Which LRZ systems were deployed?
Chung: We launched the project about a year ago and have carried out test calculations on various systems to scale them up to increasingly larger molecules. We therefore started by using 20, then finally 52 qubits from Euro-Q-Exa. We also carried out calculations using HPC resources from the BEAST test environment, which includes GPUs and direct connection to the Euro-Q-Exa system of this use case. In the pipeline of this use case, there are compute-intensive steps as post-processing that requires GPUs for acceleration. We increased the number of GPUs with the BDAI systems and finally with NVIDIA’s, in total up to 1,200. Our aim was to find out whether scaling up the systems has any benefits to the calculations or not.

However, quantum computers, HPC systems and AI systems are based on different mathematical methods.
Chung: How the simulation tasks are distributed between different architectures depends entirely on the computational complexity of each molecular sub-problem. We first get a clear picture of the required tasks, develop a pipeline for pre- and post-processing the data, and then decide which combination of algorithm and architecture is best suited to carry out which tasks. For the classical components of a molecular simulation like this, we rely on established, validated open-source software and structural models, which are standard in the science community. For hybrid computing, it is essential that this same level of abstraction and rigorous validation is maintained when transitioning to quantum architectures. While several frameworks and libraries are currently emerging to support this integration and preserve software maturity, our project specifically leveraged CUDA-Q to achieve this unified, validated environment for all technologies.

Do the results from the various systems differ? How? 
Chung: A quantum computer does not produce a single result like a classical computer, but rather a probability distribution across possible outcomes. Ultimately, however, we always end up with digital calculation results that we can work with. More important than the results is the question of which system is best suited for which part of the calculation. Supercomputers deliver the best results for computationally intensive tasks; they process large amounts of data and carry out extensive calculations, often in numerous iterations thanks to their parallelised architecture. Quantum computers are currently still limited by the number of available qubits, so they cannot yet handle very large calculations on their own. The challenge, therefore, lies in identifying the particularly complex and difficult parts of a calculation that are suitable for quantum computing.

At the moment, one more Euro-Q-Exa system with the computing power of 150 qubits is being installed at LRZ. Could the molecular simulation then be carried out entirely on a quantum computer, and would that make sense?
Chung: In principle, that could be possible, but it would require a fundamental redesign of the entire pipeline to run natively on a quantum computer. Given the current constraints around qubit counts and error rates, relying solely on quantum hardware isn't realistic yet. That’s why a hybrid approach - integrating HPC, AI, and quantum computing - is so compelling, and it’s precisely what our project delivers. Our project is specifically designed around the collaboration of multiple technologies. Running the simulation solely on a supercomputer would probably not yield results of the same quality. Whether it makes sense to perform the molecular simulation on a larger system such as the bigger Euro-Q-Exa computer is ultimately a question for chemists and biologists. I can imagine that a larger system could enable even more detailed modelling of the active regions of aGPCRs and a deeper understanding of their mechanisms and functions. (LRZ | vs)

Translation generated using AI