Cores, and 134.1 GB of system memory.

Timing. One warmup followed by three measured repetitions; tables use the median end-to-end wall time. Speedup is CPU median wall time divided by accelerated median wall time. For predict, transform, score_samples, and kneighbors, throughput is input rows divided by the same median wall time, with one row counted as one sample.

Scale and timeout policy. Algorithm-family-calibrated small, medium, and large shapes. Each backend ran every case in a separate worker with a wall-clock limit. The limit covered process startup, data preparation, estimator setup, one warmup, three measured repetitions, and correctness validation.

Execution and correctness. Successful accelerated measurements were instrumented to verify GPU-only execution. The benchmark runner applied each case’s operation-appropriate parity check. A timeout means validation could not complete.

Packages. cuml 26.10.0a69, hdbscan 0.8.44, scikit-learn 1.9.0, umap-learn 0.5.12.

Interpretation. These measurements describe the tested cases on this system; results for other workloads and systems will vary.

Continue with cuml.accel

Getting started

Installation and zero-code-change usage.

Getting Started with cuml.accel

Compatibility

Supported estimators and parameter behavior.

Accelerated Estimator Support

Profiling

Find GPU execution and fallback behavior.

Logging and Profiling