TTDFT

Official Resources

• Repository: https://github.com/ttdftdev/ttdft_public
• License: Information available in repository
• Documentation: See GitHub README

Overview

TTDFT is a high-performance Real-Space Time-Dependent Density Functional Theory code designed for modern heterogeneous computing architectures. It is written in C++ and leverages NVIDIA GPUs for significant acceleration of key operations, such as matrix multiplications in the Kohn-Sham propagation.

Scientific domain: Molecular dynamics, electronic structure, excited states Target user community: HPC users with GPU resources, developers of real-space methods

Theoretical Methods

• Real-Space TDDFT: Grid-based discretization of the Kohn-Sham equations
• Tensor Structured Algorithms: Utilization of Tucker tensor decomposition
• Tamm-Dancoff Approximation (TDA): Supported
• Real-Time Propagation: Magnus expansion and other integrators
• Chebyshev Filtering: Used for efficient eigensolvers

Capabilities

• GPU Acceleration: Native support for CUDA, cuBLAS, cuSparse
• Real-Time Dynamics: Simulation of electron dynamics under external fields
• Large-Scale Systems: Optimized for systems that benefit from tensor compression
• Ground State: Fast ground state convergence via Chebyshev filtering

Inputs & Outputs

• Input formats: text-based input files (implied from standard C++ scientific codes)
• Output data types:
  • Energy and forces
  • Time-dependent dipole moments
  • Absorption spectra (via Fourier transform of dipoles)
  • Density cubes

Performance Characteristics

• Speed: ~8x speedup reported for matrix-matrix multiplications on GPUs compared to CPU-only.
• Parallelization: Hybrid MPI and CUDA.
• Efficiency: Tensor-structured operations reduce memory footprint and computational complexity.

Computational Cost

• Memory: Tensor decomposition helps reduce storage requirements for large grids.
• Hardware: Requires NVIDIA GPUs for optimal performance (module load cuda).

Limitations & Known Constraints

• Hardware: Strongly tied to NVIDIA ecosystem (CUDA).
• Compilation: Heterogeneous build process requires careful environment setup (MPI + NVCC).
• Documentation: Primary documentation is the GitHub README and associated papers.

Comparison with Other Codes

• vs Octopus: Both real-space; TTDFT emphasizes tensor compression and C++/CUDA acceleration.
• vs GCEED: GCEED focuses on coupled Maxwell-TDDFT; TTDFT focuses on tensor algorithms and GPU compute.

Best Practices

• Compilation: Ensure cuda modules are loaded.
• GPU Usage: Target problems large enough to saturate GPU compute capability for best efficiency.

Citations

• Primary: "TTDFT: A GPU accelerated Tucker tensor DFT code for large-scale Kohn-Sham DFT calculations" (arXiv/OSTI).