Minisymposium

In 1975, the Cray 1 changed the definition of high performance computing (HPC). Essentially a single if extremely large processor, it would not be recognisable to many of those dependent on scientific computing today. Some components that we now see on-chip were in entirely separate cabinets, and hard drives resembled top-loading washing machines.

Modern supercomputers are quite different beasts, consisting of many highly interconnected processors. Vector processing is still delivering high performance, which drives investment in technologies such as GPUs and on-chip SIMD units, but high bandwidth, low-latency, reliable interconnect is paramount to large-scale scientific computation.

Quantum computing is looking more and more likely to become a realistic prospect for scientific computation during our lifetimes. Although exciting, it cannot be embraced without understanding some fundamental differences between this and what is now thought of as ``classical'' computing.

It is widely known that not all problems are are suitable for quantum computation, so it is tempting to consider any future approaches using QPUs to be similar to those taken with modern GPUs. Unfortunately, this simple substitution will not work for fundamental reasons.

In this talk, we explain this and highlight some of the differences.

Driven by the rapid development of intermediate-scale quantum devices, hybrid classical-quantum workflows have come into the focus of many leading HPC service providers including clouds and supercomputing centers. This interest is enhanced by promising early results such as variational methods for quantum chemistry and combinatorial problems but also more widely applicable approaches like circuit knitting. In this talk, we summarize experiences from early attempts to combine quantum devices and an HPE-Cray EX supercomputer by means of simple proxy applications, and how to execute a hybrid HPC-QC workflow efficiently using state-of-the-art frameworks such as the message passing interface and the Slurm job scheduler.

One of the primary goals at the LRZ is to steer quantum circuit compilation to unlock the full potential of quantum accelerators within HPC environments. During the presentation, I will explore the details of our approach with the Munich Quantum Software Stack (MQSS), specifically around quantum circuit optimization. To address these challenges, we have developed a range of heuristics to guide different types of design space explorations. Specifically, I will discuss: 1) our novel circuit cutter technique, which divides circuits into smaller segments to enhance compatibility with available accelerators. 2) Our dynamic scheduling approach, where target accelerators are chosen for each segment to optimize resource usage. And 3) our heuristic selection process for custom-made LLVM passes, aimed at finding a balance between optimization and computational efficiency. Furthermore, central to our optimization efforts is the seamless interaction with quantum devices to retrieve relevant information during the JIT compilation stages. This includes critical data such as the aforementioned supported gate sets and coupling mappings of the targeted device for accurate circuit mapping. I will explain how we interface our devices to facilitate this interaction, allowing us to access complex information such as device topologies and lists of available quantum devices.

Quantum computing systems continue to scale in size and quality. Furthermore, errorresilience approaches start to enable interesting computational regimes opening whatwe call the era of quantum utility. In this era, the integration of quantum with HPCbecomes critical to unlock the full potential of both technologies in a way that exceedsthe capabilities of either one alone. Our vision for the path forward is quantum-centricsupercomputing: quantum algorithms and routines supported by multiple quantumcircuits with the aid of classical pre- and post-processing and classical high-performanceelements.The development of tightly integrated systems with quantum and classical elements isessential for this vision. Not only will they provide an invaluable development frameworkfor computational scientists, both quantum and classical, but they will also provide atesting ground for the joint optimization of classical and quantum resources needed torun utility scale hybrid workflows.Given the fast development of quantum technologies, we believe now is the time tobridge the gap between classical and quantum computing to define and shape a newera of supercomputing.