An official website of the United States government
Low-precision arithmetic has had a transformative effect on the training of neural networks, reducing computation, memory and energy requirements. However, despite its promise, low-precision arithmetic has received little attention for Gaussian processes (GPs), largely because GPs require sophisticated linear algebra routines that are unstable in low-precision. We study the different failure modes that can occur when training GPs in half precision. To circumvent these failure modes, we propose a multi-faceted approach involving conjugate gradients with re-orthogonalization, mixed precision, and preconditioning. Our approach significantly improves the numerical stability and practical performance of conjugate gradients in low- precision over a wide range of settings, enabling GPs to train on 1.8 million data points in 10 hours on a single GPU, without any sparse approximations.
more »
« less
This content will become publicly available on December 1, 2025
Scalable computation of anisotropic vibrations for large macromolecular assemblies
Abstract The Normal Mode Analysis (NMA) is a standard approach to elucidate the anisotropic vibrations of macromolecules at their folded states, where low-frequency collective motions can reveal rearrangements of domains and changes in the exposed surface of macromolecules. Recent advances in structural biology have enabled the resolution of megascale macromolecules with millions of atoms. However, the calculation of their vibrational modes remains elusive due to the prohibitive cost associated with constructing and diagonalizing the underlying eigenproblem and the current approaches to NMA are not readily adaptable for efficient parallel computing on graphic processing unit (GPU). Here, we present eigenproblem construction and diagonalization approach that implements level-structure bandwidth-reducing algorithms to transform the sparse computation in NMA to a globally-sparse-yet-locally-dense computation, allowing batched tensor products to be most efficiently executed on GPU. We map, optimize, and compare several low-complexity Krylov-subspace eigensolvers, supplemented by techniques such as Chebyshev filtering, sum decomposition, external explicit deflation and shift-and-inverse, to allow fast GPU-resident calculations. The method allows accurate calculation of the first 1000 vibrational modes of some largest structures in PDB ( > 2.4 million atoms) at least 250 times faster than existing methods.
more »
« less
- Award ID(s):
- 2118061
- PAR ID:
- 10515622
- Publisher / Repository:
- Nature Publishing Group
- Date Published:
- Journal Name:
- Nature Communications
- Volume:
- 15
- Issue:
- 1
- ISSN:
- 2041-1723
- Page Range / eLocation ID:
- 3459
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Knowledge graph (KG) learning offers a powerful framework for generating new knowledge and making inferences. Training KG embedding can take a significantly long time, especially for larger datasets. Our analysis shows that the gradient computation of embedding is one of the dominant functions in the translation-based KG embedding training loop. We address this issue by replacing the core embedding computation with SpMM (Sparse-Dense Matrix Multiplication) kernels. This allows us to unify multiple scatter (and gather) operations as a single operation, reducing training time and memory usage. We create a general framework for training KG models using sparse kernels and implement four models, namely TransE, TransR, TransH, and TorusE. Our sparse implementations exhibit up to 5.3x speedup on the CPU and up to 4.2x speedup on the GPU with a significantly low GPU memory footprint. The speedups are consistent across large and small datasets for a given model. Our proposed sparse approach can be extended to accelerate other \revise{translation-based (such as TransC, TransM, etc.) and non-translational (such as DistMult, ComplEx, RotatE, etc.) models as well.more » « less
-
Abstract The concept of high‐entropy materials has been introduced based on the idea that multiple principal components can be mixed through the increase in configurational entropy. Implicit in this idea is that the vibrational entropy, the other component of the mixing entropy, is small compared to the configurational entropy. To explore this relationship, we examined the mixing enthalpy, configurational entropy, and vibrational entropy of two binary ceramic systems—the transition metal carbides and transition metal diborides. We computed the vibrational entropy directly using the dynamical matrices obtained from density functional theory and the quasiharmonic approximation. The mixing vibrational entropy of the mixed diborides is at least as large as the configurational entropy while it is smaller for the carbides. Utilizing the phonon density of states, we further demonstrate the origin of the high mixing vibrational entropy arises because of a large number of new low‐frequency modes that appear in the diborides. Similar modes occur in the carbides but occur at larger frequencies. These differences ultimately arise because of the structural differences where metal atoms share nearest neighbors in the diborides, while they do not in the carbides. This increased vibrational mixing entropy dramatically enhances the mixing of the diborides and demonstrates that this type of entropy cannot be neglected when considering what stabilizes mixtures and provides a new perspective on what is considered high entropy.more » « less
-
Abstract There is a pressing need for new computational tools to integrate data from diverse experimental approaches in structural biology. We present a strategy that combines sparse paramagnetic solid‐state NMR restraints with physics‐based atomistic simulations. Our approach explicitly accounts for uncertainty in the interpretation of experimental data through the use of a semi‐quantitative mapping between the data and the restraint energy that is calibrated by extensive simulations. We apply our approach to solid‐state NMR data for the model protein GB1 labeled with Cu2+‐EDTA at six different sites. We are able to determine the structure to 0.9 Å accuracy within a single day of computation on a GPU cluster. We further show that in some cases, the data from only a single paramagnetic tag are sufficient for accurate folding.more » « less
-
Sum frequency generation (SFG) necessitates both noncentrosymmetry and coherence over multiple length scales. These requirements make vibrational SFG spectroscopy capable of probing structural information of noncentrosymmetric organic crystals interspersed in polymeric matrices and their three-dimensional spatial distributions within the matrices without spectral interferences from the amorphous components. However, this analysis is not as straightforward as simple vibrational spectroscopy or scattering experiments; it requires knowing the molecular hyperpolarizability of SFG-active vibrational modes and their interplay within the coherence length. This study demonstrates how density function theory (DFT) calculations can be used to construct the molecular hyperpolarizability of a model system and combine it with the SFG theory to predict the polarization and azimuth angle dependences of SFG intensities. A model system with short peptide chains mimicking β-sheet domains in Bombyx mori silk was chosen. SFG signals of the amide-I, II, III, and A bands and one of the CH deformation modes were simulated and compared with the experimental results and the predictions from the group theory. The SFG features of amide-I and A bands of antiparallel β-sheet could be explained with DFT-based theoretical calculations. Although vibrational coupling with neighboring groups breaks the symmetry of the D2 point group, the group theory approach and DFT calculations gave similar results for the amide-I mode. The DFT calculation results for amide-II did not match with experimental data, which suggested vibrational coupling within a larger crystalline domain may dominate the SFG spectral features of these modes. This methodology can be applied to the structural analysis of other biopolymers.more » « less
