LoginSign Up
← Back to all sparks
C

CP2K

DEVOPS
Velocity3.8

Quantum chemistry and solid state physics simulation package

cp2k.org ↗

CP2K is rebuilding a legacy Fortran DFT code around GPUs, ML potentials, and k-points

dftk-pointsgpu-accelerationml-potentialselectrochemistrybuild-modernization
◆Current state
CP2K ships twice a year and each release lands a wide slate of quantum-chemistry methods rather than a single headline feature. The last two years have been dominated by three parallel threads: pushing k-point support into methods that were previously gamma-point only, wiring in external machine-learning and GPU libraries, and modernizing the build. The 2026.2 release is the first where GPU work reaches the exact-exchange hot path and where grand-canonical SCF opens electrified-interface simulation.
◆Where it's heading
The code is converging on a plugin-heavy architecture: DeePMD-kit, NequIP, DFTD4, SIRIUS, greenX, GauXC and now libGint all arrive as external libraries CP2K orchestrates rather than reimplements. Build modernization finished on schedule — the Makefile was deprecated in 2025.2 and deleted in 2026.1 — and the same discipline is visible in the steady removal of superseded modules. Method coverage is being made uniform across periodic and molecular paths, with k-points the recurring gap being closed release after release.
◆Prediction
Expect 2027.1 to continue the k-point sweep into the remaining gamma-point-only analyses and to broaden libGint's CUDA exchange beyond its initial path. The release notes flag FFTW3 as a likely hard dependency, so the next breaking change is probably build-side rather than scientific.

◆Recent moves

  1. 10d ago

    CP2K 2026.2: CUDA exact exchange and grand-canonical SCF

    ⚡ SPARK

    The clearest step yet in CP2K's GPU thread: libGint brings CUDA acceleration to Hartree-Fock exchange, the most expensive part of hybrid-functional runs. Alongside it, constant-potential grand-canonical SCF and a planar counter-charge model add a workload class the code could not previously address.

    View source ↗
  2. 6mo ago

    CP2K 2026.1 adds MiMiC multiscale interface, drops the Makefile

    A broad methods release that also completes the build migration promised a year earlier: the GNU Makefile and QUIP are removed, leaving CMake as the only path. The MiMiC interface and SF-TDDFT extend the multiscale and excited-state surfaces without redirecting them.

    View source ↗
  3. 1y ago

    CP2K 2025.2 ships GFN-xTB and RIXS, last Makefile release

    GFN-xTB and a resonant inelastic X-ray scattering module widen the method catalog at both the cheap-semiempirical and expensive-spectroscopy ends. The note that this is the final Makefile release sets up the removal that lands in 2026.1.

    View source ↗
  4. 1y ago

    CP2K 2025.1 adds Bethe-Salpeter optical spectra and Harris/EHT

    Bethe-Salpeter optical spectra and real-time BSE propagation deepen the excited-state stack, while Harris and EHT methods add cheaper approximations at the other end. PAO prediction from equivariant PyTorch models continues the ML-assisted basis thread.

    View source ↗
  5. 1y ago

    CP2K 2024.3 patches an MPI stall in MD runs

    A single-fix release for a PW environment issue that could stall MPI molecular dynamics. The maintainers note the previous release's serial binaries are unaffected, which is the clearest signal of how narrow the change is.

    View source ↗
  6. 1y ago

    CP2K 2024.2 adds DeePMD-kit, DFTD4 and OpenCL GPU support

    Three external libraries land at once — DeePMD-kit for ML potentials, DFTD4 for dispersion, and OpenCL for GPU offload — which is the pattern the code has followed since. GW for small unit cells with full k-points marks an early point in the k-point sweep that later releases extend.

    View source ↗