scikit-learn benchmark dashboard

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. Most dashboards below vary one of these (implementationCode that runs a given estimator: stock scikit-learn, scikit-learn-intelex or an Array API backend., build or hardware) and keep the others fixed. 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-intelex on 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_proba or 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 labeled lbfgs, newton-cg, or newton-cholesky: the wrapper maps every one of those names onto qn, so those cuML cases time the same algorithm.

Dashboards

Benchmark hardware