LUMI enables simulations of large-scale algorithms for quantum advantage

Quantum computing holds the promise of revolutionising computational modelling. Recent progress in constructing quantum computers has been rapid – already now, science that previously was impossible can be performed on real quantum devices. What we still need, is a larger selection of quantum algorithms and software to drive quantum-accelerated supercomputing further. For this, tools for simulating the behaviour and performance of quantum algorithms on classical computers are crucial. Now, full simulation of quantum algorithms utilising up to 44 qubits can routinely be explored on the LUMI supercomputer, pushing the frontiers for the next generation of science and discovery.

An AI-view of the LUMI supercomputer hard at work emulating the behaviour of qubits. Image: Generated with ChatGPT.

Even with real quantum computers already available, the ability to simulate quantum algorithms and circuits on traditional supercomputers is essential. Current quantum computers belong to the so called Noisy Intermediate-Scale Quantum (NISQ) category. While highly valuable for scientific exploration, NISQ devices are limited by suboptimal error rates and short coherence times. In non-technical terms, this implies that current devices are limited to calculations that can finish quickly. The challenge is to develop algorithms that can perform useful computations with a limited number of operations.

On the path towards algorithms of useful complexity, simulating quantum algorithms on supercomputers provides a crucial tool for several reasons:

1. Validation and verification: Classical simulations help develop, validate, and verify quantum algorithms before they are run on actual quantum hardware. This ensures that the algorithms work as intended and helps identify potential errors.
2. Understanding quantum systems: Simulations provide insights into the behaviour of quantum circuits, which is essential for developing more efficient quantum algorithms which will help pave a path towards prioritizing the improvements that will be made to future quantum hardware.
3. Debugging of quantum software: In contrast to physical quantum computers, simulators can provide the opportunity to, for example, follow the evolution of quantum states throughout the simulation. Physical qubits do not allow probing the state of a qubit during the calculation.
4. Development of hybrid systems: Classical simulations are instrumental in the development phases of hybrid quantum-classical systems, where both types of computing are used together to solve complex problems.

There are several different approaches to simulating quantum algorithms with classical computers. They can, arguably, be divided into two main classes: tensor network simulations and full state-vector simulations. While tensor network simulations have been used to simulate algorithms of thousands of qubits, they come with an inherent limitation: they are only suitable for quantum algorithms with a rather low degree of entanglement between the qubits. Entanglement, the “spooky action at a distance” as Einstein infamously called it, is a pre-requisite for many of the use cases where quantum computing is expected to provide tangible advantage. After all, if a quantum algorithm can be efficiently simulated on a classical computer, there is less need for a quantum computer to begin with.

For exploring algorithms utilising higher levels of entanglement, a “Schrödinger-style” full state-vector simulation becomes necessary. This type of simulation is generally much more resource intensive but offers the advantage of detailed inspection of how the quantum states change and develop during the execution of the algorithm. This is where the need for supercomputers enters.

Putting parallel resource to good use

State-vector simulation of quantum circuits is highly resource consuming, especially when the number of qubits grow. A standard laptop with 16 GB of memory cannot do a full simulation beyond 30 qubits. Even a high-end personal computer with 256 GB of memory allows chokes around 34 qubits. This is because the memory requirement of simulation grows exponentially with the number of qubits – each additional qubit practically doubles the needed memory capacity.

By distributing the simulation over several GPU nodes of the LUMI supercomputer, both the combined compute and memory capacity can be used for a single simulation. With the new simulation environment set up on LUMI, quantum algorithms of up to 44 qubits can now be simulated efficiently. A simulation of this size requires 256 TB (terabytes) of combined RAM memory and 1024 LUMI-G nodes, that is, 4096 AMD Instinct MI250X GPUs with a total of 8192 GPU Graphics Compute Dies (GCDs).
GPUs are a natural choice for accelerating full state-vector simulations of quantum circuits due to several factors:

1. Linear algebra operations and massive parallelism. State-vector simulations rely heavily on linear algebra operations like matrix multiplications and vector additions. These operations are inherently parallel, meaning they can be divided into smaller tasks that can be executed simultaneously. GPUs are designed for exactly this type of parallel processing.
2. High memory bandwidth. Simulations of quantum states can involve large datasets, which can put a strain on memory access. GPUs typically have high-bandwidth memory, which allows for faster data retrieval and minimises delays caused by memory bottlenecks, especially as the number of qubits increases.
3. Support for optimised libraries. GPUs are supported by libraries that are specifically optimised for accelerating tasks like matrix operations and Fourier transforms, which are common in quantum simulation domain. These libraries are tailored to GPU architecture, improving the performance of these operations.
4. Energy efficiency and cost-effectiveness. Compared to CPU-based systems, GPUs are generally more energy-efficient for the kind of parallel processing needed in large-scale simulations. This can reduce the operational cost for running simulations of large quantum circuits.

Summary

As a result of tailoring the supercomputer environment, it is now possible to routinely simulate quantum algorithms utilising up to 44 qubits on LUMI. This requires that several aspects of the simulation parameters need to be considered, such as total GPU node count. With this new toolset available for all LUMI users, quantum software development can now be taken to new heights. We hope that this will catalyse the development of novel quantum algorithms that can usher in the dawn of quantum advantage.

For a more in-depth, under-the-hood treatise on the large-scale simulations, check out the technical blog on the Finnish Quantum-Computing Infrastructure (FiQCI) website.

Authors: Michael Mucciardi, Senior Systems Specialist, CSC, and Mikael Johansson, Manager, Quantum Technologies, CSC.