Graphics processors are taking on a new role beyond gaming and artificial intelligence – they are now serving as surrogate quantum computers until the real hardware arrives.

The Jülich Supercomputing Centre is using GPUs and a software toolkit from Nvidia to emulate quantum computers and research algorithms for such systems. With quantum processors still under development, GPUs are the fastest circuits to play the role of fully operational quantum computers.

Jülich has been simulating quantum systems for more than 20 years, but the code has been optimized for CPUs. The supercomputing center has bought in Nvidia’s QODA to help move that compute over to graphics processors, said Kristel Michielsen, who leads a research group on quantum information processing at the center.

“This code has been optimized for CPUs. It’s a big, massively parallel code written in Fortran. We also have implemented quite recently – two years ago – on the GPUs. We have the Nvidia GPUs. We have been in discussion with Nvidia for some of the optimization and implementation,” Michielsen said.

Nvidia recently launched Quantum Optimized Device Architecture (QODA), which provides a software layer through which GPUs can emulate quantum computers and run algorithms. QODA is to quantum computing on GPUs what Nvidia’s CUDA is for AI and scientific applications on GPUs.

The road for Jülich Supercomputing Centre to practical quantum computing goes through conventional computing, including software emulators like QODA. Jülich researchers prepare an algorithm for the real device based on such simulators, which play an important role in preparation for deployment of applications on real quantum computers.

“We have the vision that in order to come to practical quantum computing – for use cases in industry and so on – we will have to go to hybrid quantum classical complement,” Michielsen said.

Jülich has declared it will house Europe’s first supercomputer that will break the exascale mark. The center already has D-Wave’s quantum annealer – the only such system in Europe – and a quantum simulator of the French startup Pasqal, which is focused on building a quantum computer based on neutral atoms. It also has Atos’ quantum learning machine, which can simulate up to about 40 qubits. The center is also looking to install or provide cloud-based access to other types of hardware, including an ion-trap and superconducting qubit system.

“The idea of Jülich is to provide our users with uniform access to a large variety of quantum computers, which have different technological operations,” Michielsen said.

The QODA software toolkit can simulate a wide range of applications, including drug discovery, weather and logistics operations. The toolchain makes it easier to simulate quantum and classical computing, something that can be challenging on real quantum hardware.

For HPC users that are not familiar with quantum computing and simulating quantum systems, QODA might be of great help in developing code for hybrid quantum-classical systems.

“It contains already familiar elements for programming HPC systems and the GPUs,” Michielsen said.

Michielsen said that one can write hybrid algorithms that will do function calls to the quantum resources, much like they do with GPUs in high-performance computing environments.

“In this respect, one can in the end see that the quantum processing unit will play a role as an accelerator, which is also the role of the GPU,” Michielsen said.

As part of fundamental research, Jülich researchers simulate real models of the quantum computers to find out how they operate.

“We start from theory, and make a model. And for this we rely on the quantum computer developers to give us parameters in a first model. And then we compare this to what an actual machine is producing. And when you look at the differences we find, usually we find some discrepancies,” Michielsen said.

Then the researchers go to the applications with the idea to see how specific systems perform on solving specific problems. For example, the D-Wave quantum annealer has been solving optimization problems with simplified and smaller problems.

The road to practical quantum computing isn’t simply translating an algorithm for a classical HPC system to work on a quantum computer as these are very different algorithms, Michielsen said, highlighting the importance of quantum algorithm development.

“If one thinks about this hybrid quantum classical computing, we do not only have to look into these variational algorithms because, by the way, they are usually not the most efficient ones. That we already noticed. What one has to do is take classical workflows for classical computing, and look at whether there are some paths which one can give to a quantum computer,” Michielsen said.

In one case, the researchers at Jülich simulated the problems that airlines have with optimization of routes at minimal cost. Airlines typically have many constraints on optimizing routes, which may include aircraft, personnel and cost. Those problems have led to many flight delays and cancellations since the start of the year.

“One has this problem then that has to translate into, in this case, a quadratic unconstrained binary optimization problem. And it’s obvious one cannot solve the problems that airlines are dealing with. One has to simplify it and make it small. We could solve our problem which was including 472 flights, and this translated into a quantum optimization problem with 40 qubits,” Michielsen said.

On quantum approximate optimization algorithms like with the airline problem, the Jülich researchers have given CPUs the classical computing optimization part of this hybrid algorithm, with the GPU playing the role of a digital quantum processor.

“Because we have emulated this with a simulator on the supercomputer, if we have a real device which is big enough…then we can really carry out this algorithm on your quantum device. In this sense, we somehow prepare an algorithm for the real device based on such emulators, so in that sense, they play a very important role,” Michielsen said.