Skip to content

September 2026 release

Latest
Latest

Choose a tag to compare

View all tags
@rokzitko rokzitko released this 06 Sep 17:51
· 1 commit to master since this release
2026.09
967adeb

NRG Ljubljana 2026.09: Major New Release

This release summarizes changes since the 2024.12 release, with major improvements in GPU accelerator support, performance, memory efficiency, physical capabilities, and workflows.

CUDA And Numerical Performance

  • NVIDIA CUDA support has been added for numerical linear algebra operations.
  • CUDA-enabled builds can select diag=cuda to use cuSOLVER diagonalisation routines (real and complex).
  • Selected operator-recalculation matrix products can be offloaded with mult=cuda.
  • Eigenvector blocks and intermediate accumulators remain cached on the GPU during recalculation.
  • BLAS/LAPACK integration has been improved, including support for ILP64 numerical libraries for really large problems.
  • Threaded BLAS, MPI scheduling, and oversubscription diagnostics are handled more consistently.
  • CPU diagonalization now defaults to the divide-and-conquer LAPACK routines dsyevd and zheevd.

Performance And Memory

  • All unnecessary matrix, eigenspace, and density-matrix copies have been eliminated.
  • Multithreading in tools (kk, hilb, broaden).
  • Seed operators, diagonalization data, and obsolete eigenvector representations are released from memory earlier.
  • FDM and DMNRG spectral kernels now reuse weights, contractions, and existing operator-block information.
  • FDM back-iteration exploits the diagonal structure of discarded-state density matrices.
  • Memory requirements for diagonalization workspaces are now reported more clearly.
  • Completed Wilson shells no longer redundantly retain full eigenvector and operator-block matrices.
  • Thermodynamic histories and density-matrix back-iteration data now use separate, compact storage.

New Physics Capabilities

  • Multiple local phonon modes are supported, including independent bosonic cutoffs.
  • Tensor-product construction of phonon bases and mode-resolved operators has been generalized.
  • Superconducting Wilson chains can be instantiated using Nambu onsite and hopping coefficients.
  • Symmetry triangle inequalities are enforced before constructing reduced operator matrix elements.
  • Floquet-NRG support introduces Floquet bases
    and quasienergy-aware truncation.
  • Three-channel QST calculations received important low-energy Hamiltonian and recalculation fixes.
  • Orbital-triplet operator generation and several SNEG symbolic-algebra operations were corrected.

Spectra And Observables

  • report.nrg can list low-lying states together with diagonal observables, aiding fixed-point and sub-gap-state identification.
  • broaden accepts arbitrary user-provided output-frequency meshes.
  • adapt --flat Gamma directly supports constant hybridization functions.
  • hilb, kk, integ, and resample can select GSL's monotonicity-preserving Steffen interpolation.
  • integ is now a unified integrator.
  • Thermal Fermi and Bose kernels are stable at extreme energies and near the Bose pole.
  • FDM partition-function accumulators consistently retain high numerical precision.
  • Level-flow energies can be reported in user-selected or physical energy units.
  • Raw HDF5 output and broadening sum-rule diagnostics received correctness fixes.

Workflows And Reliability

  • Parameter files, spectral meshes, and truncated inputs now receive substantially stricter validation.
  • The new instantiate/nrgspawn workflow can run prepared model templates without invoking Mathematica for every parameter point.
  • Basis, Hamiltonian, and operator blocks can be saved and reused in parameter sweeps.
  • Conda packaging and build coverage now include broader Linux, macOS, ARM, BLAS, and compiler configurations.

Installation

NRG Ljubljana is available from conda-forge:

conda install -c conda-forge nrgljubljana

Packages are available for Linux x86-64, Linux aarch64, macOS x86-64,
and macOS arm64. Source builds remain appropriate when CUDA support or
nonstandard numerical-library configurations are required.

Requirements And Compatibility

  • The C++ runtime requires a C++20 compiler.
  • CPU diagonalization now defaults to the divide-and-conquer LAPACK routines dsyevd and zheevd. Numerical results should remain equivalent within floating-point tolerances.
  • BLAS/LAPACK and OpenMP runtime selection is handled more explicitly. Users combining MPI with threaded numerical libraries should review their rank and thread settings to avoid oversubscription.
  • CUDA support is optional and must be enabled explicitly in source builds. It is not currently included in the standard conda-forge packages.
Assets 2
Loading