NVIDIA's CUDA-Q Reduces Resources for Quantum Clustering Algorithms

NVIDIA has announced a significant advancement in quantum computing through its CUDA-Q platform, formerly known as CUDA Quantum. According to the NVIDIA Technical Blog, CUDA-Q has been instrumental in reducing the resource requirements for quantum clustering algorithms, making them more practical for near-term quantum computing applications.

Quantum Clustering Algorithms

Quantum computers leverage the unique properties of superposition, entanglement, and interference to derive insights from data, offering theoretical speedups over classical computing methods. However, early quantum computers are expected to excel at compute-intensive tasks rather than data-intensive ones due to the absence of efficient quantum random access memory (QRAM).

Associate Professor Dr. Petros Wallden and his team at the Quantum Software Lab, University of Edinburgh, utilized CUDA-Q to simulate new quantum machine learning (QML) methods. These methods significantly reduce the qubit count required to analyze large datasets. Their research extended Harrow's work on coresets, a classical dimensionality reduction technique, to make QML applications more feasible without the need for QRAM.

What Are Coresets?

A coreset is a smaller, weighted subset of a full dataset that approximates the traits of the full dataset. This method allows for data-intensive QML tasks to be performed with significantly fewer qubits. Petros’ team chose the coreset size based on available qubits and then assessed the resulting error after quantum computations.

Quantum Approaches for Clustering with Coresets

With the input data reduced to a manageable size, three quantum clustering algorithms were explored:

• Divisive Clustering: Points are successively bipartitioned until each point is in its own cluster.
• 3-means Clustering: Points are partitioned into three clusters based on their relationship to evolving centroids.
• Gaussian Mixture Model (GMM) Clustering: Points are sorted into sets based on a mixture of Gaussian distributions.

Each method outputs a set of coresets, mapping the initial dataset to these coresets, resulting in an approximate clustering and dimensionality reduction of the dataset.

Using CUDA-Q to Overcome Scalability Issues

Exploring these QML-clustering approaches required simulations, which CUDA-Q facilitated by providing easy access to GPU hardware. This allowed Petros’ team to run comprehensive simulations on problem sizes up to 25 qubits. CUDA-Q's primitives, such as hardware-efficient ansatz kernels and spin Hamiltonians, were crucial for the Hamiltonian-based optimization process.

Early simulations ran on CPU hardware but were limited to 10 qubits due to memory constraints. Switching to NVIDIA DGX H100 GPU systems allowed the team to scale up to 25 qubits without modifying their initial code, thanks to CUDA-Q's compatibility and scalability features.

Value of CUDA-Q Simulations

Simulating all three clustering algorithms enabled benchmarking against classical methods like Lloyd's algorithm. The results showed that quantum algorithms performed best for GMM (K=2) and divisive clustering approaches.

Petros’ team plans to continue collaborating with NVIDIA to develop and scale new quantum-accelerated supercomputing applications using CUDA-Q.

Explore CUDA-Q

CUDA-Q enabled the development and simulation of novel QML implementations, with the code being portable for further large-scale simulations or deployment on physical quantum processing units (QPUs).

For more information and to get started with CUDA-Q, visit the NVIDIA Technical Blog.