AI Computing Broker and AI based quantum chemistry solution: middleware technology to maximize GPU resources and innovation through AI based quantum chemistry simulation with the middleware

In recent years, HPC system needs to accelerate not only traditional simulation applications but also AI workloads. And integrating AI into these simulations is now critical to achieve breakthroughs in performance and sophistication, making the seamless connection of HPC and AI a significant challenge.

To solve these challenges, Fujitsu is developing two technologies: “AI Computing Broker” which is middleware to improve GPU utilization, and “GeNNIP4MD” which computes quantum chemistry potential energy by AI and enables molecular dynamics simulation with chemical reaction.

AI Computing Broker maximizes the utilization efficiency of GPU resources. This technology detects idle GPU time, detaches GPUs from programs using them, and attaches them to other programs, thereby increasing utilization efficiency. GPUs are not only valuable as computing resources but also consume a significant amount of power, so maximizing GPU utilization efficiency leads to cost reduction and power saving.

Next, by incorporating AI into simulations, we are working on the development of molecular dynamics simulation technology that enables the analysis of chemical reactions, which has been difficult to achieve until now. To reproduce chemical reactions, it is necessary to perform molecular dynamics computations based on potential energy using quantum chemical simulations such as the DFT method. However, the DFT method requires a huge amount of computation time, making it difficult to achieve in a realistic time frame. Therefore, by using DFT simulation results as data and substituting energy computation with AI models called GeNNIP4MD, Fujitsu is realizing molecular dynamics simulations that can analyze chemical reactions in a realistic time frame.

To efficiently operate such workloads, a system is needed that efficiently executes traditional simulation applications, AI training, and inference. The AI Computing Broker enables efficient and simultaneous execution of AI training and inference, allowing for the efficient operation of workloads that fuse HPC and AI. This presentation will describe these technologies.