Expanding human potential: Deepin Technology uses IPU to empower molecular dynamics

December 9, 2021, Beijing, China - Deepin Technology announced today that it has completed the migration of the molecular dynamics simulation software DeePMD-kit to Graphcore (to be) IPU hardware, indicating that IPU has officially supported DeePMD-kit, a cutting-edge molecular dynamics model that won the Gordon Bell Prize. From now on, researchers can explore molecular dynamics applications and material simulations on IPU. Deepin Technology and Graphcore will continue to explore the potential of AI in scientific computing based on IPU, and jointly explore scenarios such as scientific computing, drug design, material design, and new energy based on molecular dynamics simulation. Currently, developers can obtain DeePMD-kit based on the IPU hardware migration in Graphcore's GitHub.

AI empowers HPC: DeePMD-kit brings disruptive innovation to molecular dynamics simulation

DeePMD-kit is an open source project for deep potential molecular dynamics. Molecular dynamics is a set of molecular simulation methods that mainly rely on Newtonian mechanics to simulate the motion of molecular systems. Molecular dynamics plays a supporting role in the simulation of the microscopic world and has very important applications in physics, chemistry, biology, materials science, mechanics and other fields.

For a long time, traditional molecular dynamics has faced the problem of "fast but inaccurate" and "accurate but not fast". The key to molecular dynamics is to accurately characterize the forces on atoms in the system by constructing potential functions. There are usually two methods for constructing potential functions. One is the empirical force field, which uses empirical parameters to fit the forces between atoms, and the main fitting object is the experiment. This type of method is simple to calculate, has good scalability, and can simulate a large scale. However, the process of constructing the potential function is long, takes a lot of time and effort, and has poor scalability. Every time a new system is added, the potential function needs to be reconstructed; the other method starts from quantum mechanics, makes appropriate simplifications based on quantum mechanics, and uses density functional theory and first principles to characterize the forces on atoms. This type of method is accurate, but the atomic systems that can be processed are usually relatively small.

Deepin Technology has successfully parallelized DeePMD-kit, realizing for the first time the commercial and atomic-scale application of first-principles molecular dynamics simulations with AIMD accuracy, providing a revolutionary high-performance solution for molecular dynamics. Due to its extreme innovativeness, this project of Deepin Technology won the Gordon Bell Prize, the highest award in the field of high-performance computing in the world, in 2020.

"IPU hardware and Poplar SDK environment are very suitable for molecular dynamics simulation. We worked with Graphcore to customize and optimize the framework and algorithm, and finally realized the migration of DeePMD-kit to IPU hardware. The entire migration process was very successful. All DeePMD-kit reasoning-related functions can be successfully deployed on IPU hardware, and the performance has been improved, thus realizing the application of molecular dynamics and material simulation on IPU." said Lu Denghui, algorithm researcher at Deepin Technology.

IPU brings advanced AI capabilities to molecular dynamics

IPU (Intelligent Processing Unit) is a processor designed from scratch by Graphcore for machine intelligence workloads. This time, Deepin Technology has deployed the microscopic world of molecules on the IPU, bringing advanced AI methods and hardware into the traditional scientific computing field of molecular dynamics. This is an exploration of extraordinary significance. The boundaries of scientific computing have been pushed further, and more potential has been stimulated by the leverage of AI.

深势科技选择IPU进行分子动力学任务迁移和模拟，原因在于IPU在硬件、软件以及迁移成本上都有显著优势。在硬件上，IPU硬件比较适合分子动力学模拟发展，尤其是可以提供良好的节点间通信。“此次和Graphcore公司合作，是我们向Multi-device硬件扩张上的一个积极探索，我们发现，IPU相较于传统的AI芯片，对于分子动力学模拟具有特殊的加速效果。”路登辉表示。

Poplar software helps molecular dynamics models to be easily transferred to IPU

In terms of software, Graphcore Poplar SDK is easy to use and is very suitable for molecular dynamics simulations. The process for developers to transfer DeePMD-kit to IPU is very easy and fast, and the software migration cost is low.

"After discovering that the IPU architecture is naturally suitable for molecular dynamics, we began to worry about whether our model tasks could be smoothly transferred to the IPU. As a result, in the early stage of our cooperation, two of our interns used their spare time during the summer vacation to run DeePMD-kit on the IPU." Lu Denghui, an algorithm researcher at Deepin, said, "Transferring the algorithm tasks of DeePMD-kit from GPU to IPU is like an engineer switching from C++ to Python programming. It is a wonderful process."

In summary, the significance of the cooperation between Graphcore and Deepin lies in the establishment of technical know-how for implementing molecular dynamics simulation software on IPU, and the excellent versatility of IPU and Poplar SDK. At the same time, it also enriches the hardware platform diversity of DeePMD-kit, providing more possibilities for application scenarios based on molecular dynamics simulation such as scientific computing, drug design, and material design, and also providing more possibilities for empowering traditional physical applications through machine learning.

Zhang Linfeng, Chief Scientist of Deepin, said: "Practice has proved that Graphcore's IPU hardware and corresponding SDK environment are very suitable for molecular dynamics simulation. From the perspective of the development of the entire open source community, this is also a beneficial attempt and exploration. The successful migration to IPU is just the beginning. Whether in terms of IPU hardware and Poplar software properties, or in terms of the algorithm of the DeepPMD-kit method itself, IPU has great advantages and there is still considerable potential to be explored. In the future, we will continue to explore more with Graphcore in hardware and software algorithms, adaptation, and collaboration to better promote the development of the DeepModeling open source community."

Lu Tao, President of Graphcore Greater China and Global Chief Revenue Officer, said: "Graphcore IPU is designed for the next generation of machine intelligence and is co-designed with Poplar SDK. The innovative architecture and software stack enable it to be quickly migrated across domains. Graphcore's support for DeepPMD-kit is another milestone in Graphcore's implementation in all-round business scenarios, and it is also one of the practices of Graphcore and its partners to jointly build a prosperous community. Graphcore will continue to deepen cooperation with partners including Deepin Technology to jointly promote application implementation and continuously trigger innovation."