This site tracks scikit-learn's performance across machines, builds and accelerated backends. It shows which workloads benefit from scikit-learn-intelex or an Array API backend, and when a different BLASBasic Linear Algebra Subprograms: the library that runs the matrix and vector products of NumPy and SciPy. OpenBLAS and MKL are two implementations of it./OpenMPStandard for running the loops of compiled code on several threads. scikit-learn's Cython code (HistGradientBoosting, KMeans, ...) uses it. buildOne packaging of the same scikit-learn version. Here, builds differ by their BLAS and OpenMP libraries. or machine changes anything. The comparison dashboard varies these (implementationCode that runs a given estimator: stock scikit-learn, scikit-learn-intelex or an Array API backend., build or hardware). The scalability ones look at how a fit speeds up with more CPU coresOne hardware core of the CPU. With SMT, it shows up as two logical CPUs.. Each dashboard explains how to read it in its own intro.
Overall findings:
-
On the benchmarked cases,
scikit-learn-intelexon CPUs is the most consistently fast option among the currently benchmarked set of options, especially for fitting tree-based models. Other backends and builds help in narrower, workload-specific cases.Note that scikit-learn-intelex only accelerates a subset of estimators and parameters (see its supported algorithms). Its behavior can also differ from scikit-learn, for instance in best-first tree growth, multiclass
predict_probaor missing values at predict time. These gaps are actively worked on. - For linear models, a good option for improved performance is simply to pick the MKL conda-forge buildIntel's Math Kernel Library (oneMKL): a BLAS implementation tuned for Intel CPUs, available as a conda-forge build option. of scikit-learn, instead of the PyPIThe Python Package Index, where pip installs packages from. Its scikit-learn wheels bundle their own OpenMP runtime, and NumPy and SciPy bundle OpenBLAS. one.
- Using a bigger machine doesn't help a lot for a single fit for most models except random forests and alike. But it helps a lot for multiple parallelized fits, typically for hyper-parameter tuning workloads.
-
cuML LogisticRegression always runs the quasi-Newton solverThe optimization algorithm used by fit, for instance lbfgs or newton-cholesky for LogisticRegression.
qn(L-BFGS or OWL-QN). A point may still be labeledlbfgs,newton-cg, ornewton-cholesky: the wrapper maps every one of those names ontoqn, so those cuML cases time the same algorithm.
Dashboards
| k-means: fit time against inertia | Which k-means implementation and settings give the lowest inertiaWhat k-means minimizes: the squared distance of each sample to its nearest center, here averaged over all samples. Lower is better. for a given time budget, on CPUs and GPUs. |
Benchmark hardware
Intel Core Ultra laptop
Hardware
CPU
Intel(R) Core(TM) Ultra X7 358H
X86_64, 16 physical cores, 16 logical CPUs
31 GB RAM
$1,299, released 2026
GPU(s)
level_zero:gpu:0: Intel(R) Arc(TM) B390 GPU (29 GB, integrated, priced with the CPU)
Apple M4 laptop
Hardware
CPU
Apple M4
ARM_8, 10 physical cores, 10 logical CPUs
32 GB RAM
$1,599, released 2024
GPU(s)
mps: Apple M4 GPU (32 GB, integrated, priced with the CPU)
Apple M2 laptop
Hardware
CPU
Apple M2
ARM_8, 8 physical cores, 8 logical CPUs
16 GB RAM
GPU(s)
mps: Apple M2 GPU (16 GB, integrated, priced with the CPU)
Intel Core Ultra desktop
Hardware
CPU
Intel(R) Core(TM) Ultra 9 285
X86_64, 24 physical cores, 24 logical CPUs
61 GB RAM
GPU(s)
level_zero:gpu:0: Intel(R) Arc(TM) B580 Graphics (12 GB)
High-end Intel Xeon server
Hardware
CPU
Intel(R) Xeon(R) 6787P
X86_64, 172 physical cores, 344 logical CPUs
499 GB RAM
$11,648, released 2025
GPU(s)
No GPU detected.
NVIDIA L4 VM
Hardware
CPU
AMD EPYC 7R13 Processor
X86_64, 4 physical cores, 8 logical CPUs
30 GB RAM
~$1/h (cloud VM), released 2024
GPU(s)
cuda:0: NVIDIA L4 (22 GB, priced with the CPU)
64-core AMD EPYC VM
Hardware
CPU
AMD EPYC 9555P 64-Core Processor
X86_64, 64 physical cores, 64 logical CPUs
252 GB RAM
~$3/h (cloud VM), released 2024
GPU(s)
No GPU detected.