GPU Acceleration of Collinear and Noncollinear DFT Using a Numerical Atomic Orbital-Based DFT Code
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "GPU Acceleration of Collinear and Noncollinear DFT Using a Numerical Atomic Orbital-Based DFT Code".
Mira: Here's the gist of the paper. GPU acceleration of collinear and noncollinear density functional theory in numerical atomic orbital-based codes demonstrates practical speedups for large-scale calculations The GPU acceleration achieved for collinear DFT (3.27 times speedup in diagonalization for a 1,200-atom system) falls significantly short of the theoretical peak performance ratio between the H100 (51 TFLOPS with Tensor Cores) and the Xeon 8468 (3.2 TFLOPS)—a factor of approximately 16 This discrepancy was primarily because only dense matrix multiplications and eigenvalue solvers were offloaded to the GPU The researchers implemented GPU acceleration in OpenMX version 3.9.9 by modifying routines during diagonalization, specifically focusing on matrix multiplications and eigenvalue solves For collinear DFT, the procedure involves several steps: 1.
Kai: Diagonalize S to obtain a unitary matrix U and a diagonal matrix s such that U†SU = s 2.
Mira: Construct V = Us−1/2 3.
Kai: Form H′ = V†HV 4.
Paper summary: Kai: We’ve looked at the high-level summary of this paper about GPU acceleration for collinear and noncollinear density functional theory using a numerical atomic orbital based code called OpenMX. The main point is that they implemented GPU acceleration by offloading specific parts of the math, namely matrix multiplications and eigenvalue solves, to cuBLAS and cuSOLVER libraries >
Mira: They claim this works for both collinear DFT and noncollinear DFT, which is a big step because most existing implementations were only for collinear cases >
Lev: The abstract mentions that benchmarks were run on the Pegasus supercomputer using a forty-eight-core Intel Xeon Platinum eight thousand four hundred sixty-eight CPU and one NVIDIA H100 GPU per node > <ref:2606.09058#pg1,a 48-core Intel Xeon Platinum 8468 CPU and one NVIDIA H100>
Kai: And they report specific speedups. For a five hundred twelve atom collinear case on two nodes, the GPU accelerated calculation got a two point zero two times speedup over the CPU only run with ninety six CPU cores total >
Mira: That’s two point two times for that specific setup, but they also show results for noncollinear DFT <ref:2606.09058#pg1>. For a three eighty four atom case on two nodes, the speedup is twenty-six zero times over two CPU only nodes >
Lev: So the paper is showing practical speedups on real hardware, but the comparison against the theoretical peak performance of an H100 versus a Xeon eight thousand four hundred sixty-eight shows that they aren't hitting every possible theoretical limit yet > <ref:2606.09058#pg1>
Kai: It makes sense. They only offloaded those specific matrix multiplications and eigenvalue solves, which are definitely where the GPU shines compared to just regular CPU work >
Mira: So for the listener, it means that when you need to do these kinds of complex electronic structure calculations, you can use this acceleration method to get a tangible speedup on current systems >
Lev: But we have to remember that even with this acceleration, the diagonalization step still dominates the cost in subsequent self consistent field cycles, accounting for eighty two point five percent of the total wall time >
Conclusion: Kai: Looking at these results from "GPU Acceleration of Collinear and Noncollinear DFT Using a Numerical Atomic Orbital-Based DFT Code," the authors are essentially showing that they can make these numerical atomic orbital based codes run faster on GPUs >
Mira: The title highlights both collinear and noncollinear DFT, which is important because the noncollinear case allows for spin orientation to vary three dimensions, which is much more realistic for many physical systems >
Lev: For real hardware development, the implication is that you can start testing these acceleration techniques on actual supercomputer setups like Pegasus >
Kai: The authors acknowledge a limitation in the abstract. They noted that while GPU acceleration for collinear DFT has been reported before, it hasn't been reported for noncollinear DFT, which is why this work is important >
Mira: So the main implication is that they’ve extended the utility of existing GPU acceleration techniques to cover both major types of density functional theory calculations >
Lev: It’s about showing that you can get speedups by targeting the matrix operations, but as Kai mentioned, they only offloaded those specific parts >
Kai: Exactly. So while it's a solid practical result for these specific system sizes on two nodes, we still have a gap between what this shows and the absolute theoretical maximum performance of a high-end GPU setup >
Department of Physics, Niigata University · Nuclear Science and Engineering Center, Japan Atomic Energy Agency · Institute for Solid State Physics, The University of Tokyo
physics.comp-ph
Submitted: 2026-06-08
Updated: 2026-11-11
Comments: 25 pages, 14 figures, 24 tables. v2: revised to agree with the published version; corrections to text, tables, citations and references
Journal ref: J. Phys. Soc. Jpn. 94, 124003 (2025)
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 85/100
The gist: GPU acceleration of collinear and noncollinear density functional theory in numerical atomic orbital-based codes demonstrates practical speedups for large-scale calculations The GPU acceleration…
Key concepts
- Collinear DFT
- A type of density functional theory calculation where the electronic orbitals are assumed to be collinear. This simplifies some mathematical steps in solving the equations, allowing for a more streamlined process when performing calculations on hardware like GPUs.
- Noncollinear DFT
- A more complex type of DFT calculation where the electronic orbitals are not aligned along a single axis. This requires handling more intricate mathematical procedures, such as similarity transformations and solving generalized eigenvalue problems to find the correct solutions.
- GPU Acceleration
- The process of using Graphics Processing Units (GPUs) to speed up computationally intensive tasks in scientific codes. In this study, researchers modified routines within the code to offload specific operations like matrix multiplications and eigenvalue solvers from the CPU onto the GPU for faster processing.
- Diagonalization
- A mathematical procedure used to find the eigenvalues and eigenvectors of a matrix. In DFT calculations, this step is crucial for determining key properties of the system, such as energy levels and orbital shapes. The paper focuses on how efficiently this process can be accelerated on a GPU.
Terminology
Summary
GPU acceleration of collinear and noncollinear density functional theory in numerical atomic orbital-based codes demonstrates practical speedups for large-scale calculations The GPU acceleration achieved for collinear DFT (3. p. 027 times speedup in diagonalization for a 1,200-atom system) falls significantly short of the theoretical peak performance ratio between the H100 (51 TFLOPS with Tensor Cores) and the Xeon 8468 (3. p. 02 TFLOPS)—a factor of approximately 16 This discrepancy was primarily because only dense matrix multiplications and eigenvalue solvers were offloaded to the GPU
Methodology for Acceleration
The researchers implemented GPU acceleration in OpenMX version 3.9.9 by modifying routines during diagonalization, specifically focusing on matrix multiplications and eigenvalue solves For collinear DFT, the procedure involves several steps:
-
Diagonalize S to obtain a unitary matrix U and a diagonal matrix s such that U†SU = s
-
Construct V = Us−1/2
-
Form H′ = V†HV
-
Diagonalize H′ to obtain eigenvalues ε and eigenvectors C′
-
Recover the eigenvectors of H as C = VC′
Performance Analysis for Collinear DFT
Benchmarks on the Pegasus supercomputer compare GPU-accelerated and CPU-only runs under identical settings, showing speedups for specific system sizes For a 512-atom collinear case on two nodes (two GPUs total), the GPU-accelerated calculation achieves a 2.02 times speedup over a CPU-only run on two nodes (96 CPU cores) The diagonalization step dominates the cost, accounting for 82. p. 05% of the total wall time in subsequent SCF cycles
Performance Analysis for Noncollinear DFT
For noncollinear DFT, the computational flow involves constructing noncollinear Hamiltonian HNC, coefficient matrix CNC, and overlap matrix SNC The diagonalization procedure is more complex, involving similarity transformations and solving the generalized eigenvalue problem HNCCNC = SNCCNCε In the noncollinear case, the combined share of eigenvalue solutions (36.2%) and matrix multiplications (23. p. 09%) is 60.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Offload matrix multiplications and eigenvalue solves to cuBLAS/cuSOLVER using OpenACC to accelerate density functional theory calculations in numerical atomic orbital based codes like OpenMX. This allows for significant speedup in collinear DFT, achieving 2.02 times speedup for a 512-atom case on two nodes compared to a CPU-only run.
-
Implement GPU acceleration for both collinear and noncollinear density functional theory calculations in OpenMX by modifying the routines Band DFT Col and Band DFT NonCol, allowing these codes to handle noncollinear spin orientation variation three-dimensionally. This extends GPU acceleration beyond collinear DFT implementations to a more general class of electronic structure problems.
-
Optimize the diagonalization phase on GPUs for both collinear and noncollinear DFT by replacing ScaLAPACK with cuBLAS and ELPA with cuSOLVER, which operates directly on the full matrix held by a single MPI process rather than relying on block-cyclic distribution across MPI ranks. This modification allows GPU acceleration to be applied to stages that were previously difficult to accelerate or required costly redistribution steps.
-
Address performance bottlenecks in noncollinear DFT by examining the
active time
versus total processing time, which reveals that the serialization of preceding steps and the overhead of matrix redistribution account for non-negligible time even in GPU implementations. This insight guides future work on eliminating these serial dependencies to improve overall throughput. -
Improve scaling behavior for large systems by analyzing GPU utilization and data transfer fractions as a function of system size and node count, identifying that performance saturation occurs due to reduced overlap when increasing node counts beyond four GPUs for collinear DFT. This information informs the design of future parallel strategies for ultra-large-scale simulations.
Sources
- High-Precision Numerical Simulations of Rotating Black Holes Accelerated by CUDA
- GPU Acceleration of Numerical Atomic Orbitals-Based Density Functional Theory Algorithms within the ABACUS package
Related papers
- Gradient-Based Inverse Design of Free-Energy Landscapes with Diffusion Models
- Mixture of Polyconvex Neural Potentials for Parametric Hyperelasticity: Towards Foundation Material Models
- Development and large-scale benchmarks of a protein--ligand absolute binding free energy toolkit
- Spin disorder competing with positional symmetry breaking governs the metal-insulator behavior in oxide paramagnets
- Investigation of the Asymptotic Properties of Active Impedance in Large Finite Array Antennas
- Automatic generation of exchange-correlation response kernels