通过对智能处理单元的共同设计优化加速基于图形神经网络的化学预测模型
Hatem Helal1, Jesun Firoz2, Jenna A Bilbrey3
1Graphcore, Kett House, Station Rd, Cambridge CB1 2JH, U.K.
Journal of chemical information and modeling
|February 21, 2024
概括
本研究介绍了用于训练原子图神经网络 (GNN) 的硬件软件代码设计,加速原子结构预测. 新方法提高了计算效率和性能,在专门的硬件上优于传统方法.
科学领域:
- 计算物理和化学计算物理和化学
- 材料科学 材料科学 材料科学
- 机器学习 机器学习
背景情况:
- 原子结构和属性预测的高可靠性ab initio方法在计算上昂贵.
- 机器学习 (ML) 模型,特别是图形神经网络 (GNN),提供了一个更有效的替代方案.
- 在大型原子数据库上训练GNN带来了与可变图形大小和通信模式相关的独特计算挑战.
研究的目的:
- 开发和演示一种新的硬件-软件代码设计方法,以扩大原子 GNN 的培训.
- 提高基于GNN的原子结构和属性预测的效率和性能.
- 为了优化GNN培训专门的硬件,如Graphcore的情报处理单元 (IPU).
主要方法:
- 制定了变量大小的原子图的分批处理,作为一个垃圾箱包装问题,使用硬件不可知论算法来最大限度地减少冗余和通信.
- 实现了Graphcore IPU的硬件特定优化,包括计划器,聚散运算的向量化,以及模型特定的优化 (合并通信集体,优化的softplus).
- 在Graphcore的IPU上部署了一种成熟的原子化GNN模型,并对各种原子图数据库的性能进行了评估.
主要成果:
- 与基线IPU实现相比,代码设计方法将原子化GNN的培训时间缩短了多达1.5倍.
- 在IPU上实现了1.8×的平均加速度,与Nvidia基于GPU的实现相比,用于原子化的GNN.
- 在具有不同特征 (计数,大小,稀疏性) 的原子图数据库中证明了性能和效率的提高.
结论:
- 拟议的硬件-软件代码设计有效地扩大了对结构和属性预测的原子 GNN 的培训.
- 这种方法显著提高了计算效率,导致更快的训练时间和更好的模型性能.
- 优化的IPU实现与基于GPU的解决方案相比显示出更高的性能,突出了专门硬件对科学ML的潜力.
相关概念视频
02:24Predicting Reaction Outcomes
8.4K
Kinetics describes the rate and path by which a reaction occurs. In contrast, thermodynamics deals with state functions and describes the properties, behavior, and components of a system. It is not concerned with the path taken by the process and cannot address the rate at which a reaction occurs. Although it does provide information about what can happen during a reaction process, it does not describe the detailed steps of what appears on an atomic or a molecular level. On the other hand,...
8.4K
02:27Predicting Molecular Geometry
34.3K
VSEPR Theory for Determination of Electron Pair Geometries
34.3K
02:00Molecular Models
38.3K
Physical models representing molecular architectures of chemical compounds play essential roles in understanding chemistry. The use of molecular models makes it easier to visualize the structures and shapes of atoms and molecules.
38.3K
01:27Inductive Effects on Chemical Shift: Overview
1.1K
The protons in unsubstituted alkanes are strongly shielded with chemical shifts below 1.8 ppm. Methine, methylene, and methyl protons appear at approximately 1.7, 1.2 and 0.7 ppm, while the proton signal from methane appears at 0.23 ppm. An electronegative substituent, such as chlorine, withdraws the electron density from the protons, increasing their chemical shift. Progressive substitution of the hydrogens in methane by chlorine shifts the proton signals increasingly downfield, to 3.05 ppm in...
1.1K
01:15Optimizing Chromatographic Separations
396
Optimizing chromatographic separations is crucial for obtaining clean separations in a minimum amount of time. Optimization is required for several factors, including kinetic effects related to band broadening, plate height, capacity factor, and separation factor.
Band broadening refers to spreading solute bands as they travel through the column. This broadening can impact resolution. Plate height (H) represents the length required for one theoretical plate. A lower plate height corresponds to...
Band broadening refers to spreading solute bands as they travel through the column. This broadening can impact resolution. Plate height (H) represents the length required for one theoretical plate. A lower plate height corresponds to...
396
01:27π Electron Effects on Chemical Shift: Overview
1.1K
An applied magnetic field causes loosely bound π-electrons in organic molecules to circulate, producing a local or induced diamagnetic field over a large spatial volume. As the molecules tumble in solution, the field generated by π-electrons in spherical substituents results in a zero net field. However, the net field generated by π-electrons in non-spherical substituents is not zero. The effect of this induced field depends on the orientation of the molecule with respect to B0,...
1.1K